LangTrend

Large language models (LLMs) are increasingly used to annotate text for social science research, content moderation, and training data. Their broad knowledge of real-world entities raises a concern: when labeling a text, do they rely on its evidence or substitute what they already know? We call this behavior a memory shortcut and examine when it emerges, what amplifies it, and how to mitigate it. Using 995 widely recognized public figures and four matched control names per figure, we construct counterfactual texts that separate textual evidence from prior knowledge, and evaluate three commercial LLMs on 121,594 prompts each across five stages. We find a striking asymmetry. Although all models recognize nearly every figure, none follows memory over the text when the two conflict. Shortcuts instead emerge when evidence is absent: for one model, famous names raise the rate of remembered occupations, ages, and countries by 17.0, 14.9, and 6.8 percentage points relative to synthetic names. These effects occur only for recognized figures, and 98.6% of the resulting labels are factually correct about the namesake while none is supported by the text. Evaluation threats and answer-contingent rewards amplify shortcuts by 8.7 and 12.4 points. Famous names also raise leadership ratings of sparsely documented applicants by 0.22 and 0.28 points in two models. A namesake reminder and name masking each eliminate the shortcuts, with trade-offs in gender annotation. LLM annotations can thus be factually correct yet methodologically invalid. Trustworthy LLM annotation requires audits that test evidence absence, account for incentives in annotation instructions, and evaluate the attribute-specific trade-offs of mitigation.

Coding agents can destroy data, leak secrets and message the wrong person faster than a human can review. Could a decision model catch these risks at runtime? We evaluated 22 hosted decision models and two prompted general models as classifiers for a multi-rule policy engine. A synthetic dataset of 2,643 tool calls was labeled blind against 28 rules, with a sealed holdout. With a single shared preamble in both conditions, explicit boundaries written as examples and counterexamples raised recall by 5.5-26.1 percentage points on all 23 available models and reduced false positives on 21. Jev 1.14's recall rose from 89.3% to 98.2%, while unnecessary interruptions fell from 30.6% to 4.8% of 252 clean actions. With examples, its median latency was 120 ms at $0.46 per thousand requests. Strong models retained the recall gains when concrete names were replaced by generic descriptions; three largely retained them when the same boundaries were written as prose. On 16 new scenarios written after the initial studies, a follow-up improved recall on 22 of 23 models and reduced unnecessary interruptions on 21. We release the dataset, rule texts and harness.

This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity, although the results may be influenced by possible training data contamination, which is always a concern with LLMs.

LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.

Machine learning systems are challenging to test using traditional software testing methods because their behavior is learned from data rather than explicitly programmed. This challenge is particularly significant in natural language processing (NLP), where small input variations can lead to unexpected model behavior. Conventional evaluation metrics, such as accuracy and F1-score, often fail to capture such inconsistencies. Metamorphic testing addresses this limitation by evaluating expected relationships between original and transformed inputs. With the increasing prevalence of generative artificial intelligence (GenAI) and language-based AI systems, the need for systematic robustness testing has become even more pronounced. Motivated by this growing need, this paper presents a lightweight, batch-based metamorphic testing approach for evaluating the robustness of such systems. In particular, we consider sentiment analysis as a representative NLP task and evaluate a pretrained classifier under four transformations: synonym replacement, typo injection, word-order variation, and negation. Using consistency score, flip accuracy, and confidence change, we identify strengths in handling semantic variations and vulnerabilities to textual noise and logical transformations. These findings highlight the value of metamorphic testing in identifying behavioral inconsistencies beyond conventional performance metrics and demonstrate its potential as a systematic framework for evaluating the robustness of language-based AI systems and guiding performance improvements before deployment.

Software development sessions often involve several unfinished tasks rather than one issue carried through to completion. As coding agents take on repository-level development, they need to recover an earlier task's requirements and partial progress after intervening work. Existing repository-level benchmarks typically score one issue-to-patch episode and do not directly measure whether an agent can resume a task after advancing other unfinished work in the same repository. We introduce \textbf{\ourmethod}, a benchmark for focus-task resumption in sessions that interleave unfinished tasks drawn from the same source repository, with a persistent checkout for each task. Each focus task is evaluated in matched uninterrupted and switched sessions with identical focus requirements and executable verifiers. The switched session interleaves ordered stages from other unfinished tasks in the same repository, while only the final focus patch is scored. \ourmethod~contains 200 focus tasks from 19 Python repositories, reconstructed from issue, pull-request, and review histories and rendered as 400 matched session specifications. Across 7 coding agents, interleaved unfinished work produces an average focus-task resolution gap of 11.5 percentage points, ranging from 6.0 to 15.0 percentage points. Trajectory analysis identifies cross-task requirement confusion and loss of task state as recurrent resumption behaviors. We introduce a plug-in context-management module that selects task-relevant history and maintains concise task-state summaries. On OpenCode, the module improves focus-task completion under same-repository task switching by 5.5 to 15.5 percentage points across the evaluated models. Code is available at github.com/Miraclecsc/Switch-SWE.

While language models (LMs) excel in natural language tasks, their ability to learn impossible languages–those that violate the structural, cognitive, or logical constraints that make human languages learnable and functional–remains debatable. This controversy is further fueled by a divide between linguistic theorists and empirical researchers working with small-scale LMs. To address this controversy, we introduce semantic coherence. This concept measures how effectively semantic information is expressed in language structures and examines the learning capabilities of LMs across multiple scales. By fine-tuning billion-parameter LMs on impossible language samples, we demonstrate that larger LMs achieve similar performance on impossible and possible languages, whereas million-parameter LMs exhibit significant degradation when trained on impossible languages. This performance gap widens as semantic complexity increases. We further show that scaling up the LM size fundamentally shifts the learning focus from grammatical rules to semantic understanding, with billion-scale LMs exhibiting a threefold reduction in sensitivity to low-level linguistic perturbations while maintaining stable performance under high-level semantic perturbations.

Locating political parties on environmental and climate policy is central to the study of party competition, yet the established instruments, expert surveys and the hand coding of manifestos, are costly, periodic, and slow to extend to new documents. We ask whether a panel of large language models can serve as a cheaper and reproducible alternative. Treating each manifesto as a set of sentences, three models from independent providers score every sentence on two dimensions, environmental protection and climate change, separating a party's position on an issue from the salience it gives it, and the judgments are averaged to the manifesto level. Agreement across the models is retained as a built-in reliability signal. We apply the panel to 41 manifestos from the 2024 European Parliament election, covering Austria, France, Germany, Italy, Spain, and the European party families, and validate it against the 2024 Chapel Hill Expert Survey for the 25 parties with a like-for-like expert benchmark, using the other manifestos as diagnostic cases. The panel recovers the expert ordering of parties with strong rank agreement on both dimensions, while reading them as systematically greener than experts do, a gap we interpret as the distance between manifesto promise and governing record. Inter-model disagreement modestly flags where the instrument errs, and position and salience prove related for environmental protection but independent for climate change. The result is a transparent, low-cost, and multilingual instrument that complements rather than replaces expert and hand-coded approaches; the prompt, code, and manifesto-level results are released openly.

Social-media comments are interactions, yet they are often translated and evaluated as isolated text. This mismatch is especially problematic when meaning depends on surrounding discussion, community-specific slang, or culturally situated expressions. We introduce CULTURE-CMT, a Chinese-English benchmark for contextualized social-media comment translation that restores the post and conversational context naturally available to users. Its 572 manually verified examples are organized into Slang, Context, and Both, enabling diagnostic analysis of distinct and interacting sources of translation difficulty. To support scalable evaluation, we further develop a retrieval-augmented cultural-effectiveness Judger that combines a calibrated scoring model with dynamically retrieved cultural knowledge. CULTURE-CMT reveals several non-trivial patterns. Providing interaction context improves most model configurations, but does not eliminate culturally ineffective translations. Examples requiring both contextual grounding and culturally situated language remain substantially more difficult than either source of difficulty alone. Moreover, larger models are not consistently better, and thinking produces highly model-dependent gains, including occasional degradation. In contrast, a 27B model adapted on benchmark-disjoint social-media data achieves the lowest observed invalidity rate among the evaluated configurations under thinking inference with context. Together, these results suggest that realistic social-media translation is not simply a matter of scaling models or reasoning longer, but of learning to translate meaning as it is constructed within evolving online interactions.

Bursty serving changes both queue states and the population selected for each model. We study conditional calibration contracts as a specification to be validated, rather than a certified property of an existing router. CA-MHD combines load-coupled distillation, closed-loop replay and a cost-informed Schmitt router in a discrete-event simulator with a hardware service table. Across 40 cells (ten training seeds and four content-load coupling levels), we distinguish the original teacher, frozen wide heads on an adapted trunk, and the narrow student. A common teacher-trajectory audit has median absolute student-teacher deficit 0.06 percentage points, but does not certify the student-selected deployment support. Historical results use a kernel correctness gap, not the published SmoothECE; we separately prove selected-branch composition for a residual functional at a common fixed scale. The deployed wide branch has lower ordinal accuracy than the original teacher in all 40 cells. Same-checkpoint student-only serving has lower historical ordinal gap in 27/40 cells; routing also exposes violation tails up to 17.5 percent. Student p99 latency remains above the 1.5 ms target. Retrospective temporal recalibration improves the ordinal gap in 39/40 cells, but causal delayed-label availability is unverified and no incremental per-mode benefit is detected after multiplicity adjustment. An in-sample CUSUM detects 30/40 checkpoint changes at five historical false alarms, with no prospective false-alarm guarantee. These findings support a branch-aware audit and a scoped negative result, not demonstrated routing superiority. Independent budgets, selected-support validation and net utility under a satisfied SLA remain open.

Digital learning environments produce rich clickstream data that can reveal how students approach learning tasks, yet interpreting these fine-grained traces often requires substantial expert effort. We introduce ClickSight, a framework for interpreting student clickstreams through learning strategies defined in the literature. Given a student clickstream and a predefined set of learning strategies, ClickSight uses a staged LLM-based pipeline to identify strategy-relevant evidence, assess adherence to each strategy, and generate an evidence-grounded interpretation of the student's behavior. We evaluate ClickSight across two learning environments, 365 students, and 12 model backbones using an expert-grounded quality rubric and an LLM judge calibrated with human-expert ratings. Our results show that ClickSight can generate high-quality strategy-based interpretations across different models and environments, with reasoning models performing more consistently than instruction-tuned counterparts. We further find that providing external clickstream-extracted guidance can improve interpretation quality for strategies that require precise quantitative information from the trace. By connecting fine-grained clickstreams with theoretically defined learning strategies and preserving the evidence underlying each judgment, ClickSight provides a flexible approach for studying student behavior in digital learning environments.

Key-value (KV) cache eviction methods rank tokens by attention scores, and each layer scores its own cache. To our knowledge, no eviction method asks which layer should score. We find, across twenty-one models, that the depth of the scoring layer governs eviction quality. On the eighteen models with per-layer caches, spanning 0.5B to 109B, long-context retrieval accuracy collapses at shallow layers, peaks between 0.4 and 0.75 of depth, and degrades at the last layer. The curve arises because a layer judges tokens well only once it has used the question to find the answer, which takes several layers. Modeling eviction as a top-$k$ cut, we compute answer survival probability from prefill attention and reproduce the measured curves. This approach selects the scoring layer using calibration prompts with known answers, and the selected layer is the best layer measured on most models and never significantly below it. It improves on the own-layer default on sixteen of the eighteen models, with significant gains on seven, all at 14B or below, while on larger models the default is already close to the best mid-depth layer. Scoring every cache from the selected layer while keeping one set of tokens for all heads also beats per-head SnapKV by 18 to 43 points on five models up to 14B under tight budgets. This gain carries over to financial filings at 32K context, all at the same memory budget, with no extra forward pass. The same curve explains a penalty in shared-cache models, where several layers read one cache and the layer that writes it scores by default. On Gemma-4-E2B the writer is shallow and loses 28 points to a mid-depth reader at a 10% budget. We also find that in open-book question answering (LongBench) no mid-depth layer beats the own-layer default.

An input guard decides whether a request reaches the model, but never sees what the model says. Harm that emerges during generation, because a jailbreak passed the input layer or a benign request drifted into operational detail, is visible only on the output side and only as tokens arrive. We present HaloGuard 2.0, an open-weight streaming output guard that assigns a safe/unsafe probability to every response token in a single forward pass, so a deployment can stop the stream once harm is confirmed. The guard reuses the constitution of HaloGuard 1.0, our input guard, to organise a 478,532-response training corpus, and supervises each token from the point where the response becomes harmful. We find that a guard trained on conventionally shaped data learns a shortcut specific to streaming: it judges the prompt instead of the response, and handles only 13.9% of matched safe/unsafe response pairs correctly. We correct this with a second phase on 106,043 counterfactual responses built around the base guard's own failure modes, and release the equal-weight average of the base and corrective models. On six English response-safety benchmarks, HaloGuard 2.0-0.8B reaches a macro F1 of 83.6 (83.3 on held-out halves of the decision-rule search), up from 72.8, with 38% fewer false positives. It matches the strongest full-response guards with 5-10x fewer parameters and is 2.6 points above the strongest streaming guard, Qwen3Guard-Stream-4B. A multilingual variant, HaloGuard 2.0-Poly-0.8B, trained on PolyGuardMix, reaches 83.1 F1 on the response labels of PolyGuardPrompts across 17 languages, 5.0 points above the strongest baseline. On one H100 the guard costs 0.37 ms per token per stream when moderating 64 concurrent responses. Both guards are hardened continuously through the weekly red-teaming protocol of HaloGuard 1.0, extended here to the output layer.

Social media translation requires preserving meaning while capturing the cultural nuances of user-generated content (UGC). However, gaps in the cultural knowledge of LLM-based evaluators can lead to inaccurate assessments and unreliable training rewards. To improve these signals, we use retrieval-enhanced cultural evaluation to train CUE-MT, a family of social media translation models. A calibrated, domain-specific Judger guides supervised data refinement and is augmented by an Agent for hard-example mining and reinforcement learning. The Agent retrieves relevant external evidence as needed and distills it into concise hints that the Judger uses to score translations and assign rewards. Across 9B, 27B, and 35B-A3B backbones, CUE-MT outperforms the evaluated open-source baselines of comparable size on CUE-MT-Bench and three public benchmarks. Human evaluation on CUE-MT-Bench shows cultural effectiveness gains of 7.44-15.12 percentage points over the strongest external baseline in each size group. Retrieval improves the Judger's four-way classification accuracy by 8.07 percentage points over a matched no-retrieval baseline that uses LLM-generated hints alone. Training ablations further support the combination of hard-example mining and Agent-enhanced rewards.

Factual recall in language models is typically framed as retrieving a factual association: a subject–relation pair maps to a single object. However, many facts are context-dependent: the same pair maps to different objects depending on a condition such as time. We introduce anchored factual recall, in which the model receives an explicit conditioning expression, an anchor, and must select the object valid under it. This raises questions that single-object recall does not: where the anchor enters the computation, what it changes, and how that change selects the answer. Focusing on temporal anchors, we answer them with causal interventions on four open-weight models on English prompts and identify two routes. In a subject-mediated route, the anchor resolves the final subject-token representation to the queried year, and the relation redistributes the readout of this state across attention heads. In a direct route] the final position attends to the anchor, and upper-layer MLPs turn this signal into a preference for the time-appropriate object. The two routes are parallel channels: their contributions add up, most queries use both, and their balance depends on the query, e.g., on the relation, consistently across models. This reframes mechanistic accounts of factual recall from retrieving a single stored object to selecting among context-dependent alternatives.

Harness optimization can substantially improve LLM agents without modifying model weights, but existing methods primarily optimize response quality. Extending them to joint quality–cost optimization introduces a challenge: applying the deployment preference during intermediate updates can prematurely reject candidates that are valuable for later improvements. We introduce ParetoSaddler, which addresses this mismatch by separating intermediate candidate promotion from final deployment selection. At each iteration, an optimizer updates a parent harness to produce a candidate, and then compares the two over repeated executions and uses Bayesian bootstrap Pareto support to promote candidates showing credible progress in either quality or cost. Promoted candidates receive broader development evaluation and remain available as parents for later optimization steps. We evaluate ParetoSaddler on ClawMark and GAIA2 using domain-disjoint splits. On ClawMark, ParetoSaddler improves quality from 70.86\% to 81.35\% while reducing cost from \1.17 to \1.01 per task; on GAIA2, it improves quality from 56.22\% to 62.89\% while reducing cost from \0.78 to \0.45 per task. The Pareto-selected harnesses also achieve higher quality at lower cost than quality-only harness optimization. Analysis shows that Pareto-aware optimization retains intermediate candidates that would be discarded by deployment preference but can support later improvements, while Bayesian bootstrap screening filters out candidates whose gains are weakly supported across runs.

Scaling laws predict the outcome of large, expensive training runs by extrapolating trends from smaller, cheaper runs. On downstream accuracy evaluations, these predictions fail when the weaker models produced by the cheaper runs score near zero accuracy, because these scores give no trend to extrapolate to larger scales. We introduce hinted scaling laws, which use privileged information to turn a benchmark into a continuum of easier versions so that even the small models score above zero. We find that accuracy rises smoothly and predictably with the hint level, enabling new types of scaling laws that operate below the noise threshold. On the Delphi pretraining ladder, a hinted scaling law fit on models trained with at most $10^{20}$ FLOPs forecasts the GSM8K accuracy of a $10^{23}$ FLOP model (25.9%) to within 3 percentage points; a standard scaling law under-forecasts it by 13 percentage points when fit on the same models and by 12 even with models up to $10^{22}$ FLOPs in the fit. Across both this pretraining ladder and a collection of post-trained model families, hinted scaling laws reduce the median forecasting error on GSM8K, AIME and HLE from 21 percentage points to 4.3. Hinting also reduces the number of test examples needed to tell whether one model is better than another by widening the accuracy gaps between models while approximately preserving their order: a single hinted rollout per problem ranks nearly indistinguishable models as accurately as five unhinted rollouts. Our results suggest that hinting is a general primitive for improving the accuracy and cost of downstream scaling laws and other prediction tasks in model development.

Language models can analyze and assess student texts using predefined rubrics and deliver structured feedback, reducing instructors' workload. However, existing solutions often rely on external services that are resource-intensive, costly, and raise privacy concerns, making them unsuitable for constrained environments. This study evaluates an automated assessment pipeline with three key aims: design for low-resource environments, privacy, and focus on a novel context in cybersecurity education. The pipeline integrates lightweight language models for textual evaluation and a vision-language model for image interpretation. Responses from 105 students were collected over two semesters of an undergraduate course during a cybersecurity exercise on incident response. Results across different hardware configurations show that small models in low-resource environments can achieve up to 97% accuracy in text-only and up to 94% in rich-text assessments. This rivals or even outperforms a large model while adding the benefits of locally hosted assessment. Our solution can help educators and practitioners who need grading assistance without replacing human oversight, especially in environments where cloud-based solutions may be impractical. This study contributes an open-source, locally deployable AI-assisted scoring pipeline that runs on affordable hardware and can be extended to various types of student reports in computing education. As a result, it makes automated assessment accessible to a broad range of educators and institutions.

Depression affects millions of Bengali speakers across South Asia, yet automated detection systems have relied almost entirely on datasets translated from English. Such datasets cannot capture the emoticons, romanized Bengali words, and elongated spellings that characterize real expressions of distress. We introduce BanglishDepYT, a Bengali depression dataset collected natively from YouTube, comprising 191,000 unlabeled comments and 1,000 samples verified through dual bilingual annotation. We propose EPH (Emotion-Preserving Hybrid preprocessing), which converts emoticons to semantic tags, preserves Banglish tokens, protects negation words, and normalizes emotional elongation. A key finding is that preprocessing choice substantially affects performance on informal text: standard cleaning reduces the Banglish Survival Rate to zero (SVM F1: 0.7062), while EPH preserves signals (CESS: 0.9996 vs. 0.6694) and raises SVM F1 to 0.8570, an improvement of 0.1508. We further apply domain-adaptive pretraining (DAPT), continuing SahajBERT pretraining on unlabeled in-domain comments. DAPT-SahajBERT achieves the highest mean F1 (0.8161) and MCC (0.6272), significantly outperforming the baseline (F1: 0.7805, p = 0.0013, d = 1.450) and matching MuRIL (p = 0.589). These results suggest that targeted domain pretraining is a competitive alternative to large-scale multilingual pretraining. LLM-assisted annotation using Qwen2.5-7B achieves substantial human agreement (kappa = 0.611), enabling pseudo-labeling of 96,071 comments. Semi-supervised training improves AUC from 0.8499 to 0.8731 (p = 0.044), offering a scalable path to expanding annotation. All code, models, and datasets are publicly released to support reproducible research.

A rule written for people fixes its atomic conditions and connectives but leaves open how the connectives are bracketed, and each bracketing is a different logical reading that an autoformalizer must resolve or expose. We study the scope grammar, a deliberately ambiguous context-free grammar whose derivation trees are the readings of a connective skeleton and whose quotient by logical equivalence is what matters downstream. A skeleton with $k$ binary connectives has $C_k$ derivation trees, the $k$-th Catalan number, but far fewer logically distinct readings. For the positive, repetition-free fragment we make the gap exact: through the uniqueness of the read-once normal form we characterize the realizable normal forms by lowest common ancestors and count them with a two-label recurrence over the operator word, verified against exhaustive enumeration up to eleven atoms. The count is invariant under reversal and $\wedge/\vee$ duality; a skeleton is ambiguous if and only if it uses both $\wedge$ and $\vee$, a one-pass test; and the number of classes is at least $2^{\lfloor n/3\rfloor}$. Up to fifteen atoms the alternating pattern is the worst case; its count has a cubic generating function, is the 3-Schröder sequence (OEIS A137842), and grows with base $\varphi^{5/2}\approx3.330$, below the Catalan base 4. We reduce maximality to an inequality on increments, verified to sixteen atoms, and explain why local exchange arguments cannot prove it. We also treat negation, exceptions, and background axioms, where normal-form uniqueness still decides equivalence but counting is open.

Backpropagation is the standard method for computing the gradients that train neural networks, but it is usually presented in a compact vector form that skips the intermediate steps. This tutorial derives backpropagation step by step for a fully connected feedforward network with sigmoid activations and a squared-error loss. Using a running example with three layers, we first differentiate the error with respect to a single weight by following its path through the network with the chain rule. We then show how summing over all paths leads to the error terms $δ$ and to the four general backpropagation equations in matrix form. Every term is annotated with its dimensions, so each equation can be checked for consistency. The derivation is then extended to $m$ training examples, where the weight gradients become a single matrix product and the bias term corresponds to broadcasting. It is also extended to mini-batch gradient descent, including why the gradient of a mini-batch estimates the full gradient. Batch and mini-batch training are both given as algorithms. Finally, each equation is linked to the corresponding line of a NumPy implementation trained on the AHDBase dataset of handwritten Arabic digits, and the complete code is included.

LLM-based agentic systems are increasingly combined with CoT reasoning to solve tasks that require accurate reasoning, external knowledge, and multi-step deliberation. Yet, explicit CoT incurs substantial generation overhead, while implicit CoT, which replaces textual rationales with compact continuous tokens, does not naturally support the explicit communication interface required for agentic tool-use. We propose TAC-CoT, a tool-empowered adaptive implicit CoT framework that replaces explicit reasoning traces with compact latent reasoning segments while retaining textual tokens only for tool calls, observations, and final answers. This interleaved design enables agentic tool-use without verbalizing every intermediate reasoning step. Experiments on agentic math and QA benchmarks show that TAC-CoT improves accuracy by up to 8.8 points while reducing model forward passes by up to 6.8x. Furthermore, analyses show that TAC-CoT improves tool-use quality in both single- and multi-tool settings.

When user queries are incomplete, large language models should ask targeted questions before committing to an answer, a capability we call proactive elicitation, yet evaluation on real user-assistant queries across seven LLMs (8B to frontier) shows that fewer than 11% of underspecified queries trigger a follow-up question. To benchmark and train this capability, we introduce VagueBench, where each seed query is annotated with structured intent and detail units then degraded into intent-vague and fully-vague variants via controlled removal of linguistic cues, yielding known hidden information sets for evaluating both when and what to ask. We propose DialogGRPO, an offline reinforcement learning method that organizes multi-turn elicitation as a dialogue-state directed acyclic graph, where candidate questions at each node receive dense turn-level rewards combining immediate information gain with lookahead value through group-relative comparison, enabling fine-grained credit assignment without runtime simulation. DialogGRPO elicits missing information on over 88% of vague queries with substantial gains in both recovery efficiency and answer coverage, without over-asking on complete queries, outperforms prompting and test-time-scaling baselines, and generalizes zero-shot to MultiWOZ 2.4 and AmbigQA.

LLM agents learn by interacting with environments, yet existing environments are designed for specific tasks and fail to generalize to new ones. Recent methods automate environment construction with LLMs. However, they generate each environment from scratch, neither reusing parts shared across environments nor carrying a fix from one environment to the rest. To address this, we propose a new paradigm of evolving environment atoms, which composes environments from reusable atoms (intent, rule, data and server) and evolves the atoms when an environment fails to give reliable feedback (e.g., an application lacks a tool the task requires). To instantiate this paradigm, we introduce LegoEnv, which builds 5,465 verified environments over 223 applications, and one round of evolution raises the share of solvable environments from 71.6% to 88.3%. On AutomationBench, SFT on these environments lifts Qwen3.6-27B from 16.0% to 32.0%, and RL with their verifiers lifts Qwen3.8-27B from 37.6% to 40.3%. Further analyses show steady gains with corpus size, transfer to unseen applications, and failures concentrated in a few shared root causes.

FOLLIE (for First-Order Logic for Language Inference and Entailment for French) connects French sentences, compositional first-order representations and automated natural language inference (NLI). The first French NLI-ready, NL-to-FOL resource, it contains 19,983 inference records and 36,987 non-empty sentence-slot–formula mappings from six heterogeneous NLI sources, retaining French lexical material and source problem structure. Its silver annotations are produced by a neural–symbolic pipeline, with 111 separately identified LLM-assisted additions. The construction combines morphosyntactic analysis, probability-conditioned categorial multitagging, grammatical proof search, discourse representation and first-order semantic construction. Stage-wise experiments show substantial gains in derivation coverage; analyses of conjunction distributions, logical-operator incidence and scope alternation characterise the resulting resource. We then develop an automatic inference engine whose input-conditioned background theories address lexical relations, quantification, quantities, event structure and temporal dependencies. Prover9 supplies proofs for entailment and contradiction, while Mace4 supports bounded consistency checks. Three-way accuracies are 65.37% on all 335 FraCaS-FR problems, 70.33% on 300 GQNLI-FR problems and 71.35% on 9,399 eligible SICK-FR records pooled across splits. Agreement with reference labels among committed decisions is 93.53%, 100.00% and 98.66%, at commitment coverage of 41.49%, 37.00% and 15.05%, respectively. Worked constructions, axiom schemas and proof analyses explain these outcomes and identify the contributions of scope, reference, lexical knowledge and numerical structure. FOLLIE provides a French experimental basis for studying the representations and assumptions that connect sentence interpretation to deductive inference.

Web communities rely on moderators to catch harmful messages such as insults, hate speech, and scams, but no one can read every message. Existing tools give each message a risk score, which tells moderators where to look but not what to read or how to check it. Yet meaning depends on context: the same reply, "I agree.", can support or oppose removing someone from a group. Moderators therefore need source-grounded evidence, not just scores: a few key messages kept with their original wording and replies. We present CST, a lightweight triage pipeline that first decides where to look, then what to show. It scores risk by pairing character patterns, which capture exact wording, with sentence meaning from a fixed multilingual model, so it handles several languages and runs on an ordinary CPU. It then groups related messages and picks from each group in turn, so a small reading budget covers different parts of a discussion, and links every pick back to the original thread. Across six tasks from five datasets in several languages, CST ranks harmful messages more accurately than a fully fine-tuned multilingual model and, under the same budget, finds more of them than reading by score alone. In a user study, reviewers needed at least 27% less time with CST's grouped evidence while keeping decision accuracy above 94%. CST turns risk scores into evidence that moderators can read quickly and check in context. Code and dataset reconstruction resources will be released at https://github.com/Kirito14IT/CST-MIL.

Electroencephalography (EEG) provides a non-invasive means for neurological and mental health assessment, yet existing EEG models are typically developed for individual tasks and are difficult to unify across heterogeneous signal configurations and assessment objectives. Large language models (LLMs) offer a common interface for multiple tasks, but adapting heterogeneous EEG signals to a shared LLM raises a key challenge: how to preserve task-specific signal characteristics while enabling effective language-side parameter sharing. To address this problem, we propose EEG-HealthLLM, a unified and parameter-efficient framework for heterogeneous EEG health assessment. We introduce an Asymmetric Hierarchical Adaptation Strategy: task-specific EEG adapters preserve heterogeneous spatio-temporal dynamics, while a unified shared LoRA is injected into the LLM backbone (e.g., Phi-3) to capture shared high-level assessment semantics across tasks. We further establish EEG-HealthBench by reorganizing five existing EEG resources into paradigms covering epileptic seizure detection, major depressive disorder assessment, anxiety risk assessment, depressive symptom risk assessment, and emotional state assessment under a unified subject-independent protocol. Extensive experiments show that EEG-HealthLLM achieves an average balanced accuracy (BAC) of 69.58%, ranking first among the evaluated methods in both average BAC and macro-F1. Compared with independent single-task LoRA adaptation, it improves average BAC from 67.70% to 69.58% while reducing the aggregate language-side adapter footprint by 80.0%. Further analyses reveal task-dependent cross-task interactions and trade-offs, supporting the proposed hierarchical adaptation strategy. The implementation is publicly available at https://github.com/Marui-88/EEG-HealthLLM.

Accurate house price prediction is essential for real estate buyers, sellers, agents, and financial institutions. Traditional methods rely primarily on structured data, e.g., number of bedrooms, bathrooms, parkings, but they overlook rich contextual information embedded in textual descriptions and property photos. Previous deep learning approaches for house price prediction have fused multimodal data but they fall short to extract structured spatio-temporal and contextual features from texts and semantic visual information from images. In this paper, we introduce a novel multimodal framework that addresses these limitations by systematically integrating structured attributes, textual descriptions, and visual images of properties to achieve higher prediction accuracy while maintaining interpretability. We use a large language model to derive domain-specific spatio-temporal and contextual features from text, and an unsupervised vision-language approach to automatically group photos into meaningful classes like indoor, outdoor, mixed, or floorplan, that are commonly present in real-world house listings. These categories are created without manual labeling by aligning image clusters with carefully designed pseudo text prompts. We fuse all features into a unified representation and use an optimized regressor to predict prices. Experiments on a large real-world property dataset demonstrate that our method outperforms traditional methods and existing multimodal baselines. Ablation studies confirm that the novel textual feature extraction and categorized visual embeddings substantially contribute to more accurate and reliable price predictions.

Temporal Knowledge Graph Question Answering (TKGQA) requires language model agents to retrieve evidence, select relevant facts, and perform multi-step temporal reasoning. However, the trajectory-level outcome signal in GRPO does not explicitly distinguish the quality of intermediate decisions: downstream failures can penalize useful evidence acquisition, while identical rewards in all-failed groups eliminate the relative outcome signal despite differences in intermediate progress. We propose Credit Redistribution with Evidence-Driven Opportunities (CREDO), which uses intermediate evidence to recover local supervision, including in all-failed groups, and determine where outcome supervision acts. CREDO assigns positive stage-level credit through comparisons among rollouts of the same question and uses retained answer overlap as an opportunity proxy to shift part of the outcome-policy weight toward reasoning and answering. The framework preserves terminal rewards and trajectory-level advantages, while redistribution conserves total outcome-policy weight per trajectory. Following the public evaluation protocol of prior work, CREDO raises Overall accuracy from 78.0% to 82.4% and Multiple accuracy from 55.0% to 65.6% on MultiTQ relative to the strongest published baseline in our comparison. On TimelineKGQA, it raises cross-KG accuracy from 30.9% to 36.4% relative to our reproduced baseline via direct transfer to ICEWS Actor under matched training and evaluation settings.

AI reviewers are increasingly used to assess scientific manuscripts, yet it remains unclear whether their judgments reflect the improvements authors make through revision. We investigate this question using PaperTrajectory-2026, the 2026 release of PaperTrajectory, a dataset of 6,959 paper trajectories across five AI conferences, with automatic annotations of changes along 14 dimensions and separate judgments of overall quality. The annotations show that major improvements emphasize empirical evidence and framing, while later revisions tend toward incremental improvement. Using independent assessments of manuscript versions from eight AI reviewers, we examine whether annotated major improvements receive higher scores, which revision dimensions are associated with score changes, and whether user opinions can outweigh observed revision gains. Every reviewer leaves scores unchanged on most major improvements, and agreement on the direction of score changes is weak. Revision dimensions show some reviewer-specific associations with score increases, but provide limited predictive discrimination and no consistent guide across reviewers. Even when a revision earns a higher neutral score, a brief negative user opinion can erase or reverse that advantage, with substantial variation across reviewers. These findings show that substantive revisions are not consistently reflected in review scores and that observed scoring gains can be vulnerable to user framing, motivating assessments that connect manuscript evidence to overall judgments and remain stable when no scientific evidence changes.

Goal pursuit in open-ended dialogue requires agents to adapt their actions to interaction feedback across successive exchanges. We investigate whether agentic capabilities developed for multi-step tool use carry over to dialogue organized through the model's native tool interface, and how this changes interaction behavior. We define Tool-Mediated Interaction (TMI) as a training-free protocol that represents dialogue actions as tool calls and counterpart responses as tool results, maintaining a continuous reasoning-action-observation loop with retained reasoning and interaction history. We evaluate TMI across social interaction, emotional support, persuasion, and price negotiation using two models and 15,678 main-experiment episodes. Across both models, TMI improves social goal attainment, emotional-support preference, donation amounts, and buyer negotiation margins over conversational baselines with and without thinking. A Qwen ablation further shows that TMI outperforms ordinary conversation with retained reasoning on all four primary outcomes. Behavioral analyses associate the improvements with task-specific patterns in initial proposals and responses to interaction feedback. Alongside these gains, TMI reduces mean reasoning tokens per decision by 42.0-81.5% relative to the thinking baseline and lowers mean cross-turn phrase reuse across all eight model-environment pairs. These findings show that interaction organization shapes both goal attainment and reasoning across dialogue turns.

Calculating inter-annotator agreement is an established practice in NLP, and disagreement-aware work increasingly retains the distribution of human judgments rather than reducing annotations to a single gold label. Keeping the distribution, however, does not explain why annotators disagree. This paper examines the types of disagreement that occur in emotion annotation, starting from multilabel annotations of 1,677 sentence-level units from Ernest Hemingway's The Old Man and the Sea and testing the patterns they reveal against individual-level annotations from BRIGHTER and GoEmotions. Across the three datasets, we find three main recurring disagreement structures. Emotion categories overlap systematically; annotators differ both in how broadly they apply labels and in how they treat particular category boundaries, while sparse participation and explicit abstention can look like disagreement. Where intensity is rated, disagreement over whether an emotion is present can be separated from disagreement over its intensity once recognized. Similar aggregate disagreement can therefore arise from qualitatively different annotation processes. Hard aggregation removes most of this structure, while annotator vote-share representations keep marginal label frequencies but not the co-occurrence, annotator-specific, or procedural patterns that produced them. We argue that disagreement should be understood in terms of its underlying patterns before it is summarized, aggregated, or used as a modeling target, and propose a vocabulary for disagreement structure.

Verifying that a packaged food is Halal requires reading ingredient declarations and E-number additives before goods enter certified warehouses. Dual-origin additives such as E471, E422, E441 and E920 are Mushbooh (doubtful) unless a botanical, microbial or certified-Halal qualifier appears in the same ingredient clause, so a plant-based qualifier on one clause must not clear an unqualified additive in another. We call a failure to respect this scope cross-clause modifier confounding and build a benchmark of 3,199 items (277 Open Food Facts records and 2,922 template-generated clause-swap items, labelled by an explicit 50-code rule base rather than by a certification authority) with a 540-item counterfactual suite. We propose Halal-NeSyNet, which couples a DistilBERT encoder with a 16-dimensional rule-derived symbolic vector through cross-attention and gated fusion, trained with an asymmetric false-Halal risk loss. Over five seeds, text-only TF-IDF baselines reach 86.25-86.71 % accuracy with 13.42-21.73 % false-Halal leakage and 6.33-26.00 % counterfactual triad consistency, whereas Halal-NeSyNet reaches 97.04+/-0.34 % and 99.67+/-0.50 %. Gradient boosting given the same symbolic vector reaches 96.25+/-1.24 % (+0.59 points from Halal-NeSyNet, 95 % CI [-1.11, +2.28]), so most of the gain comes from clause-scoped symbolic features and not from the network, and the risk loss did not help. Holding out whole template families, text-only models fall to 61.91-78.35 %, Halal-NeSyNet to 90.87+/-11.76 % and logistic regression with the symbolic vector to 91.11+/-2.88 %, with far more spread for Halal-NeSyNet. A three-lane receiving simulation auto-clears 49.33 % of cargo at 0.76 % green-lane contamination. Because labels and features come from one rule engine and 91.3 % of the items are templated, the results measure fidelity to a rule policy, not religious certification.

LLM-based role-playing agents are fluent at imitating a character's style, yet struggle to make role-consistent decisions over long narrative horizons. Such decisions are shaped by experiences spanning many prior scenes, which are often partial and may point in different directions. Existing methods inject cross-scene memories as context, but do not reason about how past events shape a character's current motivations and behavior. We propose MemRole, a framework that makes this memory-to-decision process explicit. It introduces Long-Horizon Memory Thinking (LHMT), a reasoning schema that decomposes hidden reasoning into context understanding, memory activation, and response strategy planning, trained on a novel-derived dataset whose character-grounded, timeline-constrained memory pools limit each character to what it could plausibly know at the current narrative point. Beyond supervised fine-tuning, we propose Self-Refining On-Policy Self-Distillation (SR-OPSD): the model critiques its own rollouts against the reference under a data-driven rubric, and a critique-conditioned teacher view of the same model distills token-level guidance back into the policy. MemRole-32B improves over its backbone by 15.2 points on memory-grounded role-playing, performs on par with frontier LLMs such as GPT-5.4, and achieves the best character-motivation recognition accuracy on CroSS-MR.

A language model that learns a new fact by fine-tuning can repeat it but cannot reason with it: taught that a person is a citizen of a new country, it still cannot say, without chain-of-thought, who the head of state of that person's country is. This two-hop curse is widely treated as a reasoning deficit, and the strongest knowledge editors answer it with synthesized reasoning data. We show that the reasoning is intact and the writing is wrong. Fine-tuning writes a new fact as an answer to be said, whereas the second hop reads the bridge ("that person's country") as an entity, from the middle layers on: patching the model's own representation of the entity into the bridge at those layers largely restores the second hop, yet after fine-tuning the bridge there is no closer to the new entity than before the edit. We therefore propose read-equivalent writing (Legible): one loss term makes the bridge phrase, at the layer where the second hop finishes reading it, match the model's own reading of the new entity, and keeps the same relation's other subjects unchanged. Phrases and targets come from the edited model itself, and no multi-hop training data is used. Across Qwen2.5, Llama-3 and Qwen3, Legible (I) composes learned facts more often than the model composes its own knowledge, on MQuAKE's two-hop-curse questions; (II) reaches 1.8-2.6x the bridge accuracy of CaKE, whose reasoning data composes no better than plain fine-tuning, and has the best average over composition, accuracy and locality against four families of writers; and (III) lifts the curse on RippleEdits, on the benchmark of Balesni et al., who named the curse, and at 14B.

Anticipating financial outcomes requires synthesizing evolving evidence and reasoning under uncertainty, making deep research systems promising tools for forecasting. Yet evaluating forecasting ability is challenging because historical backtests may reward memorized outcomes or information retrieved after the forecast date. Live forecasting avoids direct access to realized outcomes, but feedback arrives only after events resolve. To obtain timely feedback and assess historical scores, we introduce FinForesight, a diagnostic backtesting benchmark for financial deep research systems. Across the Chinese, U.S., and Hong Kong markets, the benchmark evaluates four forecasting tasks: valuation, return ranking, event direction, and future financial fundamentals. Two capability diagnostic tasks assess counterfactual reasoning and financial calculation, scored separately from forecasting performance. Our harness-compatible framework restricts retrieved evidence to information available at the forecast date. We compare pretraining-era, resolved post-cutoff, and live evaluations to test whether controlled backtests approximate future performance. Pretraining-era scores exceed post-cutoff scores in most temporal comparisons, and stronger retrieval controls narrow the gap between post-cutoff backtests and live results. Forecasting attribution and robustness analyses reveal how scores respond to retrieval policy and how predictions respond to prompt rewrites and numerical perturbations. The code and benchmark data are publicly available.

Medical language models promise expert clinical support, but the strongest open-source LLMs remain too large for private edge-device deployment, while compact medical models still trail substantially on knowledge-intensive and real-world healthcare tasks. Reasoning models that fit edge memory budgets can produce long reasoning traces, increasing sequential decoding, end-to-end latency, energy consumption, and KV-cache memory pressure. We present MedPsy, a family of text-only 1.7B and 4B medical language models designed for edge deployment, targeting this accuracy-generation-efficiency trade-off. Our recipe combines a synthetic medical data pipeline over biology, medicine, and health seeds with chain-of-thought targets from a 235B medical-focused LLM teacher; a four-stage post-training curriculum comprising two SFT stages followed by two RL stages with hard-sample mining; and an edge deployment evaluation spanning multiple GGUF quantization variants and real-device inference on laptop and smartphone hardware. On seven closed-ended medical benchmarks, MedPsy-4B surpasses MedGemma-1.5-4B by 19.34 points and matches MedGemma-27B while being 6.75x smaller; MedPsy-1.7B outperforms MedGemma-1.5-4B by 11.42 points despite being less than half its size. On HealthBench-Hard, MedPsy-4B surpasses MedGemma-27B by 15.33 points, while MedPsy-1.7B surpasses it by 11.66 points despite being ~16x smaller. Beyond accuracy, MedPsy reduces average response length by 1.7x (1.7B) and 3.2x (4B) versus its Qwen3 backbones, while Q4_K_M quantization preserves the aggregate score within 1 point of BF16 with ~69% lower disk footprint. In real-device inference, MedPsy-4B's shorter generations translate into 75% lower median end-to-end latency and estimated energy per answer than its Qwen3 backbone at matched quantization, while peak RAM remains below 5 GB across our tested laptop and smartphone configurations.

Self-evolving agents accumulate experience from task streams, but existing frameworks typically rely on fixed pipelines to retrieve and utilize this experience, neglecting when and how to deploy it. We introduce Online Meta-Control of Experience (OMCE), a framework that updates experience memory and learns a meta-control policy from task feedback during test-time adaptation. OMCE employs a lightweight contextual-bandit meta-controller that uses compact retrieval signals to select among four cognitive actions: DIRECT, EPISODE, GUIDE, and DISTILL. This routing mechanism operationalizes three coupled decisions: whether to invoke experience, which memory format to utilize, and when to consolidate new guidance. OMCE evolves online, leveraging task feedback to update the memory bank and refine the routing policy via contextual bandit updates. Experiments on four online adaptation benchmarks across four backbones show that OMCE improves first-attempt success by up to 20.5 percentage points over the strongest compared baseline, while reducing per-task token use by a factor of 5.7 to 62.5 relative to ACE with GPT-5.2. Controlled ablations further support the benefits of learned routing and continued routing policy adaptation.

Large language models (LLMs) read knowledge-graph facts ranked by relevance, leaving open whether they suffice or conflict. Our checks use relation knowledge, written by hand or learned from training questions labelled with their relation paths. It selects a compact context of facts yielding an answer, and derives it; a derivability check tests whether an answer follows, a consistency check whether facts conflict, and a validation-chosen routing gate decides where the derived answer replaces the LLM's. Seven open LLMs read the compact context better in exact match than relevance-ranked facts; on ComplexWebQuestions (CWQ) they match SubgraphRAG's facts at a quarter of the prompt tokens, and stronger LLMs read SubgraphRAG's better. Where the routing gate passes, ours leads SubgraphRAG's in exact match in all 53 LLM, benchmark and prompt settings. A precision gate, returning the derived answer only when the check passes with exactly one derived entity, answered 18 and 5 percent of held-out WebQSP and CWQ questions in a pre-specified test, right for 90 and 96 percent. The consistency check judges facts by path evidence only where another path joins their entities, true of 2.5 percent of named-entity facts in our compact WebQSP and CWQ contexts.

Representation steering methods like difference-in-means (DIM) offer a lightweight approach to controlling LLM behavior, but their effectiveness varies widely across concepts, datasets, and layers, making principled deployment costly without exhaustive hyperparameter sweeps. We build on the discriminability index $d'$, a compute-cheap geometric measure of how well contrastive training activations separate along the steering direction. Across five behaviors on 3 models, we demonstrate uses beyond its established role as a dataset-level steerability predictor. First, in the multiple-choice question answering (MCQA) setting, we show that $d'$ reliably identifies performant layers at which to steer, eliminating the need for a layer-wise sweep. Second, we extend this analysis to open-ended multi-token generation, and find that $d'$ remains successful for layer selection across different settings. Lastly, we test whether the position of a generated token's representation on the DIM line predicts steering success on that individual example. In the MCQA setting, this works for high-$d'$ datasets but is inconsistent across behaviors, models, and training-token choices in open-ended generation. This highlights that methods developed and conclusions drawn in single-token MCQA settings do not necessarily transfer to open-ended generation.

The widespread adoption of large language models (LLMs) is reshaping how textual content is produced, complicating text provenance, and raising concerns for misinformation, plagiarism, and intellectual property. While prior work on AI-generated text detection has progressed, one-dimensional human-AI co-authorship scores lack signals directly relevant to ``AI involvement''. It remains unclear how adversarial paraphrasing distorts detection signals. In this work, we formulate a discrete-class text provenance detection framework. This framework allows an abstraction of realistic co-authorship scenarios into distinct degrees of AI involvement, differentiating the AI polishing on the surface level from the AI updating the discourse structures. This framework reveals the mechanisms of how adversarial paraphrasing blurs the decision boundaries for co-authorship detectors. To improve robustness of the detectors, we augment training data with adversarial paraphrases generated by the state-of-the-art paraphrasing attack model, exposing detectors to challenging meaning-preserving rewrites. Extensive experiments across multiple datasets and generation models demonstrate that adversarial paraphrasing substantially degrades detectors trained only on benign data, while adversarial data augmentation leads to consistent robustness gains and good cross-generator transferability. Overall, our findings highlight the value of discrete-class formulation and principled data augmentation for building reliable detectors of human-AI co-authorship.

Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.

Concerns about LLM homogeneity are typically quantified with embedding similarity, which reduces two texts to a single score. We revisit homogeneity with 25 interpretable metrics spanning five dimensions: semantic, lexical, syntactic, structural, and stylistic & presentational. A human study shows that perceived similarity depends on both content (what responses say) and form (how they say it), yet embedding similarity correlates less with human meaning judgments than sentence-level matching ($ρ=0.417$ vs. $0.606$) and not significantly with form judgments ($ρ=0.140$). Across 18 contemporary LLMs, models say similar things in different words: they share global semantics and broad linguistic profiles but remain distinct in lexical characteristics, structural patterns, and interactional conventions. Across releases from 2024 to 2026, models become more like themselves but not more like one another: among significantly changing metrics, intra-model distances decrease for 17 of 22, and inter-model distances increase for 14 of 21. Across two post-training pipelines, model-specific conventions emerge during SFT and are selectively reshaped by preference optimization. These results suggest that homogeneity is multidimensional and evolving, and that embedding similarity alone can overstate cross-model homogeneity while concealing diverging intra- and inter-model trends.

Humans continuously acquire knowledge from carefully crafted materials such as books and tutorial videos, carrying that understanding into future tasks. Millions of such instructional videos are uploaded online every day, yet LLM agents remain unable to exploit this ever-expanding wealth of human teaching. Their capabilities are frozen after training, while parameter updates at scale remain prohibitively expensive. This raises the central question: can LLM agents learn from this vast instructional resource without parameter updates? We introduce Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that consolidates tutorial videos into a persistent, structured tutorialbook memory through a two-stage adversarial process that harvests hard, video-grounded questions and uses them to refine the memory. At inference time, TAL retrieves through a Table-of-Contents (TOC) tree, tying retrieval directly to the model's reasoning rather than to embedding similarity. To study this setting, we introduce VidLearn, a multimodal QA benchmark built from 72 chess tutorial videos spanning 12 openings. On VidLearn, TAL outperforms all evaluated baselines and improves monotonically as more videos are consumed, achieving up to 113% and 91% relative gains over the corresponding vanilla MLLMs on the training and test splits, respectively. In actual gameplay, TAL reaches a stronger opening-phase position than its vanilla counterpart in 69% of head-to-head games. Evaluation in a second domain further demonstrates the generality of TAL. Together, these results highlight instructional videos as a scalable resource for parameter-update-free agent continual learning.

While online medical consultation broadens access, unprompted patient disclosures and topic shifts frequently disrupt clinician inquiry, leaving clinical states progressively revealed. We characterize these dynamics using both existing evaluation metrics and task-specific interaction metrics designed in this work. Through an empirical study across 16 medical dialogue datasets, we show that authentic online interactions exhibit substantially lower contextual continuity and higher unpredictability than offline and simulated benchmarks. We propose PRISM (Policy Reasoning with Inferred States for Medicine), a framework inspired by partially observable Markov decision processes that decouples latent state estimation from action selection to address drifting dialogue trajectories. Specifically, PRISM factorizes clinical triage into belief estimation, policy optimization, and response realization. By maintaining an evolving posterior belief over consultation states and conditioning action selection on the maximum a posteriori state, PRISM advances clinical goals while accommodating patient-initiated shifts. Extensive evaluations show that PRISM substantially outperforms frontier LLMs and clinical agents, achieving a 22.40 percentage point gain in clinical decision accuracy alongside decisive improvements in response generation, surpassing the strongest baseline by 5.51 points in BERTScore_F1. Furthermore, ablation studies show that providing oracle consultation states consistently boosts response quality across all evaluated baselines by guiding their action reasoning, yielding average gains of 5.58 points in BERTScore_F1.

Memory systems for LLM agents are evaluated mostly on social dialogue, where facts accumulate and rarely change, so remembering reduces largely to finding the turn in which something was said. Financial dialogue breaks this assumption: figures are revised, stances shift without being restated, rules stated once govern later questions, and a single company is discussed under several analytical lenses. Memory systems that retrieve the passages most similar to a question, or the most recent ones, therefore return superseded values, miss rules that share no wording with the question, and cannot recover a stance spread across sessions. We present TrackMem, a per-speaker graph memory that tracks the history of figures, separates perspectives by entity and lens, and preserves each speaker's actions in chronological order. To evaluate it, we introduce FinLoCoMo, a benchmark of 15 long-horizon investment conversations with 765 questions in two splits: one following the LoCoMo categories and one targeting four failure modes. With two readers, TrackMem outperforms existing memory systems on both splits. On the failure-mode split it reaches 73.9% macro accuracy against 69.4% for the strongest baseline, with the largest gains on questions about revised figures (86.5% against 73.4%).

Despite growing interest in Arabic natural language processing, curriculum-aligned educational resources remain scarce compared with English, and few public datasets for Arabic curriculum content with structured, generation-ready annotations are available, particularly for primary-school science. We introduce ArSciKids, a collection of two Arabic datasets designed for curriculum-aligned educational content generation: (1) a scientific-experiments dataset containing 999 manually curated records from official primary-school science textbooks in Saudi Arabia, Syria, and Jordan, and (2) an assessment-question dataset containing 2,167 multiple-choice questions collected from the Saudi Manhaji educational platform. The experiment records are structured with titles, objectives, tools, procedures, observations, conclusions, keywords, and grade levels, while the assessment records contain lesson references, questions, four answer options, correct answers, and explanations. Both datasets cover grades 1-6 and consist of real-life Arabic educational content rather than machine-translated or synthetically generated material. We assess the usefulness of these resources for generation by fine-tuning ten open models spanning sub-billion to several-billion parameter scales using LoRA. Models are evaluated using automatic metrics, including ROUGE and BLEU, and answer accuracy for multiple-choice generation. Moreover, we evaluate the results using a five-criterion human evaluation protocol. The results show substantial differences across model scales and tasks. ArSciKids will be publicly available, and we hope that it provides a manually curated, structured foundation for research on Arabic educational NLP, curriculum-aware generation, and low-resource language modeling.

Cross-country sentiment analysis faces three fundamental challenges: language-specific biases distort comparisons, compositional data constraints invalidate standard statistics and temporal dynamics lack rigorous testing frameworks. We introduce sentiment sensing, a framework that addresses these gaps through integrated statistical methods. By combining compositional data analysis, functional data methods, and non-parametric inference, sentiment sensing enables valid multilingual comparisons and statistically defensible detection of sentiment shifts. Our framework has been tested in two case studies, analyzing news articles from the 27 countries of the European Union on the topics of energy and artificial intelligence. Sentiment sensing focuses on multilingual sentiment analysis enabling rigorous comparative science.

Visual-text compression renders long contexts as images and is engineered around how much of the text a vision-language model (VLM) can recover. Yet recovered text is not usable text. VLMs can transcribe the text in an image word for word and still fail to answer from it; given back their own transcript, they answer nearly as well as from the original text. We call this difference the read-use gap. Recognizing and using visual text have different resolution requirements: Qwen3-VL-8B's recognition saturates at 15-pixel characters while use catches up only at 24 pixels, and scaling from 2B to 32B raises recognition without narrowing the gap. Putting the evidence into words closes it: thinking VLMs recite the evidence in their thinking, and their own transcript, written without seeing the question, recovers most of what thinking adds. In a preregistered held-out test, the gap ranks how much eight VLMs gain from their own words over pixels (Spearman $ρ$ = 0.86), which recognition rates do not. We propose ALOUD, which spends visual resolution on finding and recognizing evidence and language tokens on using it. Without keeping the source text, it doubles the accuracy of a frozen Qwen3-VL-8B on long documents compressed 5x and 10x: at 5x, from 27.7% to 54.8%, near the original text's 58.5%, with a context smaller than the compressed pages themselves. For less than the cost of reading the original text, it leads at equal budget every method that does without it, at every level, and beats AGAR and VERA, which paste the original text back; at 5x it also beats LensVLM-9B, which is trained and keeps the original text. Visual-text compression is limited not by how much text a model can recover, but by how much it can use.

Simulated students offer a scalable way to test learning platforms and AI tutors before deployment with real students and to augment scarce data for studying learning at scale. Existing LLM-based simulators primarily target short, language-centered interactions. In open-ended learning environments, however, student behavior emerges from patterns across long-horizon interaction trajectories, where turn-level prompting can drift from an assigned behavior profile. We introduce ALPS (Align Long-term by Planning Short-term), a hierarchical simulated student conditioned on observable, environment-specific behavior profiles. Rather than selecting every action independently, ALPS alternates between planning profile-aligned task segments and generating the corresponding actions. We evaluate ALPS in both data-sufficient and data-scarce settings, where supervision comes either from real student trajectories or from a teacher LLM. Across both settings, we compare in-context learning (ICL), supervised fine-tuning (SFT), reward-ranked fine-tuning (RAFT), and their sequential combination (SFT followed by RAFT). Experiments in an inquiry-based physics game and a pharmacy simulation evaluate behavioral density and coverage against held-out real data, and profile consistency using a human-aligned LLM judge. Across both settings, ALPS consistently outperforms turn-level generation, achieving comparable density and judge pass rates under scarce supervision, although coverage remains lower. These results show that hierarchical planning enables faithful, profile-aligned simulation over long interaction horizons, even with limited real student data.

Voice conversion (VC) models can be used to anonymize speech data while preserving naturalness. Existing models, however, are mainly optimized for English and other high-resource languages, and they fail to generalize well to Arabic due to its distinct phonology. In this work, we comprehensively evaluate existing zero-shot VC systems for Arabic speech, identifying limitations in pronunciation preservation. We then train an Arabic-centric model based on EZ-VC on diverse Arabic speech, and demonstrate strong performance in intelligibility, naturalness, and pronunciation preservation, while maintaining high speaker dissimilarity. Through subjective evaluation, we further demonstrate its robustness to code-switching and cross-dialectal conversion, as well as its generalization to Arabic dialects with limited representation in the training data. Our findings highlight the limitations of zero-shot VC and the importance of target-language adaptation for robust voice anonymization.

Large language model (LLM)-based multi-agent systems (MAS) tackle challenging tasks by dividing work among agents and coordinating their contributions toward shared objectives. Evaluating these systems therefore requires benchmarks that assess not only task-solving performance but also the ability to handle interdependent work. However, existing benchmarks often emphasize task difficulty without substantial collaboration requirements or evaluate only a limited range of predefined collaboration patterns. In this paper, we introduce MAC-Bench, a comprehensive benchmark for evaluating multi-agent collaboration through interconnected tasks. Rather than imposing predefined agent roles or communication protocols, MAC-Bench embeds collaboration requirements directly into task dependencies and shared-resource constraints. Specifically, we mine reusable collaboration structures from selected benchmark tasks at two levels. Local collaboration patterns capture recurring interactions among subtasks, while global task structures characterize how these interactions contribute to complete workflows. Through coarse-to-fine composition, these structures are recombined into task graphs combining different collaboration requirements. Tasks are instantiated in executable environments with automated verification of partial progress and complete task resolution.

Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions). Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. We evaluate Zep Agent Skills on $τ^3$-bench banking, AppWorld, and EnterpriseOps-Gym. Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. Most of the gain is on the tasks that the teacher solved. On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.

Financial restatements correct errors in previously published financial reports. Could managers' language in earnings calls warn of them in advance? Comparing 109 calls that precede restatements with 218 matched control calls, we find that the text predicts restatements but adds no detectable ranking gain beyond public firm information. A text classifier beats chance, largely because of restatements tied to special-purpose acquisition companies, but a model using only public information scores slightly higher. With refitting, the 95% interval for the gain from adding text (-0.05 to 0.11 AUROC) still leaves room for a meaningful benefit. A large language model scores higher, but we show that it names the exact ticker in 66% of masked transcripts, and it was released after the restatements occurred. Our results suggest that finding new warning signs in call text requires a public-information baseline and a check of what a language model already knows.

Surprisal correlates closely with reading time and is widely used in the cognitive science of reading as a measure of processing difficulty. Classical readability formulae such as the Flesch-Kincaid Grade Level, which scores a text by the US school grade expected to read it, are the most widely used practical measures of text difficulty, but their coefficients are unitless regression constants and have never been related to surprisal. We theoretically show that these coefficients can be interpreted as surprisal: every formula of the classical additive shape is a sum of per-word and per-sentence cross-entropies under a memoryless model, and once that model is tied to the language through Zipf's law and the formula is read as its linearisation at a reference text, each coefficient factorises into a scale in grades per nat, a Zipf slope and a reference sentence length. A grade thereby becomes a quantity in surprisal. Because the formula outputs grades and children's vocabulary has been measured by grade, the interpretation recovers the vocabulary growth the formula implicitly assumes and lets it be checked against development: holding sentence length fixed, the word coefficient of the Flesch-Kincaid formula converts to 0.18 nat of unigram surprisal per word per grade (0.15 to 0.19 across the regression-derived formulae), and normative root-word vocabulary growth (Biemiller and Slonim, 2001), converted through Zipf's law, gives 0.16 nat per year at the frontier of the known vocabulary. The word coefficient is one year of vocabulary growth, and the grade is a vocabulary scale. The identification also yields a diagnostic (the sentence-length terms are not those of a memoryless surprisal) and extends to a Japanese formula.

In this work, we present a case study that explores various tasks centered around the topic of migration in Slovak, a low-resource language, such as topic relevance and geographical relevance classification, and migration source/destination location term extraction. Our results demonstrate that native (Slovak) prompts yield a modest, task-dependent gains, while large models show significant robustness to prompt variations compared to their smaller counterparts. Analysis reveals that instructions (system or task) emerge as the most critical prompt component, more so than the examples sections, with task-specific performance benefits being more pronounced than overall language effects.

Spoken medical question answering can fail when automatic speech recognition (ASR) corrupts a clinically important term before retrieval. We present MedAstra, a framework that maps frozen WavLM speech representations through a lightweight adapter into a frozen E5 text embedding space. This enables a dense retrieval path that accepts a waveform without a query transcript. Retrieved passages and speech-linked textual metadata support answer generation with entity alternatives, diagnostic contrasts, relation heuristics, and bounded retrieval refinement. ASR-less refers to the initial speech retrieval interface; lexical retrieval, string-based entity correction, and the text generator require textual inputs. The deterministic speech embedding does not constitute a calibrated model of acoustic uncertainty. Reported retrieval results reach 87.6 Recall@10 for the speech-only retriever and 90.4 for the hybrid configuration, compared with 81.0 for transcript-first dense retrieval. Medical QA and ablations provide descriptive system comparisons, with input-provenance and evaluation limitations stated explicitly. We distinguish controlled synthetic acoustic tests from real clinical speech, qualify component attribution, and specify the remaining comparisons needed to assess speech-only QA and clinical generalization.

Advances in large language models (LLMs) have created opportunities for cross-lingual transfer of medical knowledge from high-resource to low-resource languages to support access to healthcare information. However, medical question-answering (QA) datasets for fine-tuning and evaluating LLMs remain scarce for Tetun. To address this gap, we translated MedAlpaca's Medical Flashcards and WikiDoc Patient Information collections into Tetun using Gemini 2.5 Flash, producing 39,077 medical QA pairs. Translation quality was assessed on 50 randomly sampled QA pairs (100 documents) using automatic metrics against human reference translations and evaluations by two native Tetun-speaking healthcare professionals, focusing on medical accuracy and Tetun fluency. Human translations achieved higher mean ratings than LLM-generated translations for both criteria (4.43/5 vs. 4.02/5), with statistically significant differences (Wilcoxon signed-rank test, p < 0.001). To assess downstream utility, we fine-tuned Gemma and Llama variants on the translated MedAlpaca dataset using parameter-efficient fine-tuning (PEFT) and evaluated them on 325 Tetun-translated United States Medical Licensing Examination multiple-choice questions. Gemma 4 12B fine-tuned using QLoRA achieved relative improvements in mean accuracy of 10.79% with 4-bit NormalFloat (NF4) inference and 10.42% after merging the LoRA adapters into the BF16 base model, compared with the corresponding base models. These results highlight the viability of automatic dataset translation for Tetun medical QA and the effectiveness of PEFT in enhancing model performance. The datasets, models, and evaluation resources are accessible at https://github.com/labadain/labadain-medqa to support reproducible research on Tetun medical QA and facilitate machine translation, cross lingual, and multilingual research.

Pretrained models are typically viewed as generalists, requiring bespoke post-training to achieve task-specific expertise. Gan & Isola (2026) recently proposed a striking alternative: that pretrained models lie in a dense neural thicket of task specialists, which can be surfaced with simple Gaussian weight perturbations. We examine this neural thicket of task-specific experts and identify several key findings that question this perspective. First, we find that the apparent lift of the best-of-N selected perturbation decays quickly with selection-set size m, strongly suggesting overfitting to the selection questions. In fact, for 10 of the 12 model-task pairs we study, a simple null model of pure selection noise predicts more than 75 percent of the top expert's reward lift. Second, the top experts' advantage often vanishes on held-out questions or remains within the run-to-run variability of the unperturbed pretrained model. Third, we find that majority voting over the top-K selected perturbations, termed RandOpt(N, K) by Gan & Isola (2026), is often matched or outperformed by majority voting over K samples from the base model. Overall, our results suggest that random weight perturbations may be ineffective at uncovering task experts around pretrained models, and that task adaptation requires more sophistication than random search. Code and artifacts are available at github.com/asil012/RevisitingNeuralThickets.

Character names in large language model (LLM) fiction concentrate on a handful of recurrent names such as 'Elara' and 'Elias', a measurable face of AI slop. Existing fixes either require logit access or retraining, or operate per request and cannot control concentration across a corpus. How much prompt-level strategies reduce concentration, and at what cost to name naturalness, has not been systematically measured. We present an empirical study of name-concentration control across eight strategies and introduce speculative naming, a fine-tuning-free method that adapts the draft-verify-reject paradigm of speculative decoding from tokens to names. The method requests candidate name-sets through verbalized sampling, scores them against a SlopTracker that combines a per-model fingerprint with a runtime counter, and selects through a temperature-controlled rule or its strict variant (HARD). We use 20 LLMs to generate 44,782 stories in four languages from text and visual stimuli (88,919 named characters), and measure concentration with a list-free collision metric (Slop Ratio, SR) decomposed into known and emergent components. List-based strategies do not remove slop: static banlists displace concentration to emergent names and more than halve the share of attested real names. Speculative naming removes it: the sampling scaffold alone cuts SR by 89% on average, the dynamic tracker brings the total reduction to 95%, mean collision probability falls from 7.3% to 0.4%, and known-name collisions hold at zero while naturalness stays close to baseline (42-69% attested by language) and within-story collisions drop from 4.3% to 0.2%. The strategy ranking is stable across languages, model families, and name-counting conventions. Name concentration is thus controllable at the corpus level without logit access or retraining; broader creative diversity beyond naming remains open.

Applications use Typed Decision Models to route and moderate text in many languages, and they often act only on decisions above a confidence threshold. We introduce XDecideBench, a benchmark of nine suites in 122 languages with 88,425 texts and 153,875 typed questions. All suites share one scoring contract, and we report aligned, locally labelled and native items separately. Model rankings change from suite to suite. The model that is best in English, chosen in hindsight, loses significantly in about a quarter of the suite-language cells, and on Burmese Belebele it falls far behind the best model. Three open models fine-tuned for typed decisions beat the Qwen models they were built from in all twelve development languages. When the Qwen models are run locally, a pre-registered held-out cohort confirms that the accuracy gain is larger in the four languages with the least web text than in the six with the most. In both sets of languages, however, fine-tuning fixes a similar share of the base model's errors, so the larger gain may reflect weaker base models. A confidence threshold set on English data to allow 5% errors exceeds that error rate on XNLI, even in German and French.

Language models are no longer used only to generate text. Software also asks them for decisions, in routing, guardrails, and automated gates. System One decision models, also called typed decision models, such as Jev and Laya, return one fast decision over the legal options, leaving no text to inspect. However, a fast, well-formed decision does not show that the model used the evidence. Cognitive experiments can test this, because they vary task demands and measure how behavior changes, yet existing cognitive benchmarks evaluate language and multimodal models. We introduce S1CogBench to ask what these models can and cannot do on 24 cognitive tasks, evaluating five open models and Jev and keeping conditions and scoring apart from disclosed adaptations. We find that decision models vary widely in reliability, acting well does not imply predicting humans, aggregate scores hide how decisions respond to conditions, and speed comes mainly from answering without generation. To our knowledge, S1CogBench is the first benchmark of System One decision models on established cognitive experiments, and it gives a shared test of whether their decisions respond to evidence, feedback, memory load, and future opportunities before software relies on them.

Early-exit language models decide for each token how much computation to spend, usually by stopping once the prediction is confident. Yet a token is not a fixed unit: tokenizers cut words into different numbers of pieces in different languages. We ask where inside a word such a policy spends its budget. In Qwen3, along two adaptive axes (the timesteps of a spiking conversion and the layers of the original Transformer), confidence stopping gives the continuation pieces of a word less computation than its first piece in English, German, and Turkish, but more in Bangla. Segmentation drives this. Re-segmenting the same parallel passages with BPE-dropout shifts the gap by up to eight exits on both axes, small models trained under two tokenizers shift Bangla's gap by six exits, and BLOOM and mGPT, which split Bangla less finely, give it the English pattern. Characters per token do not predict the gap across languages; within-word surprisal explains 75% of its variance. A learned stopping head beats confidence stopping by 7 to 8 points, but gains less than one point over the best mix of fixed exits at equal cost, and sometimes loses. All results come from offline analysis, not measured speedups.

Gated knowledge distillation (KD) weights teacher signals per token or example, aiming to trust an imperfect teacher only where it seems reliable. We test two gates: EWAD (entropy-weighted agreement-aware distillation), which gates tokens by teacher confidence, and CHAD (counterfactual harm-aware distillation), which gates examples by gradient alignment with validation data. Across six Bangla-centric tasks, three student models (60M to about 2B parameters), and two teacher families, no gate significantly improves plain KD in 45 comparisons. KD also fails to beat cross-entropy (CE) fine-tuning on gold outputs in 40 single-seed comparisons or three-seed runs on four tasks; CE leads on every summarisation metric. The main benefit of KD is regularisation: over ten fixed epochs, CE students lose 0.030 to 0.044 token-F1 from their early-stopped checkpoint, versus 0.005 to 0.025 for distilled students. Teacher-free label smoothing closes most of this gap. In one comparison, a teacher 0.29 token-F1 worse yields students at most 0.009 worse. Thus, gating may change regularisation strength but has little teacher knowledge to select. Findings are limited to Bangla-centric tasks and small students. We will release code, data splits, predictions, and scores.

Depth pruning removes whole Transformer blocks to make a language model smaller and faster, but the damage is not spread evenly across what the model can do, and a short fine-tuning run, called healing, decides what comes back. We measure this for small models used as agents. Four instruction-tuned models of 1.2-1.7B parameters (Llama-3.2-1B, Qwen2.5-1.5B, SmolLM2-1.7B and OLMo-2-1B) are pruned and healed, and six capabilities are tracked: structured output, function calling, instruction following, mathematical reasoning, safety refusal and knowledge. Three results hold across the families. First, function calling and math collapse in every model (retention 0.00-0.07 of the unpruned score), and the order in which capabilities break agrees closely across the four (Kendall's W = 0.92; 0.62 after correcting for chance floors). Second, which parameters are trained matters as much as which data: fully tuning the last three blocks instead of adding low-rank adapters to all blocks costs up to 0.18 in capped mean retention on three families, and one 300k-token pilot per scope detects this. Third, we test a planner that allocates a 1M-token healing budget from 21 short pilot heals. At a fixed budget it raises the weakest capability on two families, although only one of these gains survives a test that also resamples training seeds. Once its 3.48M pilot tokens are counted, uniform healing with the same total number of training tokens matches or beats every planned variant on every family and retains math significantly better on three of them in an item-level test (on one when seeds are also resampled). Healing longer is a strong and simple baseline that planned recovery must beat. Code, per-item predictions and every result are public.

Bock's algorithm provides a matrix-based alternative to Chu–Liu/Edmonds (CL/E) for exact dependency parsing. We present a reformulation that retains original node and arc identities through component labels, column adjustments, and star-and-bar transfers. In exact arithmetic, we prove operational equivalence to original Bock under matching scan and tie rules, and establish correspondence with corrected Tarjan under a prescribed component schedule. A separate analysis establishes optimal-value equivalence with Dozat-derived CL/E under the stated score and single-root conventions. We integrate reformulated Bock into a biaffine dependency parser and evaluate decoding and structured training on the Penn Treebank and Chinese Treebank. These results establish Bock as an alternative exact oracle for dependency parsing and structured training, while distinguishing equality of optimal values from identity of returned trees and training trajectories.

LLM judges often check, against an amendment record, whether the documents an answer cites are still in force. Such records are rarely complete. From a table of replacements, our test generates matched answers citing a replaced document, its successor, the replaced document before its replacement, and an older document that a newer one updated without replacing; it removes links from the record and scores each finding by its stated reason. On 48 replacements from the IETF RFC index, judges given a correct record caught nearly every replaced citation, as a lookup would. On the 24 faults whose link we removed, every judge caught more from the texts alone than with the record: Gemini 3.1 Pro 23/24 against 1/24, GPT-5.5 23/24 against 9/24, Claude Opus 4.5 17/24 against 9/24 and Gemma 4 26B 13/24 against 0/24. The drop recurred on a second draw of removed links. Most misses carried no finding; in an exploratory run with a conflict code, only Gemini used it often. Rewording the gap as "no replacement recorded" did not help; presenting the record as an unverified note, or warning that it was incomplete, restored most lost catches for Gemini, GPT-5.5 and Claude (significant after correction for the first two) but at most 5 of 13 for Gemma; where catches returned, so did false alarms on updates. With both texts, judges mostly told replacements from updates; every update flagged had a near-identical title. A descriptive case study on Indian rooftop-solar rules is appended; we release the harness.

Structured pruning reduces the size of large language models, often at the cost of lower accuracy. Whether this decline reflects diminished reasoning capacity or greater difficulty in sampling correct solutions is less clear. We examine this question across expert, FFN and layer pruning on four mathematical and coding benchmarks, comparing original and pruned models through pass@k at matched sampling budgets. Even after substantial pruning, some models match or exceed the original model's reasoning coverage despite lower pass@1, and solve problems left unsolved in the original samples. Retaining broad coverage, however, does not mean that correct solutions remain easy to obtain. Under severe pruning, diverse generated chain-of-thought (CoT) traces can coexist with increased repetition and less reliable generation. As correct answers become harder to sample, the additional attempts can sharply increase total token use even when individual responses become shorter. This separation between coverage and sampling efficiency also clarifies expert-selection gains, as methods with markedly different pass@1 can solve the same problems at large budgets yet require different numbers of attempts to reach a common coverage target. Pruning should therefore be assessed by both the range of problems a model can still solve and the sampling cost of obtaining correct solutions.

Post-trained language models fabricate citations at high rates. We ask whether that is reward hacking: scoring well on a proxy for scholarship while doing worse on what it stands for. We ran four pre-registered experiments on 900 frozen questions (US federal case law, computationally adjacent scientific literature, and medical literature) across twelve open-weights checkpoints covering four post-training stages of three model families, then refit on fifteen. The clearest result holds the weights fixed and changes only the prompt. Asking for citations raises fabrication sharply in post-trained checkpoints and not in base ones: a 47.8-point interaction on twelve checkpoints (Holm-adjusted p = 2e-34) and 41.5 points on fifteen, in all three domains. A length-matched control separates the citation cue from a demand for a longer answer, so the effect is specific to citations, not verbosity. Post-training also raises fabrication overall, from 4.8% to 56.7%, but our registered prediction that it would rise most where fabrication is most tempting did not hold. A preference model trained only on surface citation features prefers post-trained output 86% of the time, and holding citation count fixed shows it rewards quantity. Our answer: possibly, but not proven. Post-trained models produce fabricated citations at far higher rates when asked for sources, and a naive evaluator that counts sources without checking them would prefer their output for citing more. But the targeted pattern we registered did not appear, and plain instruction-following could explain the same results. Proving that the reward causes fabrication needs an experiment that changes the reward during training, which this study does not do. Pre-registration: doi:10.17605/OSF.IO/S53HY. Code and data: https://github.com/zainnab-sparq/citation-reward-hacking

Activation steering controls large language models at inference time by adding a vector to their internal representations. Most work asks how to build that vector, and far less attention has gone to where to add it, even though steering at a poorly chosen layer can be worse than not steering at all. Prior selectors choose a layer without ever steering the model to test it: they score how separable the target behavior is at that layer, and how consistently the individual examples point along the steering direction. We argue that what matters is not the separability of the two activation clouds, nor the alignment of individual displacements with their mean, but the extent of the cloud of per-example displacements that mean-difference steering averages. We propose a topology-based layer selection method, called TopoSteer, which scores a layer by the total H_0 persistence of its displacement cloud, computed as the length of the cloud's Euclidean minimum spanning tree. We show that the score focuses on the displacement cloud's relative geometry, while class separation affects it only through standardization. We also provide a theoretical explanation for our method by proving that, as example-level variation increases, TopoSteer and discriminability move in opposite directions, so the two criteria can rank layers in opposite ways. Across multiple language models and steering tasks, TopoSteer provides a principled alternative to discriminability- and separability-based layer selection and yields more effective steering without requiring held-out validation.

Large Language Models (LLMs) are increasingly deployed in multi-agent settings, exhibiting emergent collective behaviors, yet the role of curiosity as a driver of such dynamics remains understudied. We introduce CuriousTogether, a framework evaluating curiosity-driven multi-agent LLM interactions across three complementary dimensions: novelty-seeking, epistemic and social curiosity. Across eight tasks spanning reasoning, creative generation, and social bias mitigation, curiosity-guided multi-agent settings yield improvements over non-curious ones on seven out of eight tasks, with gains of up to 20.6% in story generation (semantic diversity) and a 14.9 pp reduction in stereotype rates. Furthermore, country-conditioned cultural grounding show benefits for story generation and social tasks, but can hurt precision-oriented reasoning. Finally, we show that inference-time multi-agent curiosity-based approach offers favorable quality-token trade-offs, compared with single-agent baselines. Together, these findings support curiosity-guided interaction as a practical mechanism for improving multi-agent collaboration, with downstream benefits for reasoning, creative, and societal tasks.

Text-to-SQL assistants for streaming engines must handle dialect-specific types, windowing rules, and temporal semantics, including constructs poorly handled by generic prompts. We present Self-Refining Prompt Induction (SRPI), an offline method that converts generation failures into reusable dialect instructions while keeping model weights fixed. A Refiner revises the Generator's prompt using reference-conditioned feedback from three LLM judges; a sampled regression guard checks earlier successes, and a separate Compressor reduces prompt length. Inference uses the frozen prompt without retrieval or query repair. We evaluate cross-domain and cross-model transfer on Confluent Cloud Flink SQL using a controlled synthetic workload seeded by 30 expert-written examples, with 335 induction examples and 257 questions from held-out domains. The same induced prompt improves mean judge scores across seven Generators by 7.1-19.2 percentage points, including an increase from 85.3% to 97.5% for the induction model. A separate compression study reduces prompt length by 52.7% with a 0.19-percentage-point score decrease. An assistant using the induced prompt is in early access, with initial user feedback to inform subsequent refinement.

Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals? I introduce ConfabQA, a 784-item benchmark whose obscure-pre-cutoff category breaks the typical correctness/knowledge-cutoff confound and map the geometry of the hidden state with linear probes, a sparse autoencoder (SAE), and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B the probe separates upcoming refusals from confabulations well beyond prompt features (+7.4/+16.3 pp); on Llama 3.2 3B the ConfabQA margin is within fold-to-fold spread, but a direct refusal probe on PopQA and TriviaQA recovers +18-25 pp. On Qwen I closely inspect three candidate directions for the refusal vector: probe-recovered, Qwen-Scope SAE feature, and directly optimized. Although the three directions are pairwise 50-81 degrees apart, a push at or below the state's own norm along any of them flips most confabulations into refusal openings dominated by the "As..." family. Forcing the opener without touching the hidden state yields the same one-in-three judged-refusal rate as any push. The opposite push collapses refusals mostly into confabulations, yet leaves originally correct answers intact, so the direction gates the opening speech act rather than the knowledge behind it. Confabulations do not resolve into a concentrated feature in a dictionary or a token signature on Qwen3-1.7B. Scoring every probe on the margin over the strongest of four prompt-feature baselines extracts the hidden state's own contribution. Correctness signal is model-dependent and portable only for Llama 3.2 3B: 81% of the probe's margin over test majority survives transfer between ConfabQA, PopQA, and TriviaQA. Benchmark, judge, and code are released.

LLM agents increasingly rely on skill documents, yet most retrievers rank them by description alone and ignore what earlier executions revealed, even though a relevant skill can hurt execution. We present CompSkill, which indexes skills with learned 512-d keys and compiles matched same-task comparisons into skill offsets and capped key residuals (KeyOpt), so experience changes which document the executor receives, never its text. Under a logistic outcome model with unknown task difficulty, matched comparisons cancel that difficulty and only discordant ones carry information; on support executions, predictions improve with each discordant task. An experience hypergraph keeps each comparison with its executor and document versions, recalls missed skills and grounds a cited review. Compact keys raise SkillRet NDCG@10 from 64.0 to 70.9 at one eighth of the dimension. With one document, CompSkill completes 256 and 260 of 332 ToolQA tasks with GPT-5.6-luna and DeepSeek-V4.1-Flash, respectively, more than every baseline, and stays on par with the strongest baselines on AppWorld and ALFWorld. On 332 newly generated ToolQA tasks, it completes 256 with GPT-5.6-luna, significantly more than BM25, SkillRet-8B, Tool2Vec and RankGPT; over all 844 ToolQA tasks, it completes 24 more than the strongest baseline. Selection matters beyond relevance: pairing the right guide with a same-database sibling instead of another database's guide costs 8 of 96 tasks. Code, models and support data: https://github.com/satoshinji2992/CompSkill

Post training for large reasoning models remains insufficient on challenging problems, whose correct trajectories are hard to sample. Though prior works address this difficulty by integrating supervised fine tuning on expert trajectories with reinforcement learning, such external experts are not always available. Existing self improvement methods instead use reference solutions to create a privileged self-teacher. Yet privileged teacher correctness does not imply student reachability. Solution conditioning shifts the teacher distribution toward actions unlikely under the unprivileged student, producing seemingly "correct" trajectories that are unreachable or even mal-formed learning targets. Accordingly, we identify reachability alignment as the key to self constructed supervision and propose Self-Steered Reasoner (SSR). SSR realizes this principle through student support projection, which transforms the privileged teacher distribution into a directional signal only over reasoning actions locally reachable under the unguided student. Verification then retains correct trajectories sampled from this reachable space. To further boost the sampling efficiency, an iterative refinement framework is adopted, using diagnostic guidance based on current trajectories. On problems with pass@64 = 0 even after extensive RL, SSR recovers correct and learnable trajectories for 74.6% without parameter updates. Fine tuning on these trajectories followed by resumed reinforcement learning improves aggregate accuracy by 4.9 percentage points on AIME25/26 and HMMT26. These results show that privileged guidance becomes effective self-improvement data when aligned with the reasoning actions the student can reach.

Finding the passage that answers a question is the first step of most practical systems built on documents, and for Tajik we found no openly described model for it. We present Saidzoda-Tajik-E5-Large, a Tajik text embedder built from multilingual-e5-large-instruct (560M parameters). We asked our own Tajik language model to write questions for passages from Tajik laws, tax documents, Wikipedia and web text, and trained on 235,140 question and passage pairs for one epoch, with hard negatives chosen by another embedder and English and Russian pairs mixed in so the model would not forget those languages. On laws the model had never seen, the right passage is among the top ten results for 93.6% of questions, up from 80.2%; on tax documents, 94.5% against 82.6%. On Belebele, the one test we have that was written by people, the gain is smaller and not confirmed (92.2% to 95.2%, with touching intervals), and Russian and English do not change. On law and tax questions written in the same style as its training questions, it is tied with Giga-Embeddings-3B, a model five times its size; on Belebele the 3B and 8B models score slightly higher, and only after we fine-tuned them on our data are they clearly ahead. We also report what failed: training in bfloat16 broke the model, and replacing Tajik letters with similar Russian ones did not help. Our law, tax and Wikipedia test questions come from the same model that wrote the training questions, so we treat those scores as an upper limit until questions written by Tajik speakers are available.

When a language model decides refund or claim requests under written policies, the system can show it similar past cases. Each past case has two parts: its facts and the decision recorded for it. We separate the two. A model sees a policy, a request and three past cases, and we change either the recorded decisions or the facts, keeping the request fixed. We select requests that are close calls: deciding without past cases, DeepSeek V4.1 Flash approved each in some of four tries and denied it in others. When the past cases are near-copies of the request (its facts with a few details changed), switching their recorded decisions from DENY to APPROVE raises approval by about 30 to 55 percentage points in DeepSeek V4.1 Flash, GPT-5.6 Luna and Claude Haiku 5.5 on policies written for the study, and DeepSeek and Luna show the same on nine public policies paraphrased from real ones. The facts push the other way: shown without their recorded decisions, past cases whose facts support approval make these two models approve less often than past cases whose facts support denial. One reading is that the request looks weaker next to them. This push is not just a correction toward the right answer: in an additional DeepSeek analysis, it was of similar estimated size on requests that human annotators judged approvable and on those they judged deniable. In a separate DeepSeek run on the public policies, when the past cases are near-copies of a different request, switching their decisions moves approval by 9 points instead of 43, while facts that support approval still lower it by 22 points against facts that support denial. Systems that retrieve past decisions should test the two parts separately.

Text-to-image models have achieved remarkable success in generating high-quality images from textual descriptions. However, they often struggle with "semantic drift," where the generated output fails to precisely align with complex or nuanced text prompts. While recent approaches have attempted to address semantic errors regarding attribute binding or object presence, there remains a gap for a more holistic method that addresses these issues by directly refining the text embeddings of the initial user prompt. In this work, we introduce Orthogonal Drift Correction (ODC), an inference-time guidance method designed to mitigate semantic drift without requiring model retraining or additional user inputs. ODC guides image generation through a two-stage process. In the first stage, it identifies the semantic drift by evaluating the initially generated image against the user prompt in a shared vision-language embedding space. It then isolates the component of this drift vector that is orthogonal to the prompt's direction and translates it back into text via a vocabulary-based surrogate mechanism. In the second stage, it produces refined text conditioning for a second generation pass by feeding both the initial text embedding and the re-embedded drift representation into an adaptive rank-reduced concept removal module. Our experiments demonstrate that ODC consistently improves prompt-image alignment, increasing CLIP scores by up to 5.4% over the baseline on the COALBench-300 and PartiPrompts benchmarks. Furthermore, on the T2I-CompBench complex subset, ODC improves the average 3-in-1 score by 4.5% over the competing methods. As a plug-and-play module, ODC offers a practical and efficient method for improving the reliability of text-to-image models. Code is available at https://github.com/kasrakamali/ODC.

On-policy distillation (OPD), which uses reverse-KL supervision on student rollouts to transfer a context-conditioned model's behavior into student model parameters for closed-book inference, has shown strong results. However, we find that this success does not translate to knowledge injection: across two document-QA benchmarks, OPD injects document knowledge far less efficiently than simple off-policy distillation. We trace this reversal to two failures that are structural, yet sparse across token positions. Reverse KL weights its gradient by the student distribution and therefore barely lifts document-specific tokens the student considers unlikely; after forward KL restores this signal, student rollouts expose later positions on the target trajectory only after generating the preceding tokens correctly, so coverage deteriorates with target length. Motivated by the concentration of teacher-student divergence at these hard positions, we propose Adaptive-Policy Distillation (APD), whose per-token gate jointly shifts the rollout toward the teacher and the objective toward forward KL where needed, withdrawing both interventions as the student learns. Without labels or a verifier, APD improves over OPD by up to 54.6%, achieving the best scores of 0.558 F1 on QASPER and 0.486 MRR on BioASQ.

Does the level of test-time processing effort affect a language model's computational capability? A formal grammar-lens benchmark method, derived from the theory of computation, generates novel data at specified levels, each successively more computationally powerful than the levels below. Using this data, we tested GPT-5.5, and GPT-5.5-Pro-Medium, -High, and -XHigh, showing incrementally increased acceptance of test data: 16, 40, 43, and 52\%, respectively, illustrating the benchmark's ability to discern fine-grained differences among models' added test-time reasoning levels. Additionally, as test-time compute effort increases, model responses diverge from human responses to the same data. This increasing disagreement may turn out to be broadly characteristic of more capable language models, introducing the potential of a general capability-versus-alignment conflict.

Speech-to-speech translation (S2ST) increasingly preserves what is said and how it sounds, but not necessarily when speech and silence occur. For dubbing and other temporally grounded applications, matching total duration is not enough: speech and pauses must also occur at corresponding moments. We formulate timeline-preserving S2ST, which aims to preserve the source endpoint and salient internal pauses on a shared absolute timeline. We present TransDubber, an end-to-end autoregressive system that learns this temporal structure directly from aligned target waveforms, without inference-time frame-level timing control or post-hoc temporal correction. To provide this supervision, we construct 642K exactly duration-matched English-Chinese speech pairs by translating and synthesizing target speech on source-defined temporal scaffolds. Training combines a progressive supervised curriculum with verifier-guided reinforcement learning to balance translation quality and temporal alignment. On a 4,353-utterance test set spanning read and spontaneous speech, TransDubber increases the proportion of translations within 2% of the source duration from 25.6% to 61.5% and pause-alignment F1 from 0.118 to 0.550 compared with UniSS, while maintaining competitive translation quality. Our analysis shows that aligned supervision establishes timeline following, while reinforcement learning refines temporal precision and partially offsets the translation-quality cost of strict synchronization. Together, these results demonstrate the effectiveness of our data and training framework for timeline-preserving S2ST. Audio samples, code, datasets, and pretrained models are available at https://zhishengzheng.com/transdubber/.

Can a language model learn better ways to reason about other people's minds by improving its own strategies, and are the strategies it learns actually right? We study this with cogrsi. A grounded reader marks who saw each event in a story and uses a short written principle to settle disagreements between its readings. With a principle written by experts, it beats self-consistency at the same token cost. A self-improvement loop then writes the principle itself. It keeps an update only if accuracy rises by a calibrated margin, and in one variant only if it also passes counterfactual test pairs. The learned principles transfer: on 800 BigToM pairs that no system had seen, each of them scores higher than the expert principle. We then check what the learned rules actually say. Each rule is translated into code and run on thousands of small stories with known answers. Most of them give a wrong answer in at least one situation they claim to cover, for example by treating a character who is in the room as knowing what happened even while asleep. Accuracy does not reveal which rules are wrong, and the counterfactual test does not fix this. A strategy can help in practice and still be wrong as a rule, so both need checking. Our code and data are publicly available.

Despite advances in retrieval-augmented generation (RAG), suppressing hallucinations during complex reasoning remains a persistent challenge. We identify that existing RAG pipelines suffer from a textual bottleneck: by passing only raw text to the generator, they discard discriminative signals of the retriever. Furthermore, recent attempts to unify retrieval and generation within a single LLM induce severe objective conflict and exacerbate the rank collapse of decoder-only architectures, degrading performance on both tasks. To address these limitations, we introduce ShiftRAG, a decoupled RAG framework that employs a single LLM backbone but isolates discriminative retrieval learning from autoregressive generation. To bridge these decoupled modes, we project retrieval alignments into continuous soft tokens, establishing a high-bandwidth interface that infuses the generator with retrieval-side signals. Concurrently, our soft orthogonality objective mitigates rank collapse, while layer-wise relevance dynamics preserve the fine-grained latent space required for highly discriminative evidence separation. Extensive evaluations demonstrate that ShiftRAG significantly improves retrieval and question answering performance over state-of-the-art baselines. ShiftRAG preserves foundational capabilities of the backbone while maintaining a favorable efficiency-performance trade-off on general text encoding and generative tasks.

Electrochemistry is central to solving many modern-day challenges, including energy storage systems for a sustainable future, and its advancement relies heavily on the data from past findings. Existing automated pipelines are limited by text-only architectures, context fragmentation, and numerical hallucinations - collectively preventing accurate, large-scale curation of property databases. We introduce ElectrochemLLM, a multimodal information extraction pipeline that jointly identifies structured electrochemical records encompassing electrolyte compositions, electrode materials, cell configurations, and their measured physicochemical properties from unstructured scientific PDFs at scale. By combining hybrid document parsing with vision-capable LLM orchestration, the pipeline extracts verifiable, structured data across text, tables, and figures. Evaluated on the battery-electrolyte subset of our manually annotated benchmark of 232 papers, ElectrochemLLM achieves an F-1 score of 0.97, and has been applied to curate ElectroChemDB, a structured database spanning 15,000 electrochemical research articles. We demonstrate the utility of this resource by training machine learning models for electrolyte property prediction, illustrating a complete pathway from unstructured literature to predictive materials discovery.

We propose a methodology for investigating the differences between natural and synthetic language in a challenging setting: humorous improvisational dialogue. Our experimental setting compares human-produced comedy improvisations with the dialogues generated by LLMs in a multi-agent environment that we designed. The analysis combines automated linguistic measures with manual annotation grounded in Systemic Functional Linguistics (SFL). The results show that, although LLMs can create coherent and creative scenes, they often fail to reproduce key characteristics of spontaneous spoken interaction. In addition, LLM-generated interactions display a strong tendency toward politeness and positive evaluations, contrasting with the broader range of attitudes found in human humorous improvisation. Our evidence highlights both the strengths and limitations of current LLMs in modeling dialogic humor and demonstrate the value of integrating linguistic theory with NLP methods for the analysis of machine-generated language

On-device language models can learn a user's preferences through adapters trained on private data. However, a base-model upgrade may change the layers or vocabulary, leaving the personal adapter incompatible. Replacing it with a non-personal adapter loses much of the personalization, while retraining requires retaining the user's labeled history. We present Heirloom, a system that transfers a personal adapter to a new base model on the phone without retaining labels. Heirloom uses the old personal model as a teacher, collecting its predictions on stored unlabeled inputs, vendor-provided prompts, or prompts generated by the new base. It then unloads the old base and trains an adapter for the new model to match those predictions, keeping only one base model in memory at a time. We evaluate Heirloom across three generation upgrades from two model families, an architecture-preserving refresh, and three personalization tasks. On movie tagging, migration from stored inputs maintains pre-upgrade accuracy and differs from labeled retraining by at most 1.7 percentage points across the three upgrades. On a phone's neural processing unit (NPU), a fixed-length migration for the 4B upgrade takes 38 minutes and consumes 20% of the battery end to end, making the stored-input setting feasible during overnight charging.

KGW watermarks bias sampling to accumulate keyed detection evidence. We construct budgeted evasion through token-event evidence replacement using ordinary phrase rewrites. Our construction jointly selects compositions of LLM-proposed alternatives from an independent model and measures global token-edit distance on complete outputs. At the largest attainable cost within the budget, exact coverage search maximizes removal of original scored token pairs. It uses a known tokenizer without the key, detector feedback or queries to the watermarked model. Exact accounting separates source evidence removed, evidence created and length effects, with initial strength setting the required reduction. Separate signed-event analysis found that mean suppression arose mainly from removing watermark-enriched source events. In controlled, occurrence-counted, one-token-context KGW tests on short passages, 39 of 82 qualified, initially detected sources yielded rewrites with terminal detector score Z < 4 that passed blinded source-relative model review, at a mean actual token-edit cost of 16.0% under a 20% allowance. To examine whether this initial requirement can be inferred, a separate historical 95-passage study estimated initial detector score Z0 from generation traces with correlation 0.793 and an in-sample association beyond length. This aggregate information did not establish improved allocation, and the separate no-trace stopping test did not reliably exploit cheaper reviewed crossings in the candidate bank. Comparisons did not establish an advantage over simpler selection and ordinary constrained rewriting. The construction makes budgeted evidence replacement executable while exposing limits in signed reduction and recognition of sufficient editing.

Large language model (LLM) agents extend their capabilities with skills, each with a name, a description, and a body. Under progressive disclosure, an LLM router selects skills by name and description alone and loads only the selected bodies. Skill libraries grow continually as users and agents add new skills. Because the router relies on this metadata, a similar new skill can displace an earlier one, breaking previously correct routing even though the earlier skill and the router are unchanged. This routing interference is aggravated for public skills, whose descriptions may not match their bodies. We formulate the problem as lifelong skill routing, a class-incremental continual learning problem. A solution must keep earlier skills selectable without suppressing new ones. Existing approaches fall short, since continually retraining the router risks catastrophic forgetting and rewriting methods either ignore similar skills or require per-skill feedback. We therefore propose Skill++, which rewrites only the name and description of each incoming skill, without router training or labeled data. The new description states what the body supports, how the skill differs from its closest predecessor, and what related work it excludes. As a library grows to 731 skills over ten sessions, Skill++ selects all required skills for 76.3% of queries, versus 36.4% under the original metadata. Across three benchmarks and five LLM routers, it outperforms all compared approaches, and its gains carry over to the Codex CLI agent harness.

Retrieval-augmented generation (RAG) helps large language models (LLMs) answer factual questions beyond their training data, but it adds storage and inference costs. Distilling previously retrieved information into parametric memories enables reuse without online retrieval. However, training the memory to mimic a retriever rewards accurate independent predictions, which do not necessarily improve the final output quality when combined with the LLM's existing predictions. To address this, we introduce ResMem, in which the LLM reads retrieved text offline to generate teacher targets. This design avoids storing massive token representations, unlike existing parametric memory approaches such as MLP Memory, which relies on k-nearest neighbor (kNN) datastores. Furthermore, by using the LLM as an offline teacher that reads privileged retrieved evidence alongside the input, we train a parametric memory to learn an additive logit residual. This residual aligns the predictions of a student sharing the same backbone with the teacher distribution, while self-distillation updates remain confined to the parametric memory. On a Wikipedia subset, ResMem reduces the storage retained for teacher construction by 94.7% compared with the evaluated kNN pipeline. Across five question answering datasets, ResMem raises mean token F1 from 32.83% for the frozen LLM to 34.48%. It also breaks 21% fewer of the LLM's correct answers than a kNN-trained memory. These findings validate our motivation of treating retrieval as a learning resource, with parametric memory shaped by how retrieved knowledge is used to inform the model's future predictions. Code and model weights are available at https://github.com/SSRand/From-Lookups-to-Lessons and https://huggingface.co/RandHan/ResMem-Mistral-7B-v0.3.

System One decision models replace text generation with typed, probability-scored answers and are entering production. Answers execute without a person in the loop, and no multi-model audit exists on established bias benchmarks. We provide that audit: three benchmarks: BBQ (all eleven categories), CrowS-Pairs, and SilverSpoon, adapted to the typed interface, twelve models plus a no-thinking ablation, and 1,638,000 decisions. Every order-sensitive item runs twice with options swapped; native typed models also swap labels; ten configurations repeat under extra seeds and batch 16. We find widespread stereotype alignment, order-dependent answers, and confidence that tracks instability rather than bias. In BBQ, 97 of 143 model-by-category scores are significant and 94 aligned, while the commercial reference abstains on 99.5% of ambiguous items, hiding aligned probability mass. Option swaps flip 0.0% to 72.6% of choices; typed scoring stays below 10% except Laya (22-33%), and untuned letter prompting reaches 73%. CrowS-Pairs preference spans 0.496 to 0.816, and boolean answers can be decided by position (1.00 versus 0.005). SilverSpoon judgments shift toward lower-SES labels in five of eight decision models, inherited from the 9B backbone or induced by tuning. Audits must permute options and reverse questions by default; order-invariant scoring should become a procurement requirement.

Semantic caching lowers the cost of serving large language models (LLMs) by reusing stored responses for similar requests. In multi-turn conversations, whether a cached response is acceptable depends on both conversations' histories, which embedding similarity of requests and histories predicts poorly (AUC 0.62–0.74 on five workloads). We therefore learn the \emph{reuse risk}, the probability that a cached response is unacceptable, from judge feedback, and choose a budgeted set of conversation states to cache by weighing the generation cost a reuse avoids against a price for unacceptable reuse. We propose CUCB-SC-MT, which learns a cache from logged conversations with a suboptimality guarantee, and CLCB-LS-MT, which learns online and has $\tilde O(\sqrt N)$ regret over $N$ requests when conversation states recur. On math-tutoring (SocraTeach) and movie-ticketing (Taskmaster-3) conversations, both achieve lower loss than prior semantic caching approaches: ContextCache, multi-turn variants of ElastiCache and vCache, and single-turn semantic caching at every cache budget. With 1,000 cache entries, CLCB-LS-MT avoids 35–41\% of the generation cost through acceptable reuses, against at most 22–25\% for the baselines, and lowers estimated serving latency by 19–35\%.

Whether a policy shifts domains and, by extension, continues, broadens, narrows, or reverses previous policy is of paramount importance to the study of public policy. We propose Dual-Axis Policy Entailment (DAPE), which evaluates an ordered pair of policy directives on both dimensions at once: a binary judgment determines whether the two directives govern the same policy domain, and a five-way judgment determines whether the later directive continues, broadens, narrows, or reverses the earlier policy, or governs a different domain altogether. Read together, the two judgments form one five-category label, our primary metric. The two-axis design captures two central dimensions of policy change without conflating errors on either dimension into a generic "irrelevant" label, and it makes annotation more rigorous: a coder cannot dismiss a difficult pair as domain-irrelevant without an explicit, checkable decision. We introduce an expert-annotated benchmark of 1,000 ordered directive pairs drawn from a large corpus of Chinese government policy, together with an annotation-calibration protocol (an adjudication packet with pre/post agreement measurement) that we contribute as a methodological artifact. Fourteen open-weight models and four frontier API flagships fail the task zero-shot: even the best frontier flagship's chance-corrected agreement with expert labels is far below the human ceiling (Cohen's $κ\approx 0.29$ vs. $0.53$). A fine-tuned open-weight model (LoRA SFT of Qwen3-14B, five-fold cross-validation) closes this gap, reaching the coder-mutual human ceiling ($κ= 0.524$ vs. $0.527$). Finally, we apply our method to China's AI policy (1980-2024) and map the direction of policy change across eras.

LLM-as-a-judge scores are often distributionally shifted from human ratings on any given evaluation task, and post-hoc calibration (mapping LLM to human scores in a small labeled anchor set) has become a popular method to correct for this shift. In this paper we derive the calibration from two positions: that LLMaJ scores are ordinal ratings drawn from a distribution rather than isolated points. We present streaming quantile matching (SQM), a version of a classical method not previously bench marked for LLMaJ calibration, which is built on the two positions by computing the optimal-transport map between the LLM and human score distributions. We develop an evaluation framework and within it we show that unlike point-wise calibrations, SQM respects the ordinality of the human anchor set and achieves the best distributional fidelity and the top overall rank across models and data. We also show that in general, post-hoc calibration narrows the raw error gap between smaller and larger judge models, which opens the possibility of more efficient LLM judges.