LangTrend

Emotion classification from text is challenging because many affective states are distinguished by subtle linguistic signals and semantically overlapping boundaries. We investigate whether compact quantum modules can serve as targeted residual specialists that assist traditional classifiers on difficult emotion decisions. We introduce EmotiQ, a residual specialist framework for text emotion classification using a frozen transformer classifier as the primary decision maker, identifies challenging emotion boundaries from validation behavior, and trains boundary specialists, including quantum encodings, data re-uploading specialists, prototype and fidelity specialists, and kernel-based specialists. Specialists are permitted to intervene only when they satisfy validation-based admission criteria. We evaluate the framework across six text emotion classification datasets and establish traditional model baselines. Experimental results show that the admitted EmotiQ policy improves mean macro-F1 metrics over the baseline, and achieves positive gains on five of six datasets. Among the quantum components, ZZ interaction feature maps are the most consistently effective, yielding a mean macro-F1 improvement of +0.0030 and positive gains on four datasets. These findings suggest that current quantum NLP methods can be deployed as residual specialists that provide localized improvements under strong classical supervision.

Figurative language remains a major challenge for Machine Translation (MT) systems, as the intended meaning of such expressions often goes beyond literal translation. Although recent advances in Large Language Models (LLMs) improved the translation of Arabic dialects, benchmark resources for evaluating figurative language translation in Moroccan Darija remain scarce. To this end, we introduce the Moroccan Darija Figurative Language Corpus (MDFC), a benchmark dataset for MT and multiple-choice question (MCQ) selection covering idioms, metaphors, and proverbs. Using MDFC, we evaluate 19 open- and closed-weight language models and MT systems at both the expression and contextual levels under zero-shot and six-shot settings. We also conduct layer-wise similarity-based probing to determine how model representations distinguish figurative from literal meanings. Our findings reveal that closed-weight models achieve the strongest translation performance, while the benefits of contextual information vary across models and tasks. Differences between translation performance and MCQ accuracy highlight the distinction between generating an appropriate figurative translation and recognizing the intended meaning among candidate translations.

Moore's paradox \citep{moore1942} concerns sentences such as ``It is raining, but I don't believe that it is,'' which may well be true yet cannot be sensibly asserted. The absurdity lies not in the content but in the act, since to assert is to vouch for one's own belief. This makes the paradox a test of whether language models take the speech act into account when computing meaning. We present the first test of Moore's paradox in LLMs, evaluating 12 models on a curated set of Moorean sentences in varied contexts. We find that models largely recognize the infelicity of Moorean sentences, and that the judgement is not tied to a memorized form and tracks the speech act. Using the Rational Speech Act framework (RSA; \citealt{goodmanfrank2016}), we then test whether these judgements arise from a listener reasoning about a speaker. We find the models' own listeners do not reason about their own speakers, and their reasoning traces suggest that they instead recall knowledge about the paradox and apply rules therein. The models thus appear to have learned the paradox as a fact about sentences, not as a requirement on speakers.

Previous work has argued that multilingual LLMs encode language-specific features most strongly in their initial and final layers, implying that intermediate layers support language-agnostic representations. This paper reconsiders this finding by performing four different experiments: three that undertake a fine-grained analysis of representations of Portuguese and Spanish, two closely related languages, and one large-scale language identification experiment that covers over 200 languages. The results of these experiments challenge the strict separation between language-specific and language-agnostic layers in multilingual LLMs, using Llama 3.2 1B as a test case. By analyzing and manipulating contextual sentence embeddings, we demonstrate that language identity remains strongly recoverable throughout the model rather than being restricted to initial and final layers. Moreover, we find that the model tends to classify ambiguous strings as closer to Spanish than to Portuguese, revealing a skew in model behavior that differs from human judgments.

This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world's fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we investigate how nationalities, ethnic groups, and gender identities are portrayed in pornographic discourse, focusing on stereotypes and deviations from implicit norms of pornography. The results show the dominance of American pornography over other groups, which acts as the de facto norm for the adult industry. We identify persistent ethnic stereotypes reflecting colonial power dynamics, religious fetishization, and the sexualization of cultural taboos. The analysis reveals gendered representational differences across cultures, with a prevalent Madonna-Whore dichotomy in the portrayal of women and stereotypical depictions of men in physically dominant roles.

Agentic knowledge graph question answering (KGQA) requires large language models (LLMs) to solve complex questions through multi-turn interactions with a knowledge graph (KG). Outcome-supervised reinforcement learning derives verifiable outcome rewards from final answers but uniformly assigns each rollout's outcome-level advantage to all policy-generated tokens, creating a trajectory-to-operation mismatch since KG operations within a rollout can have varying effects on the search process and final answer. To address this, we propose Outcome-Anchored Credit Reallocation (OACR), which employs a training-only KG-state outcome model to estimate outcome-reward changes between adjacent KG search states and converts them into operation credit, from which token-wise weights are derived. These multiplicative weights reallocate token-level update magnitudes while preserving the sign of the outcome-level advantage at every generated token. Across three benchmarks, OACR achieves stronger overall performance than existing KGQA methods while using fewer inference-time LLM calls and tokens.

Agentic LLMs increasingly support scientific prediction by grounding decisions in external evidence. In biology, they promise accurate, interpretable predictions for important problems by integrating complex experimental and computational evidence. However, serialized tool results can contain hidden label shortcuts, such as dataset identifiers or template artifacts, that create apparent success without biological reasoning. To address this risk, we introduce a two-stage audit. Stage 1, Per-Tool TF-IDF Decomposition (PT2D), localizes label-predictive patterns to individual tool results while avoiding signal dilution in the concatenated tool trace. Stage 2, Token-Class Redaction, counterfactually tests whether the frozen LLM depends on the localized shortcut. We apply this audit to TCR-pMHC binding, where tool augmentation substantially outperforms a no-evidence baseline. At face value, this suggests that the LLM has learned to use scientific evidence. The audit overturns this interpretation: PT2D localizes the strongest signal to identifiers in one tool result, and redacting them reduces performance to chance. These findings show how hidden label shortcuts can masquerade as scientific reasoning or discovery, motivating source-level audits before benchmark gains are trusted.

Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization. Designing effective tools to support these efforts requires taking into account the community-, language-, and document-specific context. We investigate how small ($\leq$3B) vision-language models (VLMs) can be adapted as optical character recognition (OCR) tools for Lakota, a North American Indigenous language. In addition to benchmarking modeling approaches, we inspect four practical facets of low-resource archival OCR that arise in previous work and conversations with community members: constrained training budgets, segmentation noise, scan degradation, and orthographic variance. We demonstrate that fine-tuning VLMs for OCR post-correction on 35 pages of transcribed data decreases character error rate (CER) from 17.1 to 3.9\% on a bilingual corpus, and find that training only with nine pages increases CER by only 1.1\% over this. However, evaluating under realistic automatic segmentation nearly doubles the CER of the best-performing models, stressing the importance of taking realistic use cases into account. We hope this work encourages future research on developing language technologies grounded in the workflows and constraints of ongoing revitalization efforts.

Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text—especially on tasks where visualization would help, like geometry questions posed in words. We distinguish weak multimodality (the model can answer questions about an input image) from strong multimodality (the model's text representations themselves become vision-aligned, improving performance on text-only tasks with latent visual structure). To test for strong multimodality, we construct a task of shape-counting problems with paired equivalent text and image specifications. This task is trivial when the image is provided, but requires visualization when specified in text as line coordinates. We compare instruction-tuned Qwen-2.5 and Mistral models against their vision-aligned counterparts on the text variant of the task. We find that the multimodal models do no better in this setting, despite strong performance on the image version of the task. Probing experiments corroborate this: vision-aligned decoder embeddings carry no more visual structure than their text-only counterparts. Finally, we provide an optimization objective that distills in-context visual embeddings into text embeddings. This alignment induces gains on the text variant of the task without explicit task supervision, even on samples more complex than seen during alignment. This suggests that VQA-style alignment leaves available cross-modal transfer untapped.

Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose ArabPref, which to our knowledge is the first large-scale pan-Arab preference data set for the alignment of cultural behavior conditioned on the country. Our contribution includes two resources in English and Modern Standard Arabic: a culturally grounded preference dataset for training and evaluation, and a multiple-choice benchmark designed to assess culturally aligned behavior across nations. All test data are validated by native speakers to ensure authenticity. To optimize cultural alignment, we experiment with post-training methods across multilingual and Arabic-centric language models, evaluating performance on both generation tasks and our cultural multiple-choice benchmark. By releasing an evaluation benchmark, we provide a foundation for systematically studying and improving culturally aligned behavior in language models.

Small language models have improved to the point where a phone can plausibly screen its own calls, answering on an absent owner's behalf without sending call content off the device, a privacy-respecting alternative to cloud call handling. Realizing that promise takes a model that can do two things well: reply within the owner's authority during the call, and afterward tell the owner what happened and what needs attention. Released small models do neither reliably - they raise unwarranted alarms about ordinary callers and write summaries that drop material conditions or report a caller's claims as fact - and real calls supply no reference replies or summaries to learn from. We construct supervision for both tasks from synthetic calls under an information-boundary constraint: a caller writer works from a private scenario brief, a reply writer sees only the conversation up to the current turn, and a summary writer sees the completed transcript. Separate model-based reviews of the full dialogue and the summary decide which calls enter training, yielding a corpus of 3,395 calls with 15,102 targets (11,707 reply, 3,395 summary). Training a single shared LoRA adapter on both tasks, we evaluate six models (0.6-4B parameters) before and after adaptation on 192 held-out calls. Reply qualification rises from 21-54% to 94-97% and summary qualification from 0-6% to 43-72%; all twelve paired gains (six models x two tasks) have bootstrap intervals above zero, and the summary gains hold in every source family. Because one model both screened the training data and scored the outputs, these rates measure agreement with that evaluator's standard rather than human acceptance, which a planned human study will test. After adaptation, coherence and naturalness failures nearly vanish, leaving grounding and usefulness - faithfully tracking who said what, under which conditions - as the remaining challenge.

Deterministic hash-addressed conditional memory permits local post-hoc knowledge injection. Extending arXiv:2610.21553's bounded-feasibility audit rejecting under-specified generality claims with repaired probes and training-dose/capacity axes, we give, to our knowledge, the first production-scale write-content verdict. We test 124M-7B self-trained/publicly pretrained beds, Qwen3.8-Flash-Next-FP8 (176.95B; 320,001,536 FP8 rows), and DeepSeek-V4.1-Flash. Write coverage requires a sufficiently trained backbone prior, full-width readout, trainable gates; undertrained-bed intervention confirms a threshold, not a slope. At fixed budget, coverage falls from 98% (50 facts) to 12.7% (1,000); table size is not binding. Two-hop composition does not exceed the independent single-hop baseline (26% observed vs 25% expected; 15% lexical-distortion floor). Synergy remains absent across dose/load in two independent bed families. A 28-percentage-point in-context gap (given vs recalled intermediates) localizes failure to retrieval-for-reasoning, not gate access. Writes are independently retrievable, precisely addressable, reversible atoms, not composable into reasoning. A layered protocol, the first, to our knowledge, to cover compositionality in behavioral memory evaluation, separates baseline, synergy and distortion floor. Any addressed/key-value memory must report coverage, synergy ratio and distortion rate before deployment claims. Composition must be system-orchestrated; multi-constraint queries should explicitly chain written atoms. In-context controls show given intermediate entities are exploited at twice the implicit rate. Production load fits exponential interference ($R^2=0.98$); its mechanism, between budget dilution and write interference, remains open. All numbers are recomputed from archived raw logs/JSON; a fresh 52-target production rerun gives 52/52, exactly matching archived baselines.

Text processing is crucial for low-resource languages as it enables the development of tools and technologies that bridge the digital divide and provide equitable access to information and services. By empowering low-resource languages with robust text processing, we can preserve linguistic diversity, enhance education, improve communi- cation, and foster inclusivity in a rapidly digitizing world. Due to the lack of sufficient annotated datasets and lin- guistic resources, it is challenging to develop accurate and reliable language technologies for these low-resource languages. Large Language Models (LLMs) have attracted sig- nificant attention with their performance and the value they produce. However, with the ever-evolving language, adapting LLMs for a specific task is time-consuming and resource-intensive. Thus, Retrieval-Augmented Genera- tion (RAG) has overcome this limitation by augmenting LLMs with external knowledge, thereby improving their efficiency and extensive task-specific adaptation. Most applications of the RAG technique have been limited to English and has seen limited application in In- dic languages. This paper introduces the RAG-powered AI-Tutor application, Acharya 2.0, designed to produce academic assistance to Hindi-speaking students. The client-server-based web application integrates many fea- tures to support students in accessing and interacting with educational content.

Background and Objective: Clinical free text cannot leave the hospital without de-identification, yet no de-identification system or benchmark exists for Greek clinical text. We evaluated Morpheus, a fully local de-identification pipeline for Greek clinical documents, on 500 synthetic discharge summaries. Methods: The notes (300 cardiac surgery, 100 cardiology, 100 nephrology) were generated by a frontier large language model (Gemini 3.1 Pro) constrained to authentic Greek hospital document layouts. Notes were pre-annotated at generation; two clinicians then independently reviewed and corrected every note, blinded to each other, with disagreements arbitrated. Primary outcomes were recall over must-redact identifiers and over-redaction of clinical content, measured over must-survive probes. Secondary outcomes: residual leaks flagged by an independent egress-review layer, processing time and reproducibility. Results: Inter-annotator agreement was F1 = 0.998. The de-identification rate was 96.59% (7,666/7,937 identifier instances; 95% CI 96.16-96.96): 98.49% for direct and 93.42% for quasi-identifiers. No patient's own complete identifier survived in any note; of 271 residual instances, 199 (73.4%) concerned institutional identity, chiefly the hospital letterhead. The over-redaction rate was 2.82% (190/6,747 probes; 95% CI 2.45-3.24). The egress guard flagged 2 of 271 residuals, both complete telephone numbers. Two independent runs gave identical verdicts. Conclusions: A fully local, deterministic pipeline removed 98.5% of direct patient identifiers from Greek clinical text while altering under 3% of protected clinical content; most residuals concerned the institution, not the patient. To our knowledge this is the first benchmark evaluation of clinical-text de-identification for Greek and the first quantified evaluation of an independent egress-guard layer.

Reinforcement Learning with Verifiable Rewards (RLVR) can improve reasoning in large language models, but specialized non-English domains expose a data mismatch. Raw English domain text is often available, while target-language RLVR tasks with verifiable rewards are scarce. We introduce code-switched task synthesis for cross-lingual RLVR, a data-centric approach that derives target-language and code-switched RLVR task variants from otherwise unverifiable English domain text. Using fill-in-the-middle multiple-choice QA as a seed task format, we synthesize automatically verifiable tasks from domain passages and reshape them into cross-lingual variants. These variants preserve the same evidence-answer relation but change how English and target-language evidence are presented during RLVR. On Korean medical benchmarks and cybersecurity benchmarks across four target languages, code-switched training outperforms matched monolingual and bilingual-mixture baselines on average target-language performance while maintaining comparable English performance. Additional analyses provide evidence of better target-language input understanding and stronger cross-lingual alignment in both model outputs and internal representations. These results suggest that code-switched training helps target-language inputs access domain reasoning learned more strongly in English.

Large-scale conversational recommendation datasets typically require manual dialogue construction or annotation. We propose SearCl, which automatically generates conversational recommendation dialogues grounded in user search and click behaviors. SearCl segments user interactions into query sessions and derives each recommendation target from content selected after a search query. A configurable probabilistic state machine controls the dialogue flow, while two LLM agents acting as the seeker and recommender generate utterances according to turn-level constraints. To support recommendation, the recommender is provided with knowledge graphs constructed from metadata about the content items used in each dialogue, capturing their attributes and relationships. Using SearCl, we generate 3,538 dialogues from content search and click interaction data, covering 1,731 unique target items. Under an LLM-as-a-judge protocol, the resulting dialogues achieve higher quality scores than dialogues from existing benchmarks. These results demonstrate that SearCl enables controllable and knowledge-grounded dataset construction without human annotations.

System One models, introduced with TypeSafe's Jev, answer typed questions about a state with calibrated probabilities instead of generated text: choosing among options described in natural language (Choice), placing the state on an ordered scale (Score), or judging a yes/no statement (Noul). Jev is general-purpose: one set of weights, served as an API and adapted to a domain only through the request. We study the other end of the design space. A Narrow Decision Model (NDM) offers the same interface but is trained for the decisions of one organization, is small enough to run on its own infrastructure, and has an open architecture whose properties can be stated and tested. The state can be any text or structured record about which decisions are made; our running case is the routing of court communications. The architecture serializes a request as a prefix tree (the state, then each question, then each option) and runs a bidirectional encoder under a block attention mask with logical positions, so that each answer is, by construction, invariant to the order of its options and independent of the other questions asked. Routing through an organization chart is a composition of Choice decisions with an explicit "none of these" option. We describe training with proper scoring rules through a curriculum, calibration and abstention, decisions composed by rubrics, and a protocol that accepts or rejects the mechanism before held-out data is opened. A minimal validation on public and fictitious data shows that the invariances hold to rounding in a trained model, that a small instance routes on organization charts it has never seen well above lexical retrieval, and that, at equal compute, the block architecture reaches a lower loss than a pairwise cross-encoder. A first instance trained on real legal documents agrees with the LLM classifier in production on 91% of held-out documents.

Large language models (LLMs) are predominantly trained on corpora that reflect Western, anglophone worldviews. This asymmetry is not neutral: when a model answers questions about territory, community authority, or the relationship with nature, it does so from assumptions that may be alien to the communities using the system. This paper presents a system for aligning language models with the ethical principles of Indigenous communities in Latin America. Drawing on interviews with five Indigenous leaders and defenders, we built an operational cultural constitution grounded in an Indigenous cosmovision, with 12 principles. On this basis, we implemented a Constitutional AI (CAI) pipeline in two phases: supervised fine-tuning (SFT) and direct preference optimization (DPO). The base models were Qwen2.5-7B-Instruct and Qwen2.5-1.5B-Instruct. With only 24 preference pairs, improvements were modest but detectable in intergenerational transmission, linguistic caution, and consultation and consent. Qualitative analysis reveals that the SFT phase was responsible for the primary shift in alignment, while DPO refined specific categories. The system was deployed in a conversational interface with retrieval-augmented generation (RAG) that allows communities to feed the knowledge base, turning alignment into a participatory cycle. We argue the architecture is functional and will scale qualitatively once the corpus is expanded and computing constraints are resolved. The work constitutes a first iteration of an epistemic justice mechanism for value alignment in the Global South.

People ask conversational AI models about their legal rights, obligations, procedures, and personal situations that may have legal consequences. To study this behavior, we first need to distinguish requests for legal guidance from other legally themed discussions and unrelated conversations. We contribute a multi-axis taxonomy of legal queries capturing whether legal guidance is sought, how a request is formulated, and which legal domains it concerns; apply it to 2,122 naturally occurring WildChat-1M conversations; and create a human-reviewed gold evaluation set of 200 conversations. All annotations, model checkpoints, and evaluation code are publicly available. In this enriched sample, 71.1% of conversations containing legal content seek personalized or practical guidance. Rights-and-options advice is the most common primary request type among legal-content conversations, and 25.3% of guidance-positive conversations span multiple legal topics. The gold evaluation set supports a two-stage comparison of ModernBERT-base and ModernBERT-large against a TF-IDF cascade. Our framework supports large-scale studies of legal guidance in conversational AI, and the public release of the annotated data facilitates reproducible research and future model development.

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens with bidirectional context. As these models reach real-world deployment, machine unlearning becomes essential. This process aims to remove the influence of a designated "forget set" without retraining from scratch, so that sensitive, copyrighted, or restricted content can no longer be elicited while preserving retained capabilities. Masked Diffusion Unlearning (MDU) is the first publicly available unlearning method customized for these models: it tunes the model $p_θ$ to match a frozen prompt-masked anchor $p_{\mathrm{anchor}}$ on the forget set by minimizing the anchor-last divergence $\mathrm{KL}(p_θ\Vert p_{\mathrm{anchor}})$. We identify two failure modes of MDU. First, random answer masking can leave key forget tokens visible to the anchor even when the question is fully masked. Second, at a memorized (near-one-hot) prediction, the anchor-last logit gradient vanishes while the anchor-first divergence $\mathrm{KL}(p_{\mathrm{anchor}} \Vert p_θ)$ logit gradient does not. As a result, MDU training stalls on the tokens we most need to forget. Backed by theoretical and empirical analyses, we propose Enhanced Masked Diffusion Unlearning (EMDU): anchor-first KL divergence toward a frozen anchor whose entire answer span is masked. On TOFU and RWKU benchmarks using LLaDA-8B and Dream-7B, our EMDU matches or exceeds the forgetting of MDU and adapted AR baselines, while better preserving retain utility.

Long-running agents rely on memory of their past interactions for task state, such as what has been completed and which conditions still hold. Such state is usually a derived state. No record states it, and it must be computed by rules from multiple records of an evolving history. Existing memory benchmarks score only the final answer, with no reference for the intermediate states behind it. A wrong answer therefore cannot be traced to the state that was missed, and a state that silently became outdated goes undetected. We present MemDerive, a diagnostic benchmark for agent memory over evolving data. It specifies each scenario symbolically, derives exact gold answers with Datalog, records the gold derivation DAG and the supporting events of every query, and renders the history in natural language with fictional entities and distractors. Its 10 scenarios contain 4,692 events and 1,503 queries whose difficulty varies along access, evolution, and composition. On this ground truth, MemDerive evaluates answer reliability, diagnoses failures by probing the nodes of the gold derivation DAG, and tracks whether answers stay correct as events arrive. Our evaluation of four memory systems reveals failures that final-answer benchmarks hide. On derived state, every memory system falls behind simple BM25 retrieval, although the same answer model reaches 96.01% when given the supporting events. Most failures are ungathered, as a system holds the missing intermediate state but does not retrieve it for the question that needs it. Memory systems also rarely propagate implicit changes, which no new record states. These results suggest three principles for memory design. Gather inputs along the derivation, store rules with their versions, and maintain derived states through recorded dependencies.

Reasoning large language models (LLMs) benefit from test-time scaling, motivating adaptive policies that spend tokens when they help and stop early otherwise. Existing rules halt when a confidence or stability signal crosses a threshold, but such thresholds rarely transfer across models and datasets, and validation-calibrated rules lose their guarantee under shifts in test-time horizon or signal dynamics. We reframe adaptive stopping as sequential evidence accumulation along the reasoning trajectory rather than threshold calibration on held-out data. Our framework runs two e-processes on a shared per-chunk signal: a commit stream tracks the signal's position relative to its slow trend, and an abstain stream tracks its signed velocity, making the two anti-correlated by construction. Under signal-level nulls on the per-chunk increments, Ville's inequality yields an anytime-valid signal-level false-alarm bound at user-specified significance levels, with no held-out calibration. Across various model–dataset combinations, the framework matches or exceeds calibration-based stopping on the token–accuracy Pareto front using zero calibration data, and the same significance level transfers unchanged across cells where calibration becomes infeasible.

An increasing number of disciplines recognize the explanatory potential of meaning-bearing mental structures through which individuals interpret the world and guide their behavior. Concept representations constitute a core class of these structures, emerging from individuals' experiences with concept referents and encoding and organizing their understanding of them. Recognizing a private level of meaning, where concepts are not objectively defined by rules in the world but constitute subjective structures in individuals' minds, raises a fundamental question: do different individuals actually mean the same thing when they invoke the same concept? As latent structures, concept representations are difficult to access and model empirically. The present study introduces a measurement framework that makes such representations empirically accessible through textual traces elicited under standardized conditions. Using transformer-based sentence embeddings, the framework maps standardized responses elicited for concepts into a high-dimensional shared vector space, opening concept representations to the systematic analysis of interindividual variation. It develops three analytical avenues: (1) comparing representational sharedness across concepts, (2) identifying variables that correlate with representational variation, and (3) examining how focal actors or subpopulations deviate from a reference population. Each avenue is illustrated through an empirical proof-of-concept drawing on a large-scale free-association dataset. Together, the demonstrations show how researchers from across disciplines can move from a research question and data source to analysis and interpretation. An accompanying GitHub repository provides a ready-to-use implementation of the framework for existing and custom datasets.

Verifiable rewards have advanced reinforcement learning in mathematics and code. However, many valuable open-ended tasks share three properties: no automatic verifier of overall quality, multiple requirements to satisfy jointly, and multiple valid responses to each query. Training therefore needs feedback on requirement satisfaction across different valid responses. Rubric-based reinforcement learning provides this feedback by aggregating scores from explicit criteria. As learning progresses unevenly, some criteria become fully satisfied and lose discrimination while others remain informative. Dynamic methods revise shared rubrics or select evolving criteria to restore discrimination, but may also change useful criteria. To address this, we propose Reinforcement Learning with Adaptive Rubric Evolution (RARE). The core idea is to use current policy feedback to guide both which criteria change and what their replacements assess. Sample-level updates replace criteria fully satisfied across current outputs with checks of remaining quality differences, retaining all other criteria and scores. Within this rule, we compare rewarding desirable qualities with penalizing observed flaws. Video script generation exemplifies these properties and supports content production. Its demands for coherent, executable narratives under constraints make it a challenging testbed. Targeting flaws performs best across three video types with a mean gain of 3.82 points over matched static rubrics. Model and human preferences support these gains. Improvements also extend to writing and general generation benchmarks.

Large language models (LLMs) are often used to simulate survey respondents, but matching survey averages does not establish accurate demographic differences. We test whether LLM outputs reproduce within-country age gaps in loneliness using the 2022 EU Loneliness Survey, comparing adults aged 18-34 and 65 or older in France, Italy, Poland, and Sweden. We evaluate simulation with and without age information and direct distribution estimation against a probability-based sample, with a non-probability sensitivity reference. Both human samples show higher mean loneliness among younger adults. Under age-specified simulation with standard response-option order, Mistral Small 4 and GPT-5.6 Terra reverse this ordering under all three prompt framings in three and two countries, respectively, mainly by overestimating older-adult means. DeepSeek V4 Flash preserves the ordering everywhere but underestimates both group means, with errors largely cancelling in France and Italy. Against the probability-based reference, direct estimation has lower absolute age-gap error than simulation for Mistral and Terra in every country. For DeepSeek, simulation has lower gap error but direct estimation has lower group-mean error; these comparisons depend on low-validity older-adult simulation cells. Evaluating LLM outputs as proxies for age groups within countries therefore requires checking both group means and their contrast against human survey data.

Semantic map models (SMMs) construct a graph that reflects conceptual similarity by comparing cross-lingual form-meaning mappings. Experts build such models not only by obeying the data constraint, i.e., the connectivity hypothesis, which states that conceptual nodes belonging to the same form must be connected, but also by incorporating their prior semantic knowledge. Current automatic methods ignore this knowledge, which harms both accuracy and efficiency by enlarging the search space of candidate graphs. In this paper, we fuse the distributional semantic knowledge hidden in the internal representations of language models into a top-down, data-driven algorithm. This expert-like behavior re-sorts the candidates in a huge search space, guided by conceptual semantics similarity. Experiments on a self-collected benchmark of 10 datasets and comprehensive metrics show that our method builds SMMs both effectively and efficiently.

Advancing physics reasoning in large language models requires scalable training data that connect physical laws, concrete problem scenarios, and verifiable numerical answers. Expert-designed problems offer top-down control over physical laws, conditions, and scenarios, but require costly human effort. LLM-based synthesis from texts scales efficiently, but model-generated answers and assessments leave supervision vulnerable to hallucinations. We introduce TIRE, a framework for Top-Down Formula Instantiation with Reliable Execution that brings knowledge-driven problem construction to large-scale data generation. Top-down instantiation uses reusable physical formulas and their applicability conditions to control the physical relations and numerical configurations of generated problems. Reliable execution establishes reference answers before problem generation, providing a common numerical foundation for problem construction and teacher distillation. Together, these mechanisms combine control over problem content with execution-grounded supervision for both supervised fine-tuning and reinforcement learning. Using TIRE, we construct TIRE-19k, spanning nine subfields of physics. Across 3 backbone LLMs and 4 benchmarks, TIRE-19k outperforms 5 representative datasets with equal amounts of SFT data, exceeding the strongest baseline by 2.86 percentage points on average. Further GRPO raises Qwen3-8B's average accuracy by 1.52 percentage points over its TIRE-19k SFT checkpoint. These results demonstrate the value of generating physics training data from explicit physical knowledge with reliable execution.

Speech anonymization has focused on the acoustic signal, leaving open how much speaker information the transcribed words themselves retain. In this work, we study pairwise speaker and demographic-attribute verification from transcribed speech. We introduce SPARC (Spoken Profile Authorship Recognition Corpus), a benchmark built from two podcast corpora comprising over 12,500 speakers, 1.5M quality-filtered utterances, and speaker-level inferred or source-provided labels for sex, age, and accent, and evaluate five attribution models and three LLMs on identity- and attribute-verification trials with episode- and metadata-based topic proxies. LLM adversaries distinguish same-speaker pairs from text alone. Attribute-label agreement is harder to verify, but LLMs perform sex- and accent-label verification above chance. Finally, releasing only short probe texts (50 words) degrades the strongest adversaries the most, offering a partial reduction in observed linkage.

Agents now act on the Web, publishing packages, models and datasets through registry APIs, and remembering past verifications spares them the work. But a verdict rests on premises. When a signing key is revoked, everything built on it stops passing, yet the advisory names only the key, and the old verdicts of its dependents may never share a retrieval window with it. This non-locality of invalidation turns memory into a liability: in long simulated chats, an agent reading a memory of atomic facts publishes 62% of the objects whose failure lies upstream, each backed by a record that was true when written. Yet derivation is local even when invalidation is not. Tether records, at write time, the exact versions each verdict derives from, and at read time withdraws any verdict whose support has failed anywhere in the store; the agent reads the same fields as before. On Fallout, a generator of versioned worlds with hypothesis-blind simulated histories shared by every memory, Tether publishes 0.3% of upstream-failing objects across three domains when relations are stated with their verdicts, with 37% fewer verification calls than checking everything. A-Mem and Graphiti re-verify most candidates (0.83-0.86 checks per candidate), which limits their savings, yet publish 9% of upstream-failing objects; Mem0 relies on memory more and publishes 42%. Flushing every verdict on any failure is as safe as Tether but costs 32% more checks. The gain over fact memory holds with three open-weight backbones and against a whole-store window. These gains depend on capturing dependencies: when relations arrive in sessions apart from their verdicts, Tether's downstream unsafe publication rate rises to 9.4-22.1%, and a weak writer that misses edges loses safety in the same way.

Answering natural language questions over knowledge graph data is challenging due to the vast number of facts, which can be difficult to process and navigate. One potential solution for this issue is to use mined subgraphs related to the query, although this process still requires extracting these subgraphs. This research presents a solution for extracting subgraphs related to entity candidates from a question-and-answer set, which can be obtained by inferring a large language model by calculating the shortest paths between entities. The proposed approaches detail various features that can be extracted from the subgraphs and reranking models to select the most probable answers from a list of candidates. Experiments were conducted on Wikidata to evaluate the effectiveness of the proposed approaches. This involved enumerating all the main feature types that can be extracted from mined subgraphs and a detailed analysis of the proposed features and reranking method combinations. In addition, a public web application that provides a useful web tool for studying the graph space between question and answer entities has been developed to work with subgraphs. This includes visualization of the extracted subgraph and automatic generation of natural language text to describe it.

Large language models need effective ways to acquire and use external knowledge beyond what they retain from pretraining. Continued pretraining risks catastrophic forgetting, while retrieval at inference time can add substantial overhead. External parametric memory avoids online retrieval, but constructing its training data requires extensive offline search, and output-level interpolation provides only shallow fusion. We introduce Deep Fusion Memory (DFM), a parametric memory architecture that reduces this construction workload and incorporates learned memory into intermediate hidden states. A causal memory predictor learns to generate memory from the preceding context, using retrieved chunk embeddings from the corpus as supervision. The predicted memory is shared across tokens within each chunk and fused into their hidden states through gated cross-attention. Memory prediction and fusion are trained end to end with the backbone frozen, and inference requires no explicit retrieval. Compared with existing parametric memory approaches, which often construct token-level memory targets, DFM requires $32\times$ fewer retrieval queries and $2048\times$ fewer retrieved items to construct its training data. Across six QA tasks, DFM performs better over the strongest baseline by 12.71% on Mistral-7B-v0.3 and 11.19% on Qwen3-8B-Base. Beyond QA, domain adaptation yields average gains of 6.51 and 4.38 points over Qwen3-4B-Base and Qwen3-8B-Base, respectively.

Skill-augmented agentic RL has emerged as an effective approach to improving LLM agents on long-horizon tasks. Despite this progress, individual skills may follow distinct trajectories as the policy evolves, with some becoming redundant after the model internalizes their capabilities, others remaining valuable as external scaffolding, and still others turning harmful. However, existing skill library systems fail to identify each skill's true contribution due to free-riding among co-retrieved skills, difficulty confounding across tasks, and the absence of skill-specific counterfactual evidence, which prevents reliable skill lifecycle management. We propose TrackRL, an agentic RL framework where skill tracking drives skill library full lifecycle management. Specifically, within each GRPO group, we turn grouped sampling into online randomized controlled experiments. Independent random injection and within-group contrasts on the same task isolate each skill's true contribution, and reliability-weighted aggregation accumulates evidence across groups. As evidence accumulates, we continuously track skill utility, which measures the skill's causal contribution as external scaffolding, and model mastery, which reflects how far the corresponding capability has been internalized. Along this line, we formulate skill lifecycle management as an evidence-gated state-transition process that governs each skill from generation and admission through protected incubation to retention, retirement, or removal, while a soft capacity bound controls library growth. Experiments show that TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, with further analyses revealing how skill value shifts during policy learning and how evidence-driven state transitions shape skill library evolution. Our code is available at https://github.com/xizhu1022/TrackRL.

People can express the same beliefs and desires yet need different reasons to change their minds. What makes an argument relevant depends on how those views arose: experiences shape what a person finds credible, what matters to them, and what could prompt reconsideration. These personal grounds may also be expressed indirectly, through contextual clues in conversation. Benchmarks that specify beliefs and desires without these connections can reward arguments that match a description while overlooking whether they address the person's reasons. We introduce A Thousand Faces (ATF-Persu), with 1,000 personas across 200 topics and nine domains. Drawing on personality and attitude theories and conversational implicature, we construct five personas per factual context, linking experiences to judgments, expression, and grounds for revision. Evaluations show lower persuasion success than three comparison benchmarks under their respective protocols. Complete backgrounds yield broader use of supplied personal grounds and higher persona faithfulness than simplified specifications; blinded reviewers also rate ATF-Persu above CToMPersu on identifiable personal reasons, their influence on responses, and contextual expression. The dataset is publicly available at https://huggingface.co/datasets/Slandes/ATF-Persu.

We present CHSM8, a 210M-parameter chess transformer that plays moves conditioned on a free-text description of a playing style, e.g. "go for the Sicilian Defense", "solid, safe, risk-free chess" or "You are Magnus Carlsen in his prime". The model is pretrained without search on 1.07B filtered Lichess games, using a factorized move representation and legal-move masking. It reaches 1837 Elo against a strength-limited Stockfish and, in 1,000-game matches, beat every other public model we tested that was likewise trained without preference tuning, rewards or search, including one with 13 times as many parameters. We then fine-tune it to follow text on a synthetic dataset: Gemma-4-31B describes the playing style of 3M game sides, a 1B Gemma-3 model distilled from the 2.4M of those descriptions that contradict nothing on the board labels the rest, and every one of 30M selected game sides gets five descriptions. Supervised fine-tuning attaches a frozen text encoder through gated cross-attention, and direct preference optimization (DPO) uses the same moves under a mismatched description as the rejected sample. After DPO, style adherence rises on every prompt we test (for example, queens traded early 53% -> 67%, pawn moves 29% -> 32% of all moves, queenside castling 97.3% -> 99.7%), and strength without a prompt stays within noise of SFT (1792 vs. 1797 Elo). Design choices at every stage were made by fixed-budget automated search loops on top of standard architectures and training recipes. We document the full pipeline and the decisions behind it, report the ablations that failed, and evaluate the models across many style prompts and kinds of games, with 1,000 games per setting. The work is fully open source and reproducible: weights, datasets with all training labels, prompts, configurations, inference code and playable demos are public and free to build on.

Theoretical literature warns that advanced generalist agentic systems might pursue instrumental goals that are misaligned with human preferences, thus posing risks of irreversible loss of human control. The EU's Code of Practice for General-Purpose AI lists loss of control as a systemic risk that must be assessed, yet no structured assessment of the quality of the empirical evidence for this risk exists. This study fills that gap. We evaluate 19 recent empirical studies across eight properties identified from threat models (such as strategic deception, situational awareness, and autonomous replication), assessing the quality of evidence across three dimensions: construct validity, content validity, and external validity. Our central finding is that the current body of evidence supports a judgment of loss of control as "weakly plausible." In our framework, 'plausibility' denotes the strength and direction of available evidence, not a probability estimate. Emerging evidence exists across multiple properties. No study clears the validity thresholds for a stronger judgment that a property is present; a single study clears them only for a negative-direction judgment (against situational awareness). Validity gaps, particularly in external validity, are widespread. Therefore, our findings highlight the need for more rigorous evaluations of agentic general-purpose AI systems. However, for the most dangerous properties, high external validity may be structurally unachievable under responsible research practices, reflecting a ceiling on feasible evidence rather than absence of the capability. Furthermore, as AI capabilities advance rapidly, empirical assessments may systematically lag behind the capability frontier. Our framework provides a structured baseline for ongoing monitoring as capabilities advance.

Memory serves as the core component of LLM-based agents, enabling personalized responses via extracting and using user preferences from past interactions. However, its use may reduce response diversity, thereby narrowing the range of information available to users. Existing evaluations rely largely on synthetic interactions and cover few memory systems, leaving the effects of memory on response diversity insufficiently evaluated in real-world scenarios. We introduce MemBubble, a benchmark of 1000 instances built on preferences extracted from real user-assistant conversations. To measure the extent to which memory concentrates response distributions, we construct test queries in scenarios distinct from user preferences while preserving topical relevance through shared preference categories. Across multiple LLMs and memory mechanisms, entropy decreases by 34.7% on average compared with directly providing the full interaction history, highlighting the important role of memory processing in diversity loss. We also explore simple user-side interventions and changes in information availability but find that these do not effectively close the diversity gap. Our findings highlight the need to consider preserving diversity and user access to information alongside improvements in personalization when designing and using agent memory systems.

As the world changes, new information is published while older records remain online. This complicates reasoning for large language models (LLMs), which must retrieve evidence and resolve conflicts to determine which information remains valid. Existing evaluations treat information seeking as a black box, revealing little about where and why LLMs fail. To address this gap, we introduce TREK, a suite of three components: TREKTRACE is a plug-and-play framework that attributes failures to specific stages of retrieval and reasoning. We apply it to TREKBENCH-WILD, a meta-benchmark of real-world questions that require complex reasoning over rapidly evolving knowledge. Surprisingly, the main bottleneck is post-retrieval reasoning rather than retrieval: frontier models are often misled by conflicting evidence even after retrieving all the information needed to answer. To study this bottleneck, we build TREKBENCH-CTRL, a controlled benchmark with distractors spanning temporality, relevance, and trustworthiness, and show that benchmarks dominated by outdated distractors substantially underestimate task difficulty. Together, TREK provides a testbed for diagnosing and improving reasoning over rapidly evolving knowledge.

Mathematical problem solving and autonomous agents are increasingly important applications of language models, demanding stronger reasoning capabilities. Computational depth is a key resource for such reasoning, but adding model layers increases both training and inference costs, while test-time scaling incurs substantial additional inference computation. We introduce P2N (Previous to Next), which reuses the previous token's deep representations, already available during generation, to increase effective computational depth with negligible decoding overhead. Specifically, P2N adds the previous token's deep hidden state to the current token's shallow representation, allowing subsequent layers to build on previously completed deep computation. This recurrence extends effective computational depth across tokens, while introducing only a single vector addition per decoding step. We train the recurrence using parallel Jacobi updates restricted to a middle-layer core. Experiments with Qwen-style models show that, at 1.5B parameters, P2N improves GSM8K accuracy by over 10 percentage points under both zero-shot and few-shot prompting, with nearly unchanged decoding latency. The benefits also extend to language modeling and average performance on general downstream benchmarks. on general downstream benchmarks.

Large Language Models often generate factually correct responses that fail to maintain user-requested perspectives, blending multiple perspectives and forcing users to filter through unwanted information—a perspective control failure that factual accuracy metrics cannot detect. We introduce PerspectiveBench, a benchmark quantifying perspective alignment by constructing scenarios where multiple valid reasoning paths coexist, and propose the Perspective Alignment Score (PAS) to measure adherence to specified perspectives. Experiments reveal that state-of-the-art models successfully include target content but fail to exclude non-requested perspectives. Providing documents through RAG improves factual grounding but does not resolve perspective drift, as models prioritize comprehensive coverage over strict user constraints. Code and data are available at https://github.com/sunghoon014/PerspectiveBench.

Large language models sometimes reply in an unexpected language, a failure known as language confusion. We ablate individual attention heads in 12 model checkpoints spanning seven model series, using FLORES-200 continuations in five languages and Language Confusion Benchmark (LCB) chat replies. In ten models, zeroing one head causes substantially more departures from the expected language than zeroing any other head in its layer; its LCB pass-rate drop persists, though reduced, under mean ablation. Seven heads move 8.0 to 51.5 percent of all FLORES continuations out of the prompt's language, versus at most 1.4 percent for other same-layer heads; the other three mainly affect explicitly requested languages. These seven heads matter less in base models given plain text. In two models, replacing the values a head reads at the requested-language name with another language's mean makes replies fail and shifts some into that language, unlike a random change of equal size. Replacing a head's output with another language's mean redirects 84 to 99 percent of scored non-English monolingual LCB replies into that language in five models. Yet the requested language's mean improves crosslingual replies beyond every control head only in the two Gemma models, whose baselines are low. Code and results are available in the public repository at https://github.com/christian-hoang-04/anonymous-release.

Acquiring fraud labels for payment transactions is expensive and inconsistent across tasks, and high-cardinality identifiers such as merchant names make one-hot feature pipelines infeasible at scale. We present CARTA, a self-supervised encoder that learns a compact 16-dim embedding from raw transactions without labels; the frozen embedding is then reused for label-efficient fraud probing and exploratory unsupervised risk stratification without retraining the encoder. CARTA uses type-specific projections with learned embedding tables for high-cardinality identifiers, trained with token excision and InfoNCE contrastive loss. On IEEE-CIS, its linear probe at 1% of labels exceeds a raw-feature probe at 100%; on IBM TabFormer (2M rows), its frozen embedding reaches GBDT AUPRC 0.508, above frequency encoding (0.411) and the merchant-dropping baseline (0.022), though fully-supervised CatBoost is stronger. CARTA is a label-free, reusable representation that complements task-specific supervised models.

We present Stable Signer V3, an interactive high-resolution sign language avatar system built on 4D Deformable Primitive Splatting. It addresses three limitations of existing avatars: restricted facial expression, inaccurate identity-specific body deformation, and loss of fine appearance details. Shared Face Primitive Splatting adds transferable superquadric primitives for richer expression control; Test-Time Body Primitive Splatting learns an identity-specific virtual muscle layer for more faithful motion; and Test-Time Tunneling Super-Resolution uses TunnelingNet to recover identity-specific high-resolution details. Both test-time modules are optimized once per signer identity; inference then runs entirely from the stored identity parameters. Together, they reduce facial geometry error by 71.4%, body-boundary error by 35.1% across ten identities, and high-resolution RGB reconstruction error by 24.2% on held-out evaluations, while retaining interactive streaming.

In 2022, the second author published a Last Words piece in Computational Linguistics in which he pointed out that, despite the advances in AI up to that point, there were still some problems in natural language processing that on the one hand seemed trivial and perhaps boring, but on the other were hard to solve using purely neural methods. One such problem discussed in that context was text normalization, and a particular subproblem of that, namely number name reading. The piece came out just a few months before the first release of ChatGPT, and needless to say, a lot has happened since then. Indeed, as we document here, the problem (of number name reading) has practically been solved, at least by one of the latest frontier models. However, we believe that there are still open issues for such limited domain, constrained and limited resource problems, and that in some future machine learning approach, the massive resources of the latest frontier models would not be needed to solve them.

Large Language Models (LLMs) have shown strong capabilities across natural language tasks. But their ability to generate accurate written verbalizations of mathematical formulas has not been systematically evaluated. Such verbalization is important for text-to-speech systems supporting audio-based learning and visually impaired users. We evaluated five state-of-the-art LLMs across six scientific domains and constructed Read-Aloud Math, a balanced multi-domain corpus with LaTeX formulas, manually validated simple and extended transcriptions, multiple variants, textual context, domain labels, and formula complexity levels. Using multi-reference BLEU, GPT-5.2 achieved the highest score for simple transcriptions (75.9% without context, 64.9% with context), although without context it did not significantly outperform Grok 4.1 Fast or Claude Sonnet 4.5. Claude Sonnet 4.5 achieved the highest score for extended transcriptions without context (36.5%), and Grok 4.1 Fast with context (48.8%). For extended transcriptions, Statistics and Mathematics scored highest across the two context settings, while Physics scored lowest. Formula length was significantly negatively associated with BLEU for simple transcriptions without context, but not with context, whereas BLEU increased significantly with formula length for extended transcriptions in both settings. Grouping symbols were most frequently involved in transcription differences, while variable names, capitalization, and Greek letters were associated with lower BLEU scores. Overall, LLM-based formula verbalization is promising but remains sensitive to transcription type, textual context, and formula characteristics.

AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints. Constructing training experience for these capabilities requires coherent requests, service records, tool interfaces, and grading checks. We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers. Workflow requirements guide task construction, and programmatic validation with feedback repair checks generated specifications and grading behavior. Teacher executions yield curated demonstrations for supervised fine-tuning (SFT), while student interactions provide rollouts and rewards for reinforcement learning (RL). Using these resources, we train compact PocketWorker models. For Qwen3-VL-8B-Instruct, SFT raises the mean score from 46.6 to 64.1 on Claw-Eval, including a 31.9-point gain on Multi-turn tasks. An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks. We further compare RL training dynamics across configurations that vary reward processing and credit assignment. Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.

Jev and its open replicas choose among the options a caller supplies with a single forward pass, fast enough to decide inside an agent loop. We ask what such a model cannot decide, and find a simple limit: Jev cannot make a decision that an LLM makes only by writing out intermediate steps. On a benchmark that varies one factor at a time, Jev finds and selects records without error up to 20k tokens and 64 options, but once the answer depends on two or more updates of one user, accuracy falls, to 53-80% at sixteen updates, because similar records interfere. Given a stale summary of the state, Jev follows it. On real decision items graded by how many thinking tokens a reference LLM needs, Jev answers 97.5% of the items that one forward pass of that LLM solves, but only 74% of the others. Jev also does not know its limit: its confidence flags the errors it makes itself but not those it inherits from a stale input, and neither an explicit nor a trained hand-off option catches the latter. The open models we evaluate share these limits. Checking the state such models rely on requires a model that reasons.

Benchmarking large language model (LLM)-generated text detectors commonly relies on aggregate detector-level metrics such as accuracy, F1 score, and AUC. Although useful for summarizing overall performance, these metrics provide limited insight into how benchmark composition shapes detector comparisons or which texts distinguish detectors with different response profiles. We propose a measurement-based benchmarking framework for jointly characterizing benchmark heterogeneity and detector behavior using multidimensional item response theory (IRT). By treating benchmark texts as measurement items and detectors as respondents, the framework estimates text-level discrimination and difficulty-location parameters alongside multidimensional detector abilities. We apply the framework to a benchmark of 32K human-written and machine-generated texts across four domains, evaluated by 14 detectors. The estimated IRT item-parameter space reveals distinct but overlapping human-machine distributions, with substantial variation across domains and text lengths. The human-machine label shows the strongest multivariate association with this space, while domain and text length further contribute to benchmark heterogeneity. Detector ability profiles reveal multidimensional and domain-dependent response patterns that are not fully reflected in aggregate performance rankings. As a practical application, clustering texts in the estimated IRT space yields a compact subset of 2,088 texts (6.5\% of the full 32K benchmark) that largely preserves the comparative detector-performance patterns observed on the full benchmark. Overall, the proposed framework provides a principled and interpretable methodology for benchmarking heterogeneous detection technologies while jointly characterizing the measurement properties of the texts used for their evaluation.

Language model benchmarks assign each question a single outcome from one fixed wording, even though the same underlying request can be expressed in many equivalent ways. We argue that the quantity of interest should therefore not be correctness on one wording, but the accessibility of an inquiry, which we define as the probability that a model succeeds across meaning-preserving formulations. The distribution of accessibility across inquiries reveals a structure that accuracy and existing best-, worst-, or average-case prompt scores collapse: how broadly a model's capabilities extend across inquiries and how dependably those capabilities can be elicited. Two models can have the same average performance while one succeeds across more inquiries unreliably and the other succeeds across fewer inquiries but more consistently, leading to opposite preferences depending on whether the task allows multiple retries until one success or when repeated consistent success is required. Across six benchmarks and more than twenty models, this tradeoff is common: for about half of the model pairs with nearly equal mean accessibility, which model looks better depends on whether one success across five formulations is enough or all five must succeed, quantities we call Access@5 and Reliable@5, respectively. For some pairs the two measures diverge significantly, by up to 10 points in opposite directions. Accessibility also reveals that scaling, reasoning, and prompt optimization can redistribute capability across inquiries in ways aggregate accuracy obscures, sometimes broadening the set of inquiries a model can succeed on while making success on them less repeatable. Accessibility turns formulation sensitivity from a nuisance into a measurable property of capability, distinguishing what a model can sometimes succeed on from what users can actually depend on.

Alzheimer's Disease (AD) detection from speech benefits from both paralinguistic and linguistic cues, yet model training is frequently constrained by limited data. Fine-tuning a large pretrained speech encoder can easily overfit, while a frozen probe leaves the encoder's task-relevant knowledge underutilized. Therefore, we propose Dual Task Vector Merging, a parameter-efficient method that adapts a frozen wav2vec 2.0 backbone using task knowledge learned outside the AD task. Specifically, we inject two task vectors, derived as weight deltas from emotion- and ASR-fine-tuned wav2vec 2.0 models, into each transformer layer. During AD training, only layer-wise merging coefficients and a linear classification head are optimized, updating less than 0.1% of the parameters required for full fine-tuning. Speaker-independent experiments on ADReSS and ADReSSo show that dual merging improves over frozen probes, LoRA, and full fine-tuning, in the main binary settings, and gives the best results when combined with LoRA. Evaluation on PROCESS further shows that our advantages transfer to three-class HC/MCI/dementia classification when paired with LoRA. Experiments under reduced training data (20%, 40%, and 80%) confirm reliable gains in low-resource settings.

Small Language Models with quantisation and parameter-efficient adaptation are widely offered as the route to multilingual AI under compute constraints, and their efficiency is almost always reported as a single memory figure. We show that this figure is measured in English and does not transfer. Across a 131k-sentence corpus spanning eleven Indic scripts, TinyLlama-1.1B's tokeniser emits 1.34 subword tokens per Latin-script word and 22.60 per Kannada word. This is not inefficiency; it is absence. Its vocabulary contains 0% of Odia's codepoints, 1% of Telugu's and 2% of Kannada's, and for five of eleven scripts the measured fertility sits at the UTF-8 byte-fallback ceiling, on Odia at 99.6% of it, with 93.6% of emitted tokens being raw bytes. Because SentencePiece byte fallback never raises an unknown-token marker, this failure is entirely silent: TinyLlama's UNK rate is 0.0% on every script, including the one it reads as bytes. We further show that low fertility can indicate the opposite of efficiency: mBERT posts its best score anywhere (1.10 tokens/word, on Odia) while discarding 87.9% of the text as UNK. MuRIL, an Indic-vocabulary encoder with a fifth of TinyLlama's parameters, reads every script at 1.23-2.26 tokens/word with zero UNK and zero byte fallback. The binding constraint is the vocabulary, not the parameter count. The consequence is concrete: on a 2048-token context window, the same model gives an English user 1,528 words of usable context and a Kannada user 91.

Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation. Meanwhile, the medical literature offers a wealth of high-quality unlabeled documents, including clinical case reports that trace the full clinical course from initial presentation to final treatment decisions. We study whether document-grounded self-play can convert these reports into effective RL training signals. We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation}, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training. To address these challenges, we propose CasePlay, a self-play framework with two key components: (a) a Knowledge-Conditioned Proposer, which anchors question generation to case-specific knowledge points so that questions draw on a broader range of evidence from the source report; and (b) a Rubric-Judge Reasoner, which augments answer-accuracy feedback with rubric-based quality feedback on clue integration, reasoning depth, and option quality, guiding the proposer toward reasoning-intensive questions that better support reasoner training. Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.1 points on average, highlighted by a 6.3-point gain on LiveClin. This approach offers an effective path toward continually improving medical reasoning ability through self-play RL.

State media outlets often project their governments' narratives to target both domestic and foreign audiences. Because this content is plentiful and free to access, it appears widely in the web corpora that large language models (LLMs) are trained on and that retrieval systems draw from. However, the prevalence of state media in these corpora and its potential influence on LLM outputs remain largely unexplored. In this paper, we introduce WorldStateMedia, an open dataset of 118 million state media articles from 136 countries in 114 languages. We extract these articles from existing pretraining corpora and annotate each outlet with level of state control from State Media Monitor. Using WorldStateMedia to audit these corpora, we find that, in English web crawls, the share of state media pages from editorially controlled (vs. independent) outlets rose from 39% in 2013 to 93% in 2025. In several low-resource languages, state media make up a substantial share of all web data, such as 17.8% for Uyghur and 7.8% for Belarusian. To characterize the differing narratives of news outlets, we compare Russian and Ukrainian state media in English and in their native languages, revealing asymmetric topical focus and opposing framings of contested issues such as Crimea. Furthermore, in a retrieval-augmented generation (RAG) experiment on territorial disputes, retrieving a country's state media shifts model answers toward that country's claims, most strongly for English-language (foreign-targeted) media. We release WorldStateMedia to support multilingual, cross-national insights on state media at https://huggingface.co/collections/manestay/worldstatemedia

Adapting a pretrained machine translation (MT) model to new terminology usually requires new parallel data and another round of fine-tuning, which is costly for low-resource languages. We ask whether controllability can be added to MT, and the fine-tuning burden reduced, without degrading the pretrained model. GLOSS-MT looks up the source's noun phrases in an offline English-Twi dictionary, replaces every phrase found by its Twi translation in place, and fine-tunes a translation model once to translate these mixed inputs. We apply the same data and pipeline to NLLB-200 (600M, 1.3B, 3.3B), OPUS-MT en-tw and TranslateGemma 4B. On a document-level English-to-Asante Twi benchmark (84 documents, four domains), every fine-tune scores above its pretrained base in chrF2 (+2.85 to +5.48), and the three NLLB fine-tunes are the best open systems (50.12-50.80). A direct controllability test on human-translated Bible passages shows that the three fine-tunes tested write 79-97% of the dictionary terms, and that replacing every dictionary entry by a different term changes the output accordingly (70-97% of the new terms appear) with no retraining. We release the models, data and code.

Jailbreak papers report an attack success rate, the share of responses that deliver the harmful request, and read it as a property of the target model. We hold the target model, the prompts and the responses fixed and change only the grader. Three dense graders of 8B and 9B parameters, from three model families, move compliant responses to the neighbouring class, evasion, instead of scattering them across classes. One of them assigns the compliance label zero times in 900 responses. This class collapse also appears on new harmful requests, on a target model from a third family, and in a held-out 7B grader from a fourth family. Per-class label counts expose it at no cost and without human labels. The collapse decides how the rate behaves under a change of grader. On two corpora of 900 responses and eight graders, the compliance rate differs by a factor of 22 to 25, while the non-refusal rate differs by 18 to 23 percentage points. Without the three collapsing graders, compliance differs by a factor of 1.7 to 2.8 on all four corpora, and non-refusal still differs by 10 to 22 points. These results do not change when every grader is called once. We propose a reporting minimum that keeps the compliance rate and adds its per-class label counts and the named grader. Beside it goes the non-refusal rate, whose disagreement across graders is additive.

Language models behind coding agents are post-trained with outcome rewards, and recent methods concentrate credit on the turns that decide the outcome. We find that defensive actions beyond what a task requires escape both. They begin at silent forks, the turns where every next action leads to the same outcome. Such actions leave the success rate unchanged, and frontier models take the disproportionate branch there in up to 91% of decisive actions under implicit budgets. The most direct fix, a checker in the trajectory reward, reaches these forks only through fully proportionate trajectories, leaving 74% of GRPO groups with identical rewards and behavior unchanged. This motivates a rethink of where learning signals belong, namely where behavior diverges, which a behavioral checker can locate and the outcome cannot. Our method, ForkDPO, contrasts a proportionate and a disproportionate action at the same prefix, at silent forks on both sides of the required behavior. On a test set of unseen repositories, it lowers a 4B model's disproportionate behavior from 70.3% to 58.9%, against 67.7% for DPO at randomly chosen turns, establishing behavioral divergence as a learning signal for how agents act beyond whether they finish. The code is publicly available at https://github.com/Astarojth/silent-forks.

While LLM agents are increasingly capable of gathering and synthesizing evidence from a single homogeneous source, such as structured tables or unstructured data, real-world analytical questions frequently span heterogeneous sources. For example, answering one enterprise question may require both a result computed over an internal database and evidence found on the web. Recent evaluations on benchmarks like HybridDeepResearch show that even frontier agents struggle when evidence must cross this boundary. While these benchmarks have exposed the limitation, methodologies to explicitly teach cross-source reasoning remain absent. To address this, we introduce HybridResearcher, a 4B agent trained to bridge evidence across the structured and unstructured boundary. To build it, we synthesize cross-source bridge trajectories by starting with validated single-source tasks-such as existing text-to-SQL or web search queries-and fabricating the missing source around a bridge entity drawn from the seed's own output. We retain a sample only when strict construction gates confirm that neither side alone suffices, ensuring the data requires true cross-source interdependence. Finally, we train on this curated dataset using supervised fine-tuning followed by reinforcement learning. Evaluated on the existing HybridDeepResearch benchmark, HybridResearcher improves over our base model, Qwen3.5-4B, by 30.0 points pass@8 and surpasses the far larger Qwen3.5-397B-A17B by 8.3 points, while also improving over its base model on established, held-out single-source benchmarks (BrowseComp and BrowseComp-Plus), demonstrating that this complex cross-source reasoning can be taught directly to small models without triggering catastrophic forgetting.

This paper introduces BanglaBBQ, a question-answering bias benchmark designed to capture stereotypes grounded in the Bangladeshi context, in both Bangla and English. Following KoBBQ and PakBBQ, BanglaBBQ adapts the original Bias Benchmark for Question Answering (BBQ) by classifying its templates into three groups: directly translated (DT), modified to align target groups or contexts with Bangladesh (TM), and removed when deemed unsuitable for the Bangladeshi context (SR). This process retains 140 templates across six categories, comprising 76 DT and 64 TM templates. Five newly authored (NA) categories, based on stereotypes common in Bangladeshi society, are then added: marital status, age and political perception, regional origin, educational background, and mental-health status. Native Bangladeshi annotators judged all 466 newly authored templates, and 372 were retained because a strict majority of English annotators identified the intended target and rated the attitude as common enough. Each newly authored template is linked to a documented source and a verbatim quote from it. Each evaluated template also has a counter-stereotypical version of its disambiguated context. The dataset was evaluated using two open-weight models (Qwen3-4B and Qwen3-8B) and two closed models (GPT-4.1-mini and Gemini 2.5 Flash). All four lean towards the stereotyped person when the context is ambiguous, in both languages, and the effect of language differs by model, so English results do not predict behaviour in Bangla. On the adapted BBQ templates, both Qwen3 models are also less accurate when the context contradicts a stereotype than when it confirms it. On the newly authored templates, the closed models follow counter-stereotypical evidence almost perfectly, and all four models are more biased on the templates that annotators validated than on those they rejected.

Discrete diffusion language models (dLLMs) offer an attractive alternative to autoregressive generation by updating multiple tokens in parallel, but their inference efficiency is still limited by the cost of safely committing tentative predictions. In particular, speculative decoding for dLLMs is challenging because speculative tokens within the same diffusion block may depend on one another, making aggressive acceptance prone to error propagation and strict verification computationally expensive. We present PV-Decoding, a confidence-driven speculative decoding framework with progressive verification for block diffusion inference. PV-Decoding partitions masked positions into directly decoded tokens, speculative tokens, and unresolved tokens using two confidence thresholds, and then verifies speculative tokens with progressively stronger policies, including single-batch, dual-batch, and multi-batch verification. This design exposes a practical accuracy–efficiency trade-off: weaker verification minimizes overhead, whereas stronger verification improves acceptance precision and robustness. Experiments on three diffusion language models across GSM8K, MBPP, and HumanEval show that PV-Decoding reduces sequential model forward passes (MFPs) by 8%–37% relative to confidence-based decoding, while keeping the largest observed accuracy drop to 1.3%. On Dream-7B-Instruct/GSM8K, PV-Decoding (1-batch) and PV-Decoding (2-batch) improve throughput by 7.0%–13.6% under the same hardware budget, whereas PV-Decoding (multi-batch) is slower under the current runtime. These results suggest that confidence-guided speculation with structured verification is a practical path toward efficient and reliable dLLM inference.

Diffusion language models (dLLMs) generate by gradually unmasking a canvas, so their samplers decide not only which token to write but also when to commit it. We show that this second decision changes who appears in generated text, a sampler-induced bias on top of the model's own. Common commit rules, including LLaDA's default, commit the positions whose sampled tokens are most probable; this peeking favors majority options. To separate the sampler from the model, we read the model's belief about a fork, such as a protagonist's gender, in one forward pass, and show that this belief matches exact sampling from the model. Measured against it, peeking samplers on LLaDA and Dream amplify occupational gender stereotypes and collapse opinion distributions, and confidence-based ordering shifts gender-balanced occupations toward men, more so under faster parallel decoding. We propose Fork-First Sampling, which decides the fork before anything else is written: it follows the belief or any target distribution, improves fluency, and avoids the stories that remark on the protagonist's gender, which prompt-based fixes produce, at the cost of one forward pass. Fairness evaluations of dLLMs should report the commit rule.

Speech recognition in real time in far-field conversational settings still incurs significant errors. Identifying potential error spans during streaming ASR permits a recourse mechanism where the speaker can be asked to repeat the affected content under cleaner acoustic conditions. We propose \sysname, a new framework for online error localization, that identifies spans likely to benefit from this intervention. Our supervision uses clean audio to estimate how the recognizer's output might change under improved acoustic conditions. We probabilistically align noisy and clean decodes using a pair-HMM to obtain soft supervision. These soft targets train a counterfactual gain scorer by combining a suite of acoustic and language features to produce token-level scores, that are further grouped into spans for targeted recourse. An LLM uses the ASR transcript of the repetition to revise only the selected span. At a precision of $70\%$, \sysname achieves the highest recall by identifying up to 52\% more erroneous tokens compared to the strongest baseline. Targeted repetition yields relative WER reductions of up to 24\% under a budget that accounts for both interruption frequency and the lengths of re-uttered spans.