LangTrend

The largest open German text datasets are built from broad web crawls, which provide scale but limited coverage of specific domains. We introduce RefinedGAW, a processed and annotated resource derived from 26 semi-annual crawls of the German Academic Web, totaling 122 million documents and 314 billion tokens, with full crawl and capture history of every text. Documents labeled as academic research account for 84% of German RefinedGAW tokens across HTML and PDF. At 87.6 billion tokens, this exceeds the 64.9 billion tokens in the German subsets of FineWeb2 and FinePDFs, despite their almost five-times-larger combined size. RefinedGAW also contains substantial amounts of educational, math-heavy, and expert-oriented content with minimal commercial bias. Across HTML and PDF, 88% of German RefinedGAW documents have no near-duplicate candidate match in the three evaluated reference datasets, supporting complementary coverage. We release a queryable metadata index for fine-grained data selection and a rehydration tool for retrieving and processing source documents.

Large language models (LLMs) are neural networks trained on huge text collections to generate text by predicting what comes next. Engineers using ChatGPT and other generative artificial intelligence (AI) tools need a clear picture of what training changes and prompts add. The article starts with the Transformer, which connects words in context, and pre-training, where models learn to predict text. It then explains how fine-tuning adapts models to requests, with reinforcement learning from human feedback using human answer rankings. Next, retrieval-augmented generation (RAG) supplies documents, while chain-of-thought prompting encourages intermediate steps and agents use tools. FlashAttention lowers training and running costs by reducing data movement during attention; quantized low-rank adaptation (QLoRA) saves fine-tuning memory. Readers get a life cycle figure and tables comparing methods and their limits. A short checklist for judging LLM evaluation claims flags test material in training (data contamination) and model-based grading (LLM-as-a-judge). It also covers very long inputs (long context) and hallucination, meaning false or unsupported claims. Check factual accuracy separately from whether an answer sounds helpful.

Text injection trains speech recognition models on unpaired text alongside paired audio. We feed the speech encoder synthetic input derived from text so the whole model learns from it. Previously, we synthesized this input with a frozen text-to-speech (TTS) model. Here, a simple lookup table replaces the TTS model: a mean log-mel vector (or one-hot encoding) per phoneme, repeated for a random duration drawn from alignment statistics, with silence between words. Training is 1.8 times faster than with the neural TTS, achieving comparable word error rates. With an external language model, our system reaches 1.36 / 2.90 / 1.52 / 3.24 % WER on LibriSpeech {dev,test}-{clean,other}, to our knowledge state-of-the-art using only LibriSpeech paired data, text, and lexicon. We analyze components of synthetic input (durations, acoustic features), compare different HMM aligners, evaluate scaling on the 25 000 h Loquacious dataset, and show that full-encoder injection yields strong relative improvements compared to partial-encoder and decoder methods. The code of the whole pipeline behind every number in this paper is public: github.com/rwth-i6/returnn-experiments/tree/master/2026-text-injection

Estimating the correctness of large language model (LLM) predictions without access to ground truth is critical for reliable deployment in high-stakes settings. Existing approaches, such as entropy and perplexity, operate solely on final output statistics, discarding the rich information encoded in intermediate reasoning steps. To address this limitation, we propose STEP (Soft Trajectory Estimation of Prediction correctness), a training-free framework that scores each reasoning step via lightweight forward passes to obtain a soft probability distribution qi over candidate answers, from which we derive a set of soft trajectory features capturing per-step confidence and its progression toward the final answer. Extensive experiments on various benchmarks demonstrate consistent and substantial improvements, particularly in AUROC, over output-based baselines and hard-indicator trajectory methods. Notably, Temporally-Weighted Consistency, which assigns higher importance to later reasoning steps where answer distributions have stabilized, contributes the largest gains, highlighting the importance of temporal structure in reasoning trajectories. These results further validate the advantage of preserving continuous probability signals over binary discretization.

A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size $N$. On streaming replays of LongMemEval and LoCoMo, a similarity threshold tuned to 5% false alarms at $N=8$ reaches 87-100% at $N=2048$, across embedding models and for explicit and implicit cues; LLM judges show the same growth from a lower base. We prove that a gate with false-alarm rate $α$ operates at a per-item false-positive rate of about $-\ln(1-α)/N$, which pins its hit rate to the far left of the single-comparison ROC curve and, under regular variation, makes it decay as a power of $N$ with an explicit constant. The resulting prediction matches the hit rate of a per-$N$ oracle threshold to within one percentage point. The analysis explains why lexical and semantic channels need thresholds growing like $\ln N$ and $\sqrt{\ln N}$, and why per-cue normalisation roughly doubles the hit rate at matched false alarms. A cue-swap conformal gate, which compares the current cue with the user's other cues on the same store, needs no labels and has a finite-sample false-alarm guarantee under exchangeability; on LongMemEval it holds false alarms near 5% at the hit rate of a label-tuned threshold. Three annotators (Fleiss' $κ=0.63$) find that about nine in ten false alarms are genuine interruptions. Code and data: https://github.com/RiverChenStillCompiling/price-of-scale

The Babylonian Astronomical Diaries record the sky over nearly six centuries. Many tablets are broken and have lost their date formula. Experts dated many of them one tablet at a time in the standard edition, Astronomical Diaries and Related Texts from Babylonia (ADART). They used the recorded sky and other evidence, such as year formulas and royal names. Our system automates the astronomical part for all tablets with the same rules. A rule-based parser turns the recorded observations into clauses, short statements with fixed fields. A language model checks them against the text and adds observations that the parser missed. Astronomical tools test the clauses in all 701 years from 626 BCE to AD 75. The system ranks the years by how well they agree with the clauses. The system withholds published dates until its results are final, and the output lists the passages and computed positions behind each candidate year. On 120 dated tablets, the years tied at the top of the ranking include the published year in 96% of cases. For 86 of them, this top group is a single year, the published year in 84. In 119 of the 120, the reported candidate set contains it. For 56 undated entries of ADART IV, the system gives ranked candidate years. For each tablet, the results show how far astronomy alone narrows the year. An interactive page at https://tingyuansen.github.io/Babylonian_Dating/ follows each tablet and links the code, protocol and results.

Large reasoning models are costly to serve because they are large and because they generate long chains of thought (CoT). Weight pruning and CoT compression address these two costs separately, but they interfere in that pruning adds redundant reasoning on problems the model still solves and incomplete reasoning on those it no longer solves, and on harder problems makes the model less compressible by subsequent length-preference optimization. We trace this to a blind spot on each side: pruning criteria do not distinguish necessary from redundant reasoning, and preference-based CoT compression receives no training signal on problems a pruned model always or never solves. We propose Contrastive Set-Difference (CSD), which identifies neurons enriched in redundant reasoning by contrasting their activations on long and short correct solutions of the same problem and prunes them preferentially, and we extend preference pairs to the problems a pruned model always or never solves. Pruning 20\% of DeepSeek-R1-Distill-Qwen-7B, the combination uses 8–22\% fewer thinking tokens on four math benchmarks than pruning with reasoning-aware calibration (RAC) followed by the same compression, with accuracy within 0.5 points.

Large language models are increasingly being used in different stages of scientific knowledge production, including the suggestion of references. These models can operate at different levels of reasoning effort and provide summaries of their reasoning processes. This raises two questions: Does increasing reasoning effort improve the verifiability of generated references, and do the resulting reasoning summaries reveal how the model's behavior changes with greater effort? We examined summaries from GPT-5.5 as it generated scholarly references for 1,000 papers under paired low- and high-effort settings. Generated references were matched to external bibliographic records, and reasoning summaries were scored for four predefined features and explored for additional features discovered by GPT-5.5. High effort increased the mean paper-level resolver-verified reference rate from 78.3% to 84.8%, and produced longer reasoning summaries. All four predefined features received higher scores under high effort. The discovered features changed less uniformly: some appeared equally often in both conditions, whereas others were more common under high effort. Within each effort condition, neither feature set showed a consistent increase in verification rates with higher scores.

Supervised Fine-Tuning (SFT) of LLMs suffers from catastrophic forgetting, and prevailing mitigation strategies struggle to reconcile the plasticity-stability dilemma. To address this, we propose Entropic Control for Homeostatic Optimization (ECHO), a closed-loop control framework. By decomposing the KL divergence, we identify output entropy as a dominant observable component of KL variation. This enables us to leverage entropy as a computationally efficient online proxy to monitor forgetting risks in real-time, bypassing expensive reference-model inference. Consequently, ECHO employs an elastic controller that dynamically modulates the gradient weight of replay data via a proportional control law, using deviations from the pre-trained entropic equilibrium as an error signal. Empirical evaluations across 0.6B to 9B parameter architectures demonstrate that ECHO lies on, and in several comparisons extends, the observed plasticity-stability frontier and maintains high operational robustness against varying data mixing ratios.

Generative models increasingly expose a reasoning effort setting that lets users scale inference-time compute, trading higher cost for potentially improved performance. Users must choose an effort level before knowing whether extra reasoning will help, and even a successful run does not reveal whether the effort purchased was necessary. Reasoning effort is therefore a credence good: buyers observe the outcome but cannot assess the marginal value of the computation they paid for. In this work, we measure the counterfactual effect of effort levels on both proprietary and open-weight models on four benchmarks, spanning single-call reasoning and long-horizon agentic tasks. Effort, it turns out, does not consistently preserve tasks that lower effort solves, can overspend, and can change trajectory behavior. First, we find that although higher effort can raise accuracy in aggregate, it frequently fails on tasks that lower effort solved. Second, on the subset of easy tasks that every effort level solves, higher effort is generally more costly, with the highest effort level costing up to 42x as much as the lowest. Lastly, effort changes not only how many tokens are used, but also how agents solve tasks: at higher effort, agents verify their own work more heavily and more often search for solutions than work through problems directly, both of which lengthen trajectories without necessarily altering the outcome. These findings show that naively choosing the highest level is not only cost-inefficient but also offers no guarantee of the best instance-level performance. A simple alternative, starting at the lowest effort and escalating only after a verified failure, matches or improves on fixed top effort across all models and benchmarks at lower cost for most proprietary models.

Structured semantic parsing is typically judged by whether the correct category is assigned, which leaves value recovery and instance-level associations unexamined: a system may recognize an event yet bind an argument to the wrong instance, or identify an intent yet supply an incorrect slot value. We present USP, a cross-family semantic contract that represents categories, fields, values, and instance associations through a common interface for named-entity recognition, relation extraction, event extraction, and intent-slot parsing. The contract is paired with five diagnostic matching views—category agreement, field recognition, field-value pairing, instance-bound assignment, and complete-instance recovery—that ask distinct questions about the same output. Using this contract, we normalize 61 datasets into a unified corpus with shared schemas, prompts, and matching rules, and compare six models on a fixed test subset. Our fine-tuned Xing4.0-29B-Intent attains 57.70 task-balanced Field-Value Micro-F1 (equally weighted across the four task families) and ranks first in every family and on all five views; its margin over the strongest alternative widens from 2.93 points for category matching to 20.73 points for complete-instance matching. This widening gap shows that category agreement alone understates structural deficits. USP thus offers the community a shared representation, a normalized 61-dataset corpus, and a diagnostic evaluation protocol that make structural agreement measurable and comparable across extraction and intent-slot tasks.

Extracting salient information from long-horizon dialogue histories remains a significant challenge for Large Language Models (LLMs) due to the restricted context windows and efficient retrieval challenges. Current methods typically rely on history organization or local token-level optimization, suffering from information loss, retrieval redundancy, and myopic attention. Inspired by the view in cognitive science that dialogue constitutes sense-making and memory association, we propose CoLink, a framework that enhances LLMs' global context awareness in domain-specific multi-turn dialogues through information linking and compression. CoLink first clusters historical contexts into topic graphs to establish cross-semantic associations with external knowledge. Guided by the adaptive control of thought, it then retrieves and activates knowledge via synergizing semantic relevance and graph-structural topology. Crucially, to achieve effective compression of history and knowledge, we ground multi-turn dialogues in information theory and derive a principled optimization objective for extracting the core sense-making information. Coupled with a competitive slot mechanism, CoLink condenses the global context into dedicated memory and knowledge slots, efficiently distilling lengthy, redundant inputs into an extremely compact latent representation. Extensive empirical results on both public and our MMD datasets validate CoLink's superiority and generalizability in knowledge linking and sense modeling.

The stance expressed in responses from retrieval-augmented generation systems is significantly influenced by the stance distribution of the retrieved materials. Existing research on knowledge poisoning has primarily targeted factual errors. In contrast, this study focuses on injection attacks in which the facts are correct but the stance on a controversial issue is one-sided. At a mean topic-level poisoning share of 18.9% within the post-injection ProCon documents, all four generation models from different vendors exhibited stance shifts aligned with the poisoning direction; across four concentration levels, the shifts increased monotonically. This paper proposes a Stance-Balancing Middleware positioned between the retriever and the generator. It consists of stance detection, fixed balancing (2 supporting + 2 opposing + 1 neutral), and a debiasing prompt, without introducing additional large language model calls. Across all eight concentration-direction settings, the complete method reduced the magnitude of stance deviation by an estimated 0.106 to 0.350; the six settings at poisoning shares of 11.9%-18.9% remained significant after Holm adjustment (adjusted p <= 0.009). In the cross-model experiment, all eight model-direction combinations moved along the stance axis in the corrective direction, and seven of eight showed a clearly detected reduction in absolute distance from 0. In the concentration sweep, the post-defense intervals included 0 in all eight settings, while the cross-model pro-poisoning results retained small positive residual point estimates of +0.045 to +0.063. For 20 candidate documents, stance detection took approximately 1.5% of the median generation time on a single GPU and approximately 16% on CPU.

Large Language Models(LLMs) have shown strong performance on tasks with verifiable answers, but many real-world queries involve subjectivity, disagreement, and conflicting evidence. Most existing benchmarks for such queries remain English-centric, overlooking the vast diversity of perspectives held by users worldwide. To systematically study how LLM perspectives vary across languages, we curate PerspectivesQA, a manually verified multilingual dataset of globally contested topics covering 18 languages, and propose ROOTS, a framework that profiles linguistic, perspective, and framing diversity. We evaluate a diverse set of eight models covering both open-source and proprietary families at varying scales, establishing that LLMs hedge their answers in English but commit to positions in lower-resource languages. Larger models maintain a more consistent cross-lingual worldview, while smaller models are more easily swayed. All models also exhibit high sensitivity to minor surface-level input variations, producing notable stance even in larger models.

Entity matching is the task of deciding whether two records describe the same real-world entity. Current entity matching methods rely on pre-trained language models (PLMs) which require task specific fine-tuning or generative large language models (LLMs) which do not require task-specific fine-tuning but are slow and expensive at matching large record sets. Decision models are a new class of LLM-based models that return a probability for each predefined answer option rather than generate text. This paper investigates decision models as an alternative to generative LLMs and fine-tuned PLMs for entity matching. We evaluate six decision models in a zero-shot setting across eight benchmarks and compare their performance with that of the generative LLM GPT-6 Luna and the PLM-based matcher Ditto. Our experiments show that, without task-specific training data, the four strongest decision models achieve mean F1 scores close to that of GPT-6 Luna, with the best model slightly surpassing it. This model also outperforms Ditto on six of the eight benchmarks. All decision models are faster than GPT-6 Luna, and most are less expensive. For three of the four strongest decision models, the predicted match probabilities are well calibrated without requiring any task-specific recalibration. A model cascade that routes the least confident record pairs to an ensemble of generative LLMs improves the mean F1 scores of the four strongest decision models by 1.9-4.0 percentage points. Using the best decision model and its best-performing prompt for each benchmark, the cascade surpasses the highest published F1 scores from literature on four of the eight benchmarks.

This work proposes a self-attention mechanism based on the fully connected quantum Boltzmann machine (FC-QBM), and further constructs a fully connected quantum Boltzmann machine language model (FC-QBM-LM) using the mechanism as the core engine. By performing unitary evolution on semantic quantum states, the mechanism implicitly implements positional encoding and self-attention computation, eliminating the need to decompose continuous semantic streams into discrete matrices for explicit matrix multiplication, with the measurement results naturally satisfying the normalization condition. Specifically, this work rigorously derives the computational principle of the self-attention mechanism based on FC-QBM, and demonstrates that this mechanism is mathematically equivalent to the classical self-attention mechanism. The mathematical principles governing the training and generation of FC-QBM-LM are thoroughly derived, and the two corresponding algorithms are subsequently developed. An 8-qubit FC-QBM-LM is finally constructed and trained on the corpus of Tang poetry. Under the 8-qubit setting, the quantum layer implements the entire positional encoding and self-attention computation pipeline with only 44 parameters. In the noiseless scenario, the model exhibits favorable convergence. Furthermore, the quantum circuit proposed in this work, which terminates with Y-axis rotation gates, achieves better convergence than the counterpart terminating with X-axis rotation gates. Under the small-sample regime and across eight distinct noise intensity levels, the model demonstrates strong noise robustness, retaining over 99% of its noiseless convergence performance. Experiments on five-character quatrain generation reveal that the model preliminarily exhibits three core capabilities: associative and divergent thinking, content generation across imagery and human-aligned word ordering.

In formal systems such as Lean 4, efficiently finding answers to mathematical problems and generating verifiable proofs is an important task in automated mathematical reasoning. In the traditional way where an upstream solver provides multiple candidate answers, the common workflow requires searching for proofs for each candidate separately, incurring redundant computational cost. To this end, we propose a joint proof framework that integrates candidate answer selection, answer supplementation, and formal proof generation into a unified reasoning process. At its core is a trained prover that selects given or proposes new answers and generates a proof for at least one answer in a single call. Joint proof can also reuse generated proofs to check for faithfulness errors. For problems with a unique answer, if two distinct answers both obtain verifiable proofs under the same formal statement, this confirms the statement does not faithfully preserve the original problem semantics. On MATH-500, AMC 2023, and AIME 2024/2025/2026, our framework achieves the same formal proof success rate as per-candidate proof generation with about one sixth of the token cost, and identifies at least 6.1% of the formalized statements as unfaithful without requiring additional judges. Our model and code are available at https://github.com/Linppz/Gobble-Prover.

"Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which simple example-level metrics, drawn from prior work on data selection, text quality, and instruction tuning, best predict training utility under a fixed finetuning budget. For each metric, we sort instruction-response examples by score, partition them into low, middle, and high bins with roughly equal token mass, finetune separately on each bin, and compare validation perplexity and LLM-as-judge scores for the models' generated answers as a function of cumulative training tokens. Our results suggest that answer-dense supervision, as captured by response ratio, is a stronger and more portable predictor of instruction-tuning utility than commonly used alternative proxies.

Large language models increasingly collaborate with humans and peer AI agents whose contributions may be incomplete, inconsistent, or incorrect. Effective collaboration requires calibrated trust, whereby decisions to rely on, verify, or assist a partner adapt rationally as empirical evidence accumulates. We investigate whether models translate partner evidence into appropriate reliance policies. We introduce a benchmark of 324 scenarios across 23 collaboration categories. Each scenario involves three sequential decisions with explicit transition dynamics, feedback emissions, and cost schedules, permitting exact analytical evaluation of choices against their expected consequences over the remaining interaction. Evaluating seven models alongside 36 controlled-history probes, we dissociate evidence sensitivity from normatively appropriate action. While the strongest frontier model approaches near-optimal decisions, other models exhibit recurring pathologies: procuring excessive safeguards, purchasing verification tests with zero expected value of information, prioritizing immediate savings over cooperative learning, and misallocating assistance to suboptimal channels. Supplying explicit posterior distributions resolves certain errors while introducing new failure modes. Calibrated trust in autonomous agents critically demands coupling empirical evidence with forward-looking decision utility.

Insertion-based masked diffusion models enable flexible-length, parallel generation by interleaving mask insertion with token unmasking. However, because their generation decisions are irreversible, aggressive parallel decoding can introduce errors that subsequent steps cannot correct, necessitating conservative decoding with more forward passes (NFEs). Moreover, they exhibit higher entropy than standard MDMs because they marginalize over multiple alignments during training and therefore require even more conservative decoding in the absence of a correction mechanism. To address this limitation, we introduce $δ$MDM, which makes intermediate generations revisable by equipping an insertion-based model with substitution and deletion operations. During training, we expose the model to errors caused by aggressive unmasking and insertion, teaching it to correct its own mistakes. At inference time, we interleave insertion and unmasking with correction steps that can revise previously unmasked tokens and remove unnecessary masks. This mechanism makes aggressive decoding practical, improving both generation speed and quality. $δ$MDM Pareto-dominates FlexMDM, a vanilla insertion-based MDM, on the speed–quality trade-off. On high-diversity unconditional text generation on OpenWebText, it halves the generative perplexity relative to FlexMDM at the highest inference budget. It also achieves lower generative perplexity at all inference budgets compared to MDM trained from scratch with variable-length padded sequences. On low-diversity conditional text generation on TinyGSM, it matches FlexMDM accuracy with half as many NFEs.

Best-of-N (BoN) inference returns the candidate with the highest proxy reward, so a larger $N$ tests the proxy's extreme tail more than its average ranking quality. We show that the exact BoN value is a Beta-weighted integral of conditional true reward over within-prompt score quantiles. It follows that BoN depends on scores only through their ranks. It also follows that BoN can decline only if the calibration curve in the tail turns downward, and the budget where the decline appears depends on where that turn lies. Under a bounded additive model, the error tail sets the limit: Gaussian errors keep improving, Laplace errors reach a plateau, and heavy-tailed errors return to the prior mean. When error variance is largest at low true reward, even Gaussian noise can push BoN below random selection. Synthetic phase maps agree with this taxonomy. On real candidate pools spanning several verifiable domains, base policies, and learned reward models, tail failures are rare and all occur in code generation. A held-out tail calibration statistic reliably predicts where a cross-fitted budget cap helps. It does about as well as rank accuracy, and the cap is competitive with a Soft-BoN baseline. Tail calibration therefore does not replace ranking metrics. It is an audit that is unchanged by any monotone rescaling of the reward scores and shows which budgets are at risk.

"Why is my period late?" Sometimes, the most useful response begins with another question. Yet conversational AI systems may respond to incomplete health queries without seeking essential context or grounding their guidance in verifiable evidence. This paper introduces HerHealthGPT, an evidence-grounded conversational AI system designed to address these challenges in women's reproductive health. HerHealthGPT integrates retrieval-augmented generation (RAG), context-aware clarification, and source validation within an interactive web interface. Rather than responding immediately to incomplete queries, the system maintains conversational context, asks targeted follow-up questions, checks for premature conclusions, and grounds its responses in retrieved evidence. We evaluate the AI components using 540 English messages derived from 90 reproductive-health cases and conduct a prototype questionnaire study with 30 adults. Evidence retrieval improves concern-category accuracy from 54.1% to 59.3% compared with fine-tuning alone. The clarification router achieves 85.6% recall and 86.2% specificity on constructed targets. In the user study, 83.3% of participants preferred the retrieval-enabled configuration, while 25 of 30 considered its follow-up questions relevant. These findings suggest that integrating evidence retrieval with interactive clarification offers a promising approach to more transparent, context-aware reproductive-health information systems, while further clinical validation remains necessary.

Long-context language models can increasingly process entire documents, making full-context prompting a practical alternative to passage retrieval. However, greater capacity determines how much information a model can receive, but not which information best supports a query. This distinction motivates rethinking retrieval: when the full document already fits, can selecting and organizing its evidence improve answering? The challenge is to retain supporting facts and relations while keeping the context focused on the question. To reconcile these goals, we introduce \ours, a training-free framework that uses semantic abstraction to guide selection while retaining original text for answering. \ours constructs a reusable semantic hierarchy from the full document in a single generation, linking node descriptions to source spans. Given a query, structural–semantic scoring and fine-grained lexical attribution nominate shared source nodes; their union with ancestor context preserves structural relations and complete local text. Across four frontier answer models, \ours improves average QA-F1 by 1.72 points over full-context prompting on three LongBench tasks, surpassing the evaluated baselines on the four-model average while reducing answer-context length by 58.55\%. Our code is available at the $\href{https://anonymous.4open.science/r/treefuse-ECFE}{repository}$.

Professionals send working documents to external large language models. A common safeguard is reversible pseudonymisation: personal data are replaced with placeholders before the request and restored locally in the answer. We present an open benchmark of five existing open tools and a regular-expression baseline on Russian and UK English, and an open engine built on the two strongest tools. The benchmark excludes each tool's own authors' dataset from its verdict, compares the engine with each tool on identical datasets for detection, reports masking precision, and tests on a held-out set of synthetic documents whose files, hashes, decision rules, run plan and stopping rule were published before any run. Two local LLMs test whether placeholders survive. All evaluation data are synthetic, and the benchmark was run by the author of the engine; code, data, predictions and raw answers are released. On external datasets no existing tool meets all predefined thresholds: in UK English the identifier and leak thresholds are not met, in Russian the identifier threshold is borderline. On the last held-out set of 1,000 documents per language, the best existing tools leave a whole value visible in 61% (Russian) and 58% (UK) of documents. On these documents, from the generator it was developed on, the engine does so in 0.3% [0.1; 0.9] and 0.2% [0.1; 0.7], leaves any part of a value visible in 0.8% and 0.2%, and meets all detection thresholds in both languages. In UK English it also restores at least 99% of its placeholders after 760 answers. In Russian it restores only 94.6% [93.6; 95.6] of name placeholders, against a 99% threshold, mainly because the inflection it borrows from pii-guard produces wrong surname forms. The engine was developed over six earlier held-out sets from the same generator; real documents were not tested.

The Voynich manuscript is a fifteenth-century book in a script that no one can read. One account holds that its text means nothing and that a scribe made it by copying his own words and changing them a little. The other holds that a real language lies hidden behind a cipher. We write each account as a program that fills the pages of the manuscript, fit every program with the same budget, and check it on statistics the fit never saw. The copying process of Timm and Schinner predicts these statistics far better than any cipher with one key. But ciphers that switch keys between the manuscript's two dialects, or that encode only the word classes of a real text, predict them equally well. Word statistics alone cannot decide. We look for a trace of language that needs no decipherment. In 28 texts in 16 languages the same word pairs keep sharing a line. The manuscript shows this pattern at half the strength of the weakest of these texts. A word code keeps the pattern above the manuscript unless its plaintext has a weak pattern or many words are miscopied. Solvers that decipher control texts find no readable text in the manuscript under a letter cipher, a word code or 28 codes with one letter, letter pair or syllable per word. Word codes of abbreviated Latin, the plaintext their fits prefer, show which word follows which more clearly than the manuscript unless two words in five are miscopied. One cipher remains possible, a nomenclator with a second key whose scribe miscopies more than half of the words, and our solvers cannot exclude it in 9 of 16 languages. Most of the structure a solver finds comes from how the end of one word shapes the start of the next, which reflects how tightly glyphs are bound to each other. Planted messages are found reliably only when they are long, each letter always becomes the same word and these words are adjacent, so other messages could hide.

As machine translation systems approach human-level quality, traditional evaluation methodologies struggle to detect subtle translation errors. We critically examine limitations in current gold-standard approaches (MQM and ESA), including inconsistencies from variable annotator expertise, excessive categorization complexity, coarse severity granularity, accuracy bias over fluency, and time constraints. To address this issue, we introduce a high-quality dataset consisting of human evaluations for English-Russian translations from WMT24, created by professional linguists. We show that expert assessments without time pressure yield substantially different results from standard evaluations. To enable consistent and rich annotation by these experts, we developed the RATE (Refined Assessment for Translation Evaluation) protocol. RATE provides a streamlined error taxonomy, expanded severity ratings, and multidimensional scoring balancing accuracy and fluency, facilitating deeper analysis of MT outputs. Our analysis, powered by this expert dataset, reveals that state-of-the-art MT systems may have surpassed human translations in accuracy while still lagging in fluency - a critical distinction obscured by existing accuracy-biased metrics. Our findings highlight that advancing MT evaluation requires not only better protocols but crucially, high-quality annotations from skilled linguists.

More people are seeking out general-purpose AI models for medical advice, including diagnosis. Yet, AI models are susceptible to sycophancy, like validating users' beliefs even when they are untrue, which could lead to false positive diagnoses. Here, we evaluate how nine AI models change their diagnostic decision when a patient presents a belief that they have a specific (but incorrect) medical condition. We vary how close the user belief is to the clinical evidence presented. Models sycophantically adopt the patient's belief on average 23.2\% of the time—sometimes even when the patient proposes a fictional, made-up condition. We also varied different personas, including a Rare Disease Claimer that got models to accept their belief 43.0\% of the time. Additional analyses find that medical fine-tuning and model scale makes models less susceptible to changing their diagnoses. Sycophantic bias in medical diagnosis raises implications for safety, especially as more people are using AI for health support.

Text-to-SQL systems achieve remarkable accuracy on academic benchmarks, yet performance drops dramatically in enterprise deployments. This gap stems from a fundamental assumption: existing approaches require complete schema documentation, but enterprise databases typically lack comprehensive documentation, featuring cryptic naming and implicit relationships. We present the first text-to-SQL framework that operates without schema documentation, instead learning database structure from historical query logs. Query logs encode implicit structural knowledge: which tables are queried together, how they join, and what columns map to business concepts. We construct a knowledge graph from parsed query patterns and introduce entity-driven GraphRAG retrieval that grounds user questions in domain terminology rather than relying solely on embedding similarity. Validated on both an industrial deployment with expert evaluation and the Beaver benchmark, our approach substantially outperforms similarity-based retrieval methods, the dominant industry approach, in both executability and answer quality. On Beaver, we achieve 95.7% execution validity compared to 60.8% for embedding-based retrieval, with expert data engineers preferring our system 70% to 10% on internal deployment. Our work demonstrates that text-to-SQL can operate effectively in schema-ambiguous enterprise environments, enabling natural language database interfaces "in the wild".

For agentic data analysis, language-model agents often require a large number of annotated tasks to learn how to interact effectively with executable environments (i.e., databases). While high-quality annotated tasks require heavy manual effort, an executable environment already provides rich interaction structure and verifiable outcomes. In this sense, the environment itself offers a free lunch, an additional set of verifiable tasks that require no human annotation. We propose EnvLunch, a warm-up stage that turns this environmental signal into agent training without human-written or model-generated task descriptions, before downstream agentic reinforcement learning. EnvLunch samples valid action trajectories, executes them to obtain final observations, and trains an agent to recover a functionally equivalent trajectory given only its final observation. This label-free warm-up enables the model to outperform the baseline by over 9 points on average across all downstream training budgets, and with few labels, the baseline needs 8 to 32 times as many annotated examples to match the accuracy that the EnvLunch model reaches at its smallest training budget. Letting the environment alone define the task yields a larger gain than having the model write its own tasks over the same environments, and this warm-up also improves fine-tuning on benchmarks beyond data analysis. EnvLunch therefore saves annotated tasks and provides a better starting point for fine-tuning agentic models when such tasks are limited.

While Multimodal Large Language Models (MLLMs) excel at zero-shot reasoning, their spatial capabilities are primarily evaluated through static, passive observation. In complex, human-centric scenes, occlusion often hides crucial geometric evidence, and passive benchmarks force models to guess rather than recognize their own visual uncertainty. To bridge this gap, we introduce SATE-Bench, a novel interactive benchmark comprising 316 real-world scenes and 2,525 multi-turn QA pairs to evaluate visual self-awareness and active exocentric view-exploration. It combines a Passive multi-view control with an Active exploration setting, where models start from a severely occluded view and must iteratively query adjacent viewpoints until they gather enough evidence to fully ``sate'' the question, scored by a step-penalized efficiency reward. Evaluating state-of-the-art MLLMs reveals that they terminate exploration prematurely while reporting high confidence, and struggle with fine-grained 3D geometric tasks; matched-view controls further attribute the Active–Passive gap primarily to insufficient evidence acquisition. Moreover, fine-tuning with Group Relative Policy Optimization (GRPO) shows that rewarding view-seeking increases active exploration but lowers overall accuracy, as models over-explore and fail to fuse expanding multi-view contexts across turns. These findings indicate that scaling passive multi-view perception does not intrinsically yield active spatial reasoning, highlighting the need for models that know where to look, how to integrate what they see, and when to stop. Code and data are available at https://gtc-gh.github.io/SATE-Bench/

Mobile-use multimodal agents are typically evaluated by task success from a clean initial state, which does not isolate their ability to recover from intermediate errors. Because mobile environments change dynamically, a general mobile-use agent must complete tasks robustly even from unexpected states and in unseen environments. This requires knowing when and where an agent makes errors, and whether it can recover from errors and subsequent states. To evaluate this capability, we introduce AndroidFlux, a failure-recovery benchmark built on AndroidWorld. AndroidFlux restores recorded trajectories to controlled checkpoints before, at, and after the first recoverable error, so that different agents continue from the same device state and history. On these checkpoints, a complementary reward-modeling benchmark evaluates recovery-action preferences using downstream rollout outcomes. We further release an offline reinforcement learning (RL) dataset that improves both error recovery and overall task success. Our evaluations show that recovery performance only partially correlates with clean-start success, and that explicit reasoning helps only models trained with GUI-domain specific RL.

LLMs are increasingly being used by users from diverse demographics with pluralistic values and distinct backgrounds. Maintaining safety while respecting value-laden personas is uniquely challenging because persona-conditioned requests are often not overtly malicious, unlike those in existing safety benchmarks. Alignment methods aimed at adversarial requests therefore may miss the harm that emerges from persona use, leaving models to either refuse legitimate value expression or blindly follow unsafe personas. We construct a dataset of 11k controversial topics paired with value-based personas mapping to value dimensions from the Schwartz Value Survey. Through an agentic pipeline that iteratively refines the topic and persona, we surface and collect harmful responses from well-aligned commercial LLMs with a more than 47% success rate. We use UNESCO's global citizenship principles as a minimal common ground for distinguishing valid value expressions from harmful behaviors, and introduce a controllable alignment model that lets users balance persona adherence and safety at inference time. The resulting model is safer and more persona-adherent than baselines on most samples.

Streaming automatic speech recognition transcribes at low latency but produces an unattributed word stream, while real-time applications such as customer-service analytics need each word tied to a speaker role. Role attribution is usually a separate stage, either an offline or streaming speaker-diarization model aligned to the words or a text-based large language model applied to the transcript, and role labels for training are scarce. We present RoleHear, a semi-supervised pipeline that makes a single cache-aware streaming transducer emit roles inline through lightweight in-vocabulary role tokens: a fast turn-boundary marker followed by a per-role token a few words later. An offline transducer teacher, fine-tuned into a role-attributed recognizer, relabels unlabeled audio, and a multimodal large language model supplies silver role labels where gold annotations are unavailable. The streaming student is fine-tuned on the mixture in a single pass. Across customer-service calls with gold roles and earnings calls with silver roles, RoleHear attributes roles accurately while preserving transcription quality and narrowing the gap to the offline teacher.

Users of conversational assistants rarely state all requirements up front: they add constraints, amend them, withdraw them, digress, and return. Instruction-tuned LLMs fail here in two ways: active constraints stop being honored as turns accumulate (persistence failure), and withdrawn constraints keep being enacted (relapse). The only existing repair we know of is an external ledger that re-injects a compiled specification every turn. We propose StateWise, a self-improving approach that teaches a single LLM to maintain its own revisable constraint state, a typed state with tombstones for revoked constraints, so persistence and relapse are measured by executable checkers and ground-truth states are compiled for free. StateWise fine-tunes the LLM on two complementary skills, summarizing a dialogue into its constraint state and responding conditioned on that state, then self-trains on its own trajectories, keeping only turns that pass an exact round-trip check against the compiled state. This teaches 7-8B models to track the state almost perfectly but, surprisingly, not to obey it: on Llama-3.1-8B every supervised variant roughly doubles relapse. We therefore add a preference stage on self-generated relapse contrasts, where the rejected response enacts a revoked rule and the chosen one passes the round-trip check. On two backbones with held-out constraint families and unseen phrasings, StateWise cuts relapse by 8-22 points relative to supervised self-training, lifts conversation success by 8-19 points over the backbone, and beats an oracle compiled-ledger baseline on relapse with no inference-time scaffolding. We also characterize the recipe's cost on general instruction-following benchmarks and study weight interpolation as a mitigation.

De-identifying clinical notes with a hosted language model discloses the information the step is meant to protect. We present EdgeDeID, a 0.6B-parameter decoder-only de-identifier that runs locally and is trained without real patient text: three LLM teachers write synthetic notes, a single labeler re-annotates them into one convention, and the model learns to emit a typed XML schema of the identifiers it finds. On a 100-note synthetic in-domain benchmark with manually verified annotations it reaches 0.994 micro-F1 and leaves no missed identifier in 99% of notes, against 0.857 and 12% for the strongest of eight off-the-shelf systems. Quantized, it runs on CPU in 615 MB. Of three additions to supervised training, only a verifier adapter helps: it separates complete from incomplete outputs of other systems at 0.94 AUROC, whereas reinforcement learning with a recall-weighted reward and a rule-based post-processing layer leave in-domain results unchanged. On MIMIC ICU nursing notes F1 falls to 0.48. The errors follow two properties our synthetic notes lack, telegraphic style and low identifier density, and three hand-written in-context examples recover 0.08 F1 without annotated target data.

Retrieval-augmented generation (RAG) lets large language models (LLMs) answer knowledge-intensive questions from retrieved evidence, but adapting a general-purpose RAG system to a specialized domain such as medicine is difficult when only an unlabeled corpus is available. Self-training approaches such as SimRAG address this by letting the LLM write its own question-answer pairs from domain passages and fine-tuning on the pairs that survive a round-trip consistency filter. We observe that these approaches treat every retained example identically: they never ask whether the student already knows the answer, already extracts it correctly from the retrieved context, or fails - and if so, whether the failure lies with the reader or with the retriever. To tackle this, we propose TriageRAG, a failure-aware self-training approach that probes the student on each self-generated question with and without retrieved context, triages it into one of four categories (known, solved, context-failure, retrieval-failure), and routes each category to the training signal that addresses it, including a parametric route that teaches the reader to answer when retrieval misses. Probing is cheap and repeated every round, so the training mixture follows the student, and the triage histogram doubles as a label-free diagnostic of where the domain gap lies. Experiments on 11 datasets spanning three domains with a Qwen3-4B backbone, and on CS-Bench with Qwen3-8B, demonstrate that TriageRAG outperforms a budget-matched SimRAG by 1.2-6.4 points and is the only method that recovers the closed-book level in every case where retrieval hurts the base model. A single ablation identifies the mechanism: round-trip filtering removes every training context that lacks the evidence, and a reader that never sees such contexts learns to over-trust noisy retrieval.

Educational taxonomies organize educational content, such as exercise problems, into concepts, which are also called skills or knowledge components. These taxonomies define learning objectives in educational applications such as intelligent tutoring systems. Developed through human expertise and iterative refinement, they also provide fundamental structure for curricula, assessment, and models of student learning. Recent advances in semantic text embeddings make it possible to construct alternative educational taxonomies computationally from the content itself. In this paper, we investigate this possibility by clustering embeddings of math questions to construct embedding-based educational taxonomies. We compare these taxonomies with the existing Common Core State Standards (CCSS)-based knowledge-component taxonomies in two ASSISTments datasets and evaluate their utility in Deep Knowledge Tracing, a downstream task that relies on taxonomic mappings of questions to knowledge components. The embedding-based taxonomies exhibit substantial but incomplete correspondence with the existing taxonomies, revealing both preserved and alternative organizations of educational content. When used for knowledge tracing, they improve predictive performance by approximately 3-5 AUC percentage points. These results demonstrate the potential of semantic embeddings for automatically constructing educational taxonomies that both recover structure present in human-developed taxonomies and provide useful alternative representations for downstream educational modeling.

Large language models increasingly act as partners in everyday communication. In these interactions, useful assistance may require not only answering, but also asking the right questions that surface knowledge gaps, assumptions, and alternatives. In this study, we refer to such epistemic question-asking as model curiosity, and introduce a theory-grounded taxonomy that distinguishes curiosity-oriented questions from generic model questioning. Across four conversation datasets, we find that curiosity-oriented questions remain rare (17-28% of all questions), but are associated with more exploratory and epistemically rich subsequent user responses, with varying patterns across datasets. We further conduct a human study to evaluate the usefulness of model curiosity, finding that curiosity can encourage exploration, especially in advice, creative, and research contexts. These findings have implications for building LLMs that ask more selectively to support deeper exploration and human-AI collaboration.

When an image and its caption disagree, can a user reliably control which one a vision-language model (VLM) follows? Accuracy alone cannot answer this question, since a model might give the correct answer from parametric knowledge without following the requested evidence. We therefore evaluate modality grounding control through conflict-preserving counterfactual interventions. Specifically, Flip measures whether an initially correct answer follows an edit to the requested modality, while Stay measures whether it remains stable when the competing modality is edited. To learn this control, we build on neologism learning by introducing new words for visual and textual grounding, training only their embeddings while keeping the pretrained model fixed. We also introduce Counterfactual Twin-Margin (CTM) supervision, which uses paired interventions to teach how answer probabilities should change when the requested evidence changes. Across three VLMs on three existing benchmarks and our multi-attribute Ground-CF benchmark, we find that learned modality words improve control over natural-language instructions. Surprisingly, a single learned embedding per modality provides strong lightweight visual control and shows little average degradation on held-out values. LoRA remains strongest overall, although its visual advantage narrows under held-out evaluation. Moreover, counterfactual supervision improves control in several settings where standard supervision is weaker. Finally, we test our method on real-world news video and find that visual control improves when the relevant evidence is made easier to access. Overall, our results show that reliable modality control requires testing not only whether an answer is correct, but whether it continues to follow edits to the modality of choice.

Multimodal large language models (MLLMs) have become strong visual reasoners, yet remain limited in open-world, multi-step search that requires grounding images, selecting tools, and integrating web evidence. Existing approaches typically construct and filter fixed offline corpora whose question distributions cannot adapt as the target deep-search agent improves. Building on MTA-Agent's image-grounded QA construction, we present MTA-Zero, an online multimodal QA-Solver co-training framework for self-improving multimodal deep-search agents that requires no precomputed multi-hop corpus. Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online. A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions. To focus generation on the Solver's learning frontier, we derive a Goldilocks reward for the QA Generator from the Solver's group-wise rollout success rate: questions solved by all rollouts or by none receive zero reward, whereas questions near the current frontier are favored. We optimize the two policies independently with GRPO, without sharing parameters or gradients. Across six multimodal deep-search benchmarks, MTA-Zero improves a Qwen3-VL-8B Solver from 36.91% to 51.58% mean accuracy and achieves performance comparable to GPT-5 under the same agent workflow. The same framework also improves a newer Qwen3.6-27B Solver by 10.82 points. These results show that verified online question generation provides a policy-specific curriculum for continually improving multimodal deep-search agents.

Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long time horizons, yet current evaluations largely assess individual task performance rather than improvement through accumulated experience. We introduce Continual Learning Exam (CLE), a large-scale benchmark developed by 50+ domain experts to measure experience-driven improvement in realistic professional workflows. CLE comprises 120 human-authored tasks, each organised as a five-level ladder, yielding 600 levels spanning 89 professional roles across 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on experience accumulated during earlier assignments. An evaluation agent assesses each submission against human-authored rubrics. We compare a single-submission setting with one that provides diagnostic feedback after failure and allows one revision. Across 17 harness-model configurations using frontier LLMs, feedback and revision yield a median completion gain of 5.3 percentage points. Yet the best-performing system clears only 47.7% of levels and completes all five levels in just 23.3% of tasks, highlighting the difficulty of sustained learning from experience for existing frontier LLMs. As such, the dataset can also be used to post-train LLMs to strengthen their continual learning capabilities.

Memory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multiple benchmarks and model backbones, we find many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value. Taking RAG as a reference, we dissect the net value of memory into the gains and losses over direct retrieval, termed memory benefit and memory tax. These gains and losses arise because memory process changes how the raw information is organized and accessed. We view such changes as the double-edged deviations from the raw history. They are originally designed and introduced to benefit the task, for example by reorganizing facts or distilling preferences, but some may also remove or obscure evidence, giving rise to memory tax. Since memory benefit reflects the expected gains of its design, we then focus more on the overlooked memory tax, study how it arises from the system and evaluation side, and generalize the concept to more application scenarios of memory. We hope the notion of memory tax can provide a new lens and shift more attention from standalone performance to the net value of memory, informing more reliable memory designs and evaluations. We will open-source the code and results to the community. Project Page: https://hexuan21.github.io/MemoryTax/

Combining language models is usually done by averaging weights, which works only when the models share a common starting point. We ask whether five frozen, independently trained models with different internal designs (four transformers from four organizations and one state-space model) can be related through their activations alone, with no weight merging. Using permutation tests, an independent replication probe set, depth sweeps, pooling ablations and confound checks, we find statistically robust correspondence between how the five models organize the same text. It is strongest at each model's final layer and weaker for the state-space model, whose weakest pairing is still 35-43 times chance. We use this to build a shared space with Generalized Procrustes Analysis, in which no model is a fixed reference. At the final layer, mean held-out retrieval accuracy over the twenty ordered model pairs is 0.91 to 0.93 on two probe sets (chance 0.03), beating pairwise linear maps (0.82 to 0.85), but the space is worse than those maps in the middle layers. It transfers to ordinary English of unfamiliar grammatical form and fails on mathematical and logical notation. We then test whether the space supports a router that picks among the five models on gradeable tasks. Across four task types we find no case where routing beats always using the strongest model, and we report this as a negative result. A correctness signal that looked strong turned out to be category membership in disguise. All experiments ran on one laptop without a GPU.

Tool-using large language models (LLMs) must decide whether external information warrants action. We investigate whether their expressed credibility judgments reliably indicate behavioral resistance to misinformation. We evaluate five LLMs on 196 constructed false notices paired with legitimate counterparts and 204 archived attacks adapted to simulated tool use. When asked to act on the constructed notices, models issue harmful tool calls in 27%–99% of cases they have rated untrustworthy. Removing the preliminary rating step increases harmful execution by 15 percentage points on average, showing that credibility elicitation itself changes the behavior being evaluated. Interventions that improve credibility discrimination do not consistently yield safer actions. A simple forewarning significantly reduces harmful execution in four of five models, but neither tested generated-example prebunk outperforms it in the main evaluation. These reductions can also compromise legitimate task completion. Supervised fine-tuning on generated false notices reduces harmful execution on held-out archived attacks under the rating-first protocol, with model-dependent utility costs and weaker protection without rating elicitation. Our findings motivate evaluating misinformation defenses through both harmful actions and legitimate task completion, while accounting for the behavioral effects of the evaluation protocol.

Multilingual BabyLM systems usually fix language sampling proportions even though the weakest language can change during training. We present ARMOR-DP, a modular framework that diagnoses imbalance arising from representation, exposure, capacity, and objective pressure, and uses per-language validation loss to update sampling. Controlled experiments show that equal character budgets create a severe representation artifact for Chinese; byte-level encoding removes much of it, while static Chinese protection can subsequently make English the weakest language. The dynamic controller tracks these transitions and achieves a better score–balance tradeoff than fixed protection. On official evaluation, curriculum-dynamic checkpoints outperform a matched static Chinese-protective baseline while using fewer Chinese epochs; scaling capacity yields the strongest internal ARMOR-DP system. These results support a simple conclusion: multilingual weakness is a training-time state rather than a fixed property of a language. ARMOR-DP therefore separates diagnosis from intervention, allowing protection to follow the current bottleneck instead of a predetermined language label.

Hallucination detection is increasingly important for evaluating and deploying large language models. However, existing methods are often compared across benchmarks that define and measure hallucination in substantially different ways, making reported scores difficult to interpret across settings. We study hallucination detection as a benchmark-conditioned measurement problem and organize existing benchmarks along three axes: provenance, which describes how model outputs and evidence conditions are obtained; typology, which captures the factuality and faithfulness failures represented by benchmark labels; and granularity, which specifies the unit at which hallucinations are annotated, predicted, and evaluated. Using this framework, we analyze how benchmark construction, annotation, and evaluation choices shape the underlying prediction task, and relate detector mechanisms and assumptions to the settings in which their performance is measured. We also examine span-level evaluation, where variation in annotation boundaries can substantially affect reported results, and propose Span Coverage F1 as a complementary containment-based metric. Our analysis clarifies what detector scores capture, highlights limitations of cross-benchmark comparisons, and motivates more explicit evaluation of generalization across benchmark conditions.

Depression severity estimation is crucial for mental health assessment and intervention, yet traditional approaches depend on clinical interviews or self-report questionnaires that require users' active participation. With the widespread use of online social media, users leave rich social traces that may reflect their mental states, motivating the task of social media-based depression severity estimation. Existing efforts have progressed from textual-feature-based classification to LLM-based questionnaire completion over standardized scale items, yet they still fall short in user-specific mental priors modeling, diversified semantic retrieval, and fine-grained reasoning over mental evidence. To address these limitations, we propose UMer, a user-specific mental evidence reasoning framework for social traces-based depression severity estimation. UMer first distills a user-specific mental prior from historical social traces to capture the user's psychological tendencies, behavioral patterns, and self-expression style. It then performs scale-guided multi-query retrieval to construct evidence memories from item-relevant posts. Finally, it conducts coarse-to-fine evidence-aware reasoning to distinguish direct and indirect evidence and predict item-level scores with interpretable rationales. Extensive experiments on real-world benchmarks demonstrate that UMer not only achieves competitive performance over existing baselines, but also provides evidence-grounded interpretability and generalizes to other scale-based tasks.

An image is rarely described only once. Retrieval benchmarks and many collections attach several captions to each picture, yet dual encoders still judge every image-caption pair in isolation. We ask what these neighboring captions contribute when a frozen encoder consults them at inference time. Fusing image similarity with similarity to an image's own captions improves text-to-image R@1 by 3.4 to 16.8 points across nine frozen encoders when the captions come from the same annotators as the queries. The gain falls to 1.9 to 7.2 points when an off-the-shelf captioner writes them, and to at most one point when they come from a different annotation source; a paraphrase-trained text encoder strengthens the text side further. A natural refinement, scoring the part of the image that the captions leave unexplained, appears to add much more under a leave-one-out evaluation. We show that this advantage arises because the correct candidate's supports are built differently from those of the negative candidates, with set size and query removal each contributing about half, and that under a fixed support index the residual reduces to a reweighting of what the captions already provide. On compositional benchmarks, the captions help by matching candidates to human descriptions of the image, without improving image-text grounding. The code is available at: https://github.com/nhatkhangcs/clip_consensus

Language-model agents must adapt as users revise their requirements and priorities during complex tasks. We introduce IntentDelta, a human-constructed multi-turn benchmark with 638 trajectories and 3,766 turns across WebShop and TravelPlanner. Each turn pairs a user utterance with a cumulative intention state and human reference actions. The trajectories are cross-verified and cover feasibility- and preference-driven changes in constraints and priorities. We evaluate eleven agents on intention recovery and action quality. Average Intention Recovery is 33.1\% in TravelPlanner and 44.7\% in WebShop, while average Action Score is 31.6% and 41.0%, respectively. Agents struggle to preserve earlier requirements and distinguish soft preferences from must-have constraints. Correct intention recovery also does not guarantee successful execution, highlighting challenges in translating evolving user intentions into actions.

LLM-based evaluators have rapidly become a standard tool for machine translation evaluation, yet the design choices that shape their behavior remain poorly understood. We present a systematic study of how instruction design, few-shot demonstrations, backbone model choice, and multi-step workflows affect span-level and ranking-based MT evaluation. We collect new expert RATE annotations on WMT24 and WMT25 English-Russian translations, with transfer validation on MQM- and ESA-annotated benchmarks across five additional language pairs. Our experiments reveal four robust findings: instruction structure shapes evaluator behavior, high-quality demonstrations matter more than quantity, frontier backbones exhibit distinct evaluator profiles (recall-, precision-, or ranking-oriented), and fusing outputs across diverse backbones outperforms single-evaluator multi-call setups. Although evaluator design exposes a large combinatorial space, several discovered design principles transfer consistently across annotation protocols and language pairs. We synthesize these findings into ORBIT, a family of LLM-based MT evaluators. Our strongest configurations achieve expert-level performance in error span detection and in Accuracy and Fluency ranking at roughly 20x lower cost than expert human annotation.

Large Language Models (LLMs) are increasingly used for medical self-diagnosis. However, the way demographic cues shape outputs in these interactions remains poorly understood. This is particularly consequential for diagnoses like ADHD, where persistent gender disparities are not fully aligned with the clinical evidence. We adapt the Context Association Test (CAT) to evaluate whether GPT-4o responds to gendered ADHD associations in a zero-shot, self-diagnostic setting. We show that explicit cueing consistently amplifies certain ADHD subtype associations while implicit cueing produces near proxy-aligned distributions. These divergences suggest that explicit demographic cues may activate bias-mitigation mechanisms that over-correct rather than eliminate bias, whereas implicit cues appear to reflect the model's default behavior. Our results demonstrate that evaluating gendered bias in LLM-mediated self-diagnosis requires careful attention to the demographic associations a model encodes.

As the applications of AI continue to expand across various multimodal and multilingual domains, ensuring safe and responsible deployment of these systems has become an important concern. A wide-range of interventions have been proposed to make AI responses safer and more reliable. They often incur substantial computational or annotation costs, or cannot be extended to multilingual settings. In this work, we propose Selective Detection And Targeted Editing (SEDATE), a training-free approach based on detect-edit-verify strategy. SEDATE first detects the responsible neurons, edits them as and when necessary and finally verifies the model responses; all without any additional model training. We evaluate SEDATE across 7 languages and 4 model families. We further evaluate its effectiveness across 4 out-of-distribution languages and 4 additional datasets. We show that SEDATE achieves substantially superior safety scores across all languages and models over the best-performing baseline while retaining the general capabilities and incurring negligible overhead. Our results are robust, with the trends remaining similar for out-of-distribution languages and datasets.

Pipelines that require structured outputs from hosted language models often specify a Pydantic schema in the prompt, validate the response, and retry when validation fails. Although native structured-output modes provide an alternative, application-side validation remains important for enforcing custom constraints. Yet pipeline reliability is often summarized by a single validity rate, which obscures why outputs fail and which interventions may help. We introduce a fine-grained taxonomy of schema-conformance failures and a deterministic classifier that assigns failure categories from model responses and Pydantic validation errors. We apply this framework to 3,997 generation attempts across nine OpenAI and Anthropic model/API configurations, using questions from HotpotQA and FinanceBench. Our results show that similar validity rates can mask different failure patterns, and that dominant failure categories vary with schema constraints and task. Markdown-fenced JSON is the most common failure: among 681 fenced failures, only 48.0% pass validation after fence removal, while the remainder conceal additional violations, predominantly field-length violations. In an exploratory prompt ablation, variants that suppress fences appear to shift failures toward length violations, but the shift is consistent with exposing violations that were already present, illustrating how aggregate validity rates can obscure changes in failure composition. We derive four engineering practices for validator, schema, retry and evaluation design, three supported by our data and one, on retry feedback, proposed but not tested, and release the classifier, completions, and annotations.

Dictionary-like definition generators are known to be useful for lexical semantic change detection (LSCD). However, they suffer from overly granular sense distribution, producing excessive number of definitions for one and the same lexical sense. In this work, we show that fine-tuning with Direct Preference Optimization (DPO) helps to overcome this problem. We test our approach both quantitatively on a curated academic LSCD benchmark and qualitatively on real-world data for a practical task of auditing diachronic changes in sustainability terminology within project documentation of EU development finance. All models, code and predictions produced in this study are publicly available at https://github.com/ltgoslo/dpo-for-definition-generation.

Patent examination is a challenging real-world task that requires agents to discover sparse prior art, iteratively refine queries, inspect technical evidence, and produce novelty assessments documented in Office Actions (OAs) across heterogeneous scientific and technical domains. However, existing benchmarks largely focus on static and limited candidate sets, leaving the full search–reasoning–assessment process underexplored. We present PatEx-Gym, the first framework for developing and evaluating agents for patent examination, which integrates three components. First, PatEx-Bench provides 1K realistic examination tasks, two large-scale retrieval corpora containing over 1M patents and 52K scientific papers, and a harness runtime supporting iterative search, document visits, reasoning, and OA generation. Second, we introduce PASS, an annotation-free Prior-Art Agentic Search Trajectory Synthesis framework. Specifically, it exploits document visits as implicit relevance signals and evidence-acquisition anchors to model intra-turn query behavior, inter-turn search progression, and trajectory-level evidence and outcome quality, yielding PatEx-SynData. Third, we train PatEx-Agent at the 4B scale via supervised fine-tuning on the curated trajectories, establishing a lightweight baseline. Extensive experiments show that final decision accuracy alone does not fully reflect examination quality, while prior-art discovery remains difficult even for strong frontier models. Our PatEx-Agent achieves competitive performance against substantially larger general-purpose agents. These results show that training on PatEx-SynData enables a compact 4B model to serve as a competitive baseline for prior-art search and patent examination. We view patent examination as a demanding testbed for agentic search and deep research agents.

Selecting a compact training subset is an important step in adapting medical language models from existing question-answering banks. We investigate perturbation-based data selection using two signals derived from Gaussian perturbations of the final-block representation: prediction-change frequency and non-flipping margin contraction. After a shared entropy prefilter, Representation-Sensitivity Prioritization (RSP) combines the signals using automatic root-mean-square scaling, while RSP-Drop ranks candidates by contraction alone. Both selectors use only questions and answer options during selection. We evaluate them with MedGemma-4B on five medical QA datasets at retained-set budgets of 64, 128, and 256 examples. At the primary budget of 128, RSP and RSP-Drop obtain dataset-averaged Macro-F1 scores of 47.55 and 48.28, respectively, compared with 45.11 for entropy-pool-matched random selection. Their relative performance varies across datasets and budgets. A selected-set case study identifies contrasting clean-margin profiles among the examples exchanged by the two rules, while a full-pool ablation shows that entropy prefiltering is not uniformly beneficial. These retrospective results characterize the empirical utility and limitations of perturbation-based selection under a fixed adaptation protocol. Evaluation uses one backbone and benchmarks with documented training-family overlap; it does not establish clinical effectiveness.

Personal email agents decide, for every message, what it is, whether it is urgent, whether to draft a reply, create a task, or ignore it. Their error rate is rarely measured: the stored labels were written by the system itself. We present Foyer, a self-hosted personal inbox agent in daily use, and a measurement protocol (stratified sampling, frozen snapshots, thread-grouped splits, pre-registered hypotheses, human labels in the agent's review loop) that yields ground truth before any training. On 739 real Italian/English emails labelled by the owner (231 unbiased, 500 adversarially sampled), we compare nine systems under thread-grouped cross-validation: the production keyword rules, a TF-IDF + logistic-regression baseline, local typed decision models (Tev1, Nimble), a hosted decision model (Nimble v4-max), and hosted generative LLMs (Qwen2.5-7B, Gemma 4 26B/31B). Four results. The production rules score 28% on message type against a 58% majority class: the field shown most prominently carried no information. With ~200 labels, logistic regression has the highest accuracy on every field (type 0.79, urgency 0.74, needs-reply 0.83, is-noise 0.90), though it never predicts the two rarest classes and is statistically tied with Gemma 4 on type; few-shot prompting does not close the gap. The sender's relationship (client, supplier, service) is recovered from other emails of the same contact with 98.6% accuracy where the contact has been seen, and is not inferable from a single message by any model: it belongs in the contact graph, not the prompt. Finally, accuracy is not the product metric: composed into actions, the most accurate system most often ignores emails needing action (18% vs 4% for the best LLM); a cost-chosen threshold on its noise probability halves that rate and makes it the cheapest system.

Despite advances in multilingual text representations, effective dense retrieval for Persian remains constrained by the limited availability of language-specific training resources and compact embedding models. We present Yalda Embedding, a 595.8-million-parameter Persian retrieval model adapted from Qwen3-Embedding-0.6B using more than one million query-document pairs. On eight Persian retrieval tasks from FaMTEB, the model achieves the highest macro-average among the evaluated 0.6B models, scoring 0.5488 and outperforming Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task. It also approaches Qwen3-Embedding-4B, winning on four tasks, tying on one, and trailing by 0.0075 in macro-average. An incremental ablation study examines contrastive batch size, optimization stabilization, and relational knowledge distillation from F2LLM-v2-8B. Under the one-epoch training setup, contrastive fine-tuning alone yields a modest mean improvement over initialization and regressions on two tasks. The final distilled configuration improves all eight tasks over the non-distilled model, raising the mean from 0.4813 to 0.5488 without increasing inference-time model size. These gains include hard-negative evaluation tasks despite training without explicitly mined or supplied hard negatives. We release model weights, training and evaluation code, the pairwise dataset, and its teacher-embedded counterpart, enabling inspection of data composition and training procedures, independent evaluation, and reuse without repeating expensive teacher inference.

Training large language models (LLMs) for multi-turn strategic persuasion faces two coupled obstacles: (i) realistic simulation of psychologically grounded persuadees, and (ii) dense, faithful supervision for credit assignment in strategic dialogues. Outcome-only rewards are sparse and uninformative about which conversational moves drove success, while LLM-judge process rewards are expensive and decoupled from the underlying interaction dynamics. We address both challenges with a unified framework. First, we introduce PersuArena, a persuasion simulation environment that produces realistic, non-monotonic trajectories. Second, building directly on this environment, we propose Process Rewards via Online Belief Elicitation for Relative Policy Optimization (Probe-RPO), an online RL framework that converts the persuadee's internal state into dense process rewards. At each turn, three introspective probes are read from the frozen persuadee's logits, yielding a multi-dimensional reward signal that is faithfully aligned with persuasion dynamics. With extensive experiments on PersuArena spanning healthcare, financial advising, workplace management, education, and personal relationships, we demonstrate that our Probe-RPO achieves the full acceptance rate of 52.5%, compared with the 35.8% for the strongest GRPO ablation and 25.8% for the best off-the-shelf LLM. We also conduct comprehensive analyses on the generalization across persuadee models and the persuasion dynamics across conversations. Our work offers a promising paradigm that can provide effective process supervision for strategic dialogue.

This article argues that instruction-fine-tuned language models should be understood as implementing a genre, what is proposed as the instructional-conversational genre. This genre is an amalgamation of discursive language from prior conversations between generative models and humans recursively run through these massive neural networks. It is instantiated in applications through instruction fine-tuning, the inclusion of a pretext known as a system prompt, and the formalization of inputs in mark-up languages characterized by the use of special tokens and role assignment in turntaking structures.

Automated semantic complexity estimation remains underexplored in NLP, despite important applications in linguistics and interpretability. This paper defines semantic complexity as the code length required to uniquely specify a text's semantic meaning, and operationalises it via SCAPE-T, an attention-derived metric based on encoder-only transformers. Existing measures either model reader effort (readability, lexical difficulty), which correlates with surface features (word length, syntax), or quantify surface features directly (type-token ratio, word frequency). SCAPE-T targets semantic integration directly by combining three submetrics: (i) inverse-weighted self-focus (token-to-self), (ii) CLS focus (token-to-CLS), and (iii) inter-head redundancy (cosine similarity across heads), using layer-weighted aggregation; data-driven learning is confined to hyperparameter tuning for calibration. Development uses philosophical prose vs. MCTest stories; evaluation is on children's stories vs. technical Wikipedia articles, with pairwise accuracy on contrasts as the primary criterion. Baselines include readability, syntactic/lexical, and LLM-derived measures. SCAPE-T substantially outperforms all tested baselines under word-length control, with a pairwise accuracy of 78.5% (95% bootstrap CI: 76.1-80.4, N=10k pairwise comparisons; +12.7 percentage points over the strongest baseline); without that control, readability heuristics overtake, consistent with their reliance on surface features. On a short-terms abstractness dataset, SCAPE-T also leads. Ablations show each submetric contributes materially. Overall, the results show SCAPE-T isolates conceptual structure from lexical/syntactic confounds better than existing methods, and generalises across two held-out domains.

Can AI agents autonomously train useful language models and improve them using feedback from users? We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement. Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. We evaluate their models on 20 benchmarks and through blind human evaluations. All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.

French users may omit apostrophes and diacritics or make keyboard typos in everyday input. The effects of these common variations on language model performance and prediction reliability remain insufficiently understood. To address this gap, we construct a controlled French surface-variant benchmark covering seven task types, with 4,200 original-low-high triplets (12,600 input instances). Each high-intensity variant retains all low-intensity edits and adds further changes. We evaluate six models from two LLM families. The results show that accuracy decreases and Brier score and negative log-likelihood (NLL) increase with perturbation intensity. Larger models show smaller accuracy losses under high intensity, but increases in Brier score and NLL do not consistently decrease with scale. Keyboard typos cause the largest losses in accuracy and the quality of predicted probabilities at the predefined high-intensity settings. Finally, calibration temperatures fitted on original inputs transfer effectively to surface variants. Combining calibration with input restoration further improves Brier score and NLL over either method alone while preserving restored accuracy. Our benchmark provides a unified framework for jointly evaluating task accuracy and prediction reliability under common French surface variations.

Linguistic productivity, as evidenced by the use of words in novel contexts, is one of the most striking facts about human language. However, how abstract the knowledge that enables such productivity is remains unclear. We compare people's and computational language models' linguistic productivity in contexts where successful generalization depends on form-based information. We focus on extending the use of a novel verb learned in the active voice to passive contexts. We find people extend active to passive uses in a way consistent with abstract generalizations concerning linguistic form, even when no plausible meaning can be assigned to the novel verb. In contrast, large language models succeeded only when a meaning like an existing verb's could be assigned to the novel verb. We conclude that computational models' syntactic knowledge remains tied to the distribution and meaning of specific lexical items in a way that human syntactic knowledge is not.

Natural Language Autoencoders (NLAs) express a model's internal activations in the form of readable text. The verbalizer turns an activation into an English explanation; the reconstructor then converts that explanation back into an activation, and the explanation is considered successful if the round trip brings back the original activation. We claim that this criterion evaluates whether the information is sufficient and does not address descriptive faithfulness. A sentence can contain sufficient information to make it possible to identify an activation even if it wrongly explains what the activation does: in the case of reconstruction, the primary aim is to accurately identify the activations and not to describe them. This issue has the same structure as the chain-of-thought faithfulness problem, in which verbal output that has been trained to go with a process is taken as a report on that process, and it parallels the dissociation between an individual's introspective report and the actual governing process as documented in human psychology. Although proposed as interpretability tools for the purpose of alignment, NLAs reintroduce, through unverified explanations, the epistemic problem that they were intended to resolve.

Large language model (LLM) agents work in scenarios with very different memory needs. A question about past conversations needs the exact facts it names, while a multi-step task needs a trace of the steps already taken together with details that the next step cannot name. Existing memories are each designed and tested for one kind of scenario. Retrieval memories such as A-Mem miss steps that the current query does not express, summaries drop exact strings, and pairing the two fixes the split of the budget in advance. We propose FlexA-Mem, a general memory that writes each unit of history once into an immutable store and renders a fresh memory for every call. The rendered memory is a compact timeline of records plus detail that a fixed rule fits into the remaining room without an LLM call. The detail is either relevant units as raw text or excerpts or the verbatim pieces that best match the query. With one store, FlexA-Mem serves both kinds of scenario. On AppWorld challenge tasks at 4k, it outperforms the second-best method by 11.5%, and on 147 unused tasks with predictions fixed in advance it outperforms A-Mem with a summary by 12.2%. On LongMemEval and LoCoMo, it matches a dedicated keyword retriever that writes no memory, while memories built on text written by an LLM fall far behind at tight budgets. Ablations show that the timeline carries AppWorld completion, whereas recovery mainly saves steps there and decides accuracy on exact facts. Code is available at https://github.com/WujiangXu/Flex-Amem

For agricultural advisory, generating an answer is only part of the problem: the system must also control unsafe requests, work with limited connectivity, and accommodate changing guidance. KRISHOKTECH is a Bengali agricultural advisory system built around an explicit authorization boundary around generation. Before retrieval, normalization and rule-based admission checks refer policy-covered high-risk requests and ask for a crop when treatment intent is underspecified. The system then binds retrieval to a single crop and asks for confirmation when the text and image indicate different crops. After generation, a narrow verifier removes dosage claims not supported by retrieved BARI/BRRI evidence. Agricultural guidance is kept in an external evidence index that can be updated without retraining the generator, while browser-side vision and local retrieval reduce connectivity requirements. The interface makes these decisions visible as ASK, CONFIRM, REFER, and DROP, together with their evidence. On 200 farmer queries and 400 stress queries, the controls halted underspecified requests and reduced cross-crop retrieval violations from 36.5% to 0% under the tested alias filter. Mismatch detection caught 453 of 454 text-image conflicts, while the verifier caught 114 of 118 inserted dosage errors. The demonstration shows how these controls govern the path from request to delivered advice.

In LLM-agent opinion simulations, each agent reports the probability that a claim is true after seeing its own previous probability and the probabilities reported by its neighbors, and that number becomes its state in the next round. If the number describes the agent's position rather than the sentence it was shown, then rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, so the two reports should sum to one. We test this with three small open models: Qwen3-4B, Llama-3.2-3B and Gemma-3-4B. On 2,304 agent states taken from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they again exceed one. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it persists under repeated sampling and across peer group sizes and report strengths. When the reports are fed back into two populations that start from complementary states, the populations do not remain mirror images: their divergence is positive after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction and lowers Brier loss averaged over the two wordings, but at the final round it beats the better of the two original wordings for only one of the three models. A schedule-by-targeting experiment on the same model gives a small, stable interaction because such sweeps keep the claim wording fixed. Consistency under rewording therefore needs its own test before simulated probabilities are read as beliefs

It is essential that we communicate with the entirety of an affected population during a crisis, including those who do not speak a region's dominant language(s). Commercial Translation tools offer a tempting solution to broaden access and can provide a lifeline when no other resources are available. However, Emergency response managers often operate in the dark regarding the quality and adequacy of these automated outputs. In this paper, we conduct a nationwide survey of emergency managers to understand their current workflows, perspectives, and operational challenges. We then evaluate commercial translation tools against these needs, shedding light on their utility and offering practical guidance on how and when these tools can be safely used in emergency response.

Hospitals regularly buy software that clinicians and patients depend on. Because these purchases are costly, hard to reverse and can affect patient care, buyers look for evidence of which peer institutions already use a product and which systems it connects to. This evidence is public but scattered across vendor websites, press releases and third-party pages. Language-model agents can now collect it at low cost, but the correctness of the extracted claims is difficult to establish, and a wrongly listed customer can mislead buyers and misrepresent an organisation. We present a pipeline that builds a knowledge graph of customer and integration relations from the open web under one rule: every edge must be supported by a verbatim quote from its source page. Pages come from a deterministic crawl of each vendor's site and a budgeted ReAct agent that searches the rest of the web. A permissive extractor proposes relations with quotes, and a layered grounding stage decides which to keep, using a referent step that reads the whole page when a quote names only one side, deterministic checks, and a panel of five model calls voting on nine criteria. On 61 vendors, the pipeline reaches 0.94 precision and 0.71 recall against a page-level key covering the 6,428 candidate relations proposed by its extractor, whose verdicts were drafted by one model and re-checked with a second. Against the golden reference, the marketplace's own curated list, it scores lower, but 89% of the relations it keeps that the reference lacks are stated on a page. A confidence score built from the panel's votes keeps more correct relations than the fixed rule at the same precision, and reaches 0.90 precision at 0.90 recall. We also discuss how narrow, bounded model calls, string-matched evidence and voting address long-context degradation and run-to-run variance.

Repeating Transformer layers can improve language-model performance without adding parameters. A pretrained model can be adapted to support both single-pass and looped inference by reusing middle-layer weights for additional computation. Because both paths share weights, training to improve looped predictions can degrade single-pass quality. We study joint training strategies using weighted losses on both outputs, with guidance from the frozen original model. Joint supervision substantially reduces single-pass degradation compared to looped-only training, while preserving looped gains over ordinary continued training. Most of this benefit comes from adjusting the single-pass loss weight, with guidance from the frozen original model further improving the looped output. With validation-selected weights, our method preserves single-pass quality within a predefined tolerance of ordinary continued training while improving looped predictions.

Modern vision encoders are widely used as semantic representations, but the relationship between embedding similarity and human-visible semantics can become unclear far from the natural-image manifold. We probe this regime in a black-box setting by applying sparse pixel edits to blank canvases with a simple pixel-level hill-climbing procedure. Across ImageNet and Flickr30k, we evaluate whether these edited blank canvases can retrieve chosen targets in multimodal contrastive encoders and self-supervised visual encoders. On OpenCLIP, inputs optimized toward class prompts achieve strong top-1 and top-5 targeted retrieval performance; on DINOv3, inputs optimized toward class prototypes achieve high target alignment and near-top prototype ranks. These results show that sparse edits on otherwise blank, visually non-semantic inputs can enter target semantic basins in both language-aligned and self-supervised visual representation spaces. Our findings identify a shared off-manifold failure mode of modern vision encoders and suggest sparse blank-canvas optimization as a simple probe of representation geometry.

Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a target, the student inherits the teacher's errors. We show that using the same pseudo-labels as a reward reference for Group Relative Policy Optimization (GRPO) is more robust to label noise. In the banking domain, GRPO yields 29.3% to 56.2% relative WER improvements over SFT on Whisper-generated labels, and the advantage holds across four Whisper sizes, four teacher families, three domains and four languages. While SFT reproduces almost 100% of teacher errors verbatim, GRPO passes through only 27%. Rollout statistics explain why, erroneous pseudo-label words rarely appear in on-policy rollouts, so group-relative advantages stay aligned with the clean transcript, whereas SFT forces them into the policy's support and collapses the reward variance GRPO relies on.

Do preferences Large Language Models (LLMs) express in standard elicitation tasks predict their strategic behaviour? LLM agents interacting with humans and other AI systems face potentially conflicting objectives, such as individual welfare, group welfare, and environmental concerns. Prior work characterizes LLM preferences through behavioral-economic and psychological elicitation or strategic games, but rarely tests consistency across settings and prediction in other strategic environments. This requires modeling how competing objectives jointly shape behavior, whereas existing approaches typically measure preference dimensions in isolation. Drawing on utility theory, we introduce the Preference-Behavior Model (PBM), which combines a preference representation with mechanisms capturing imperfect utility maximization. We formulate estimation as an inverse-game problem, jointly inferring preferences and behavioral parameters from observed choices. We evaluate PBM in three social dilemmas that expose conflicting motives: a Public Goods Game, a Collective-Risk Dilemma, and a Common-Pool Resource dilemma. We find that while preferences estimated from established behavioral economic elicitation instruments do not consistently predict behavior in these dilemmas, models fitted in one dilemma retain predictive value in others, suggesting that inferred preferences capture meaningful aspects of LLM decision-making that elicitation instruments may miss. Finally, we assess alignment with persona instructions, finding that inferred preferences broadly reflect intended priorities with specific divergences across dimensions, LLMs, and games being identified by our approach. Our methodology provides a framework for studying how consistently LLM preferences explain behavior, assessing alignment with intended priorities, and revealing trade-offs and behavioral patterns that standard elicitation may miss.

To understand the emergence of grammatical behavior in language models, we track circuits for subject-verb number agreement in pythia-1b and ten seeds of pythia-410m across training. We find that subject-verb agreement is learned early, and is initially handled by a diffuse collection of neurons. However, after agreement behavior has stabilized, the mechanisms responsible for it continue to reorganize. The attention heads which mediate agreement often shift during training: in the case of pythia-1b, behavior is controlled by two heads early in training but by a third (different) head later. Furthermore, the circuit shrinks over training, with pythia-1b's final circuit being strikingly sparse. We take these results to be informative about how linguistic knowledge emerges and is represented in a language model.

Fine-tuning on datasets of unknown origin can create a backdoor in the trained model when the data includes poisoned examples. Difficult to localise and remove, backdoors can persist through safety training. While parameter-efficient fine-tuning methods capture the full parameter changes introduced during training, they are not easily interpretable. We present the full method of TopKLoRA, extending the preliminary proposal of Masiak et al. (2025), an interpretable-by-design LoRA variant in which only the $k$ largest adapter latents are active at each token. Unlike SAEs, TopKLoRA is trained purely on downstream loss, where optimisation pressure incentivises learning functionally causal latents. We demonstrate this sparse LoRA adapter in a sleeper-agent backdoor case study where we create model organisms by fine-tuning gemma-2-2b, Qwen2.5-1.5B, and Qwen2.5-7B on a poisoned instruction-following dataset. For each organism, we identify and ablate a circuit that is both necessary and sufficient to implement the backdoor. Furthermore, our new fine-tuning method produces sparse circuits that are up to an order of magnitude smaller than a dense LoRA baseline, allowing surgical ablation of the malicious capability without sacrificing instruction-following performance. Finally, we propose a technique to construct model organisms with localised circuits using TopKLoRA latents and gradient-routing, which we use to validate the findings of our circuit search algorithm. We consider these findings as a step toward interpretable-by-design architectures that provide faithful units of analysis without compromising model performance. We share our code: https://github.com/interpretable-finetuning/TopKLoRA.