LangTrend

Sentiment analysis has been of long-standing interest in psychotherapy research. Recently, the Transformer deep learning architecture has produced text-based sentiment analysis models that are highly accurate and context-aware. These models have been explored as proxies for emotion measurement instruments in psychotherapy, but not investigated as stand-alone psychometric tools. Using proposed utterance-level and session-level sentiment features derived from a fine-grained sentiment model on a large corpus of psychotherapy sessions (N = 751), we investigate the distribution of session aggregated sentiment scores. Further, we characterize the relationship of these features to individual components and the overall score of the OQ-45 instrument and find that this sentiment feature is most strongly correlated to components related to emotional valence in directionally intuitive ways. Finally, we report that there are statistically significant differences between the sentiment distributions for patients flagged as at risk of deterioration or dropping out of care via either the OQ Rational or Empirical outcome models. These correlations to a fully-validated psychometric instrument demonstrate that these proposed sentiment features are, at least, adjunctive measures of client distress and deterioration.

The Metacognitive Probe is an exploratory five-task, 15-slot diagnostic that decomposes an LLM's confidence behaviour into five behaviourally-distinct dimensions: confidence calibration (T1-CC), epistemic vigilance (T2-EV), knowledge boundary (T3-KB), calibration range (T4-CR), and reasoning-chain validation (T5-RCV). It is evaluated on N=8 frontier models and N=69 humans. The instrument is motivated by Flavell (1979) and Nelson and Narens (1990) but operates on observable confidence-correctness alignment; it is not a validated cross-species metacognition scale, and the pre-specified human developmental hypothesis was falsified. Composite benchmarks (MMLU, BIG-Bench, HELM, GPQA) ask whether a model produces a correct response. They are silent on whether the model knows when its response is wrong. A model can score 80 on a composite calibration benchmark and still be wildly overconfident in narrow pockets the aggregate cannot surface. The Metacognitive Probe surfaces those pockets. Our headline is a 47-point within-model dissociation in Gemini 2.5 Flash: panel-best within-task calibration (T1-CC = 88; Spearman rho = +0.551, 95% CI [+0.14, +0.80], p = 0.005) and panel-worst cross-task difficulty prediction (T4-CR = 41; sigma_conf = 1.4 across twelve factoids).

Production LLM coding agents drift over long sessions: they forget user-specified constraints, slip into mistakes the user already flagged, and confabulate prior agreements. White-box approaches such as persona vectors require model weights and so cannot be applied to closed APIs (Claude, GPT-4) that most users actually interact with. We present Nautilus Compass, a black-box persona drift detector and agent memory layer for production coding agents. The method operates entirely at the prompt-text layer: cosine similarity between user prompts and behavioral anchor texts, aggregated by a weighted top-k mean using BGE-m3 embeddings. Compass is, to our knowledge, the only public agent memory layer (among Mem0, Letta, Cognee, Zep, MemOS, smrti verified May 2026) that does not call an LLM at index time to extract facts or build a graph; raw conversation text is embedded directly. The system ships as a Claude Code plugin, an MCP 2024-11-05 A2A server (Cursor, Cline, Hermes), a CLI, and a REST API on one daemon, with a Merkle-chained audit log for tamper-evident anchor updates. On a held-out test set built from real Claude Code session traces and labeled by an independent LLM judge, Compass reaches ROC AUC 0.83 for drift detection. The embedded retrieval pipeline scores 56.6% on LongMemEval-S v0.8 and 44.4% on EverMemBench-Dynamic (n=500), topping the four published EverMemBench Table 4 baselines. LongMemEval-S 56.6% is ~30 points below recent white-box leaders (90+%); we treat that as the architectural ceiling of the no-extraction design. End-to-end reproduction cost is $3.50 (~14x cheaper than GPT-4o-judged stacks). A paired cross-vendor behavior A/B accompanies these numbers as preliminary system-level evidence. Code, anchors, frozen test data, and audit-log tooling are MIT-licensed at github.com/chunxiaoxx/nautilus-compass.

We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and linear RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. It can also be implemented in a hybrid solution in tandem with LRNN layers in place of traditional attention, to supplement the LRNN with improved sublinear memory growth context length usage and long context decoding. We release our code at https://github.com/featherless-ai/KVM-paper and trained models at https://huggingface.co/collections/featherless-ai/kvm-paper under the Apache 2.0 license.

Large language models (LLMs) are often evaluated based on their stated values, yet these do not reliably translate into their actions, a discrepancy termed "value-action gap." In this work, we argue that this gap persists even under explicit reasoning, revealing a deeper failure mode we call "Pseudo-Deliberation": the appearance of principled reasoning without corresponding behavioral alignment. To study this systematically, we introduce VALDI, a framework for measuring alignment between stated values and generated dialogue. VALDI includes 4,941 human-centered scenarios across five domains, three tasks that elicit value articulation, reasoning, and action, and five metrics for quantifying value adherence. Across both proprietary and open-source LLMs, we observe consistent misalignment between expressed values and downstream dialogues. To investigate intervention strategies, we propose VIVALDI, a multi-agent value auditor that intervenes at different stages of generation.

A vision-language model can look at a knot diagram and report what it sees, yet fail to act on that structure. KnotBench pairs an 858,318-image corpus from 1,951 prime-knot prototypes (crossing numbers 3 to 19) with a protocol whose answers are checked against Regina's canonical knot signature. Its 14 tasks span four families, equivalence judgment, move prediction, identification, and cross-modal grounding; an image-versus-symbol split locates failures along the perception-operation gap. We score Claude Opus 4.7 and GPT-5, each with and without thinking, under a 64K output-token budget matched on both vendors. Across 56 (task, model) cases, 15 sit at or below a random baseline and 8 of 14 tasks have a best score under 1.5x random. On diagram-to-symbol transcription, no model produces a strictly correct string, and permissive Regina decoding recovers the knot in 0 to 4 of 100 items. Thinking-mode reasoning lifts overall accuracy by 1.65 points for Claude and 9.25 points for GPT-5, narrowing the gap only modestly. Read together, the four families suggest current vision-language models hold features of a diagram but lack apparatus to simulate moves on those features.

The implicit policy of maintaining relatively stable acceptance rates at top AI conferences, despite exponentially growing submissions, introduces a critical structural vulnerability. This position paper characterizes a new systemic threat we term Agentic Denominator Gaming, in which a malicious actor deploys AI agents to generate and submit a large volume of superficially plausible but low-quality papers. Crucially, their objective is not the acceptance of low-quality papers, but rather to inflate the submission denominator and overwhelm reviewing capacity. Under a relatively stable acceptance rate, this dilution can systematically increase the publication probability of a small, targeted set of legitimate papers. We analyze the practical feasibility of this threat and its broader consequences, including intensified reviewer burnout, degraded review quality, and the emergence of industrialized automated agent mills. Finally, we propose and evaluate a range of mitigation strategies, and argue that durable protection will require system-level policy and incentive reforms, rather than relying primarily on technical detection alone.

Recent advancements in Neural Machine Translation (NMT) have significantly improved translation quality. However, the increasing size and complexity of state-of-the-art models present significant challenges for deployment on resource-limited devices. Knowledge distillation (KD) is a promising approach for compressing models, but its effectiveness diminishes when there is a large capacity gap between teacher and student models. To address this issue, we propose Evolving Knowledge Distillation (EKD), a progressive training framework in which the student model learns from a sequence of teachers with gradually increasing capacities. Experiments on IWSLT-14, WMT-17, and WMT-23 benchmarks show that EKD leads to consistent improvements at each stage. On IWSLT-14, the final student achieves a BLEU score of 34.24, narrowing the gap to the strongest teacher (34.32 BLEU) to just 0.08 BLEU. Similar trends are observed on other datasets. These results demonstrate that EKD effectively bridges the capacity gap, enabling compact models to achieve performance close to that of much larger teacher models.Code and models are available at https://github.com/agi-content-generation/EKD.

Disagreement in annotation is a common phenomenon in the development of NLP datasets and serves as a valuable source of insight. While majority voting remains the dominant strategy for aggregating labels, recent work has explored modeling individual annotators to preserve their perspectives. However, modeling each annotator is resource-intensive and remains underexplored across various NLP tasks. We propose an agreement-based clustering technique to model the disagreement between the annotators. We conduct comprehensive experiments in 40 datasets in 18 typologically diverse languages, covering three subjective NLP tasks: sentiment analysis, emotion classification, and hate speech detection. We evaluate four aggregation approaches: majority vote, ensemble, multi-label, and multitask. The results demonstrate that agreement-based clustering can leverage the full spectrum of annotator perspectives and significantly enhance classification performance in subjective NLP tasks compared to majority voting and individual annotator modeling. Regarding the aggregation approach, the multi-label and multitask approaches are better for modeling clustered annotators than an ensemble and model majority vote.

How should hidden states generated autoregressively be collapsed into a representation that reflects a language model's internal state? Despite tokens being generated under causal masking, we find that mean pooling across their hidden states yields more semantic representations than any individual token alone. We quantify this through kernel alignment to reference spaces in language, vision, and protein domains. The improvement through mean pooling is consistent with information being distributed across generated tokens rather than localized to a single position. Furthermore, representations derived from generated tokens outperform those from prompt tokens, and alignment across generation reveals interpretable dynamics in model behavior.

Reliable detection of personally identifiable information (PII) is increasingly important across modern data-processing systems, yet the task remains difficult: PII spans are heterogeneous, locale-dependent, context-sensitive, and often embedded in noisy or semi-structured documents. We present GLiNER2-PII, a small 0.3B-parameter model adapted from GLiNER2 and designed to recognize a broad taxonomy of 42 PII entity types at character-span resolution. Training such systems, however, is constrained by the scarcity of shareable annotated data and the privacy risks associated with collecting real PII at scale. To address this challenge, we construct a multilingual synthetic corpus of 4,910 annotated texts using a constraint-driven generation pipeline that produces diverse, realistic examples across languages, domains, formats, and entity distributions. On the challenging SPY benchmark, GLiNER2-PII achieves the highest span-level F1 among five compared systems, including OpenAI Privacy Filter and three GLiNER-based detectors. We publicly release the model on Hugging Face to support further research and practical deployment of open PII detection systems.

Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.

In high-stakes domains such as healthcare, the reliability of Large Language Models (LLMs) is critical, particularly when generating clinical insights from incident reports. This study proposes a tag-based few-shot example selection method for prompting LLMs to generate background/causal factors and preventive measures from details of the medical incidents. For our experiments, we use the Japanese Medical Incident Dataset (JMID), a structured dataset of 3,884 real-world medical accident and near-miss reports. These reports are variably annotated with a wide range of tags–some include descriptive information (e.g., "medications," "blood transfusion therapy"). We compare three few-shot example selection strategies–random sampling, cosine similarity-based selection, and our proposed tag-based method–using GPT-4o and LLaMA 3.3. Results show that the tag-based approach achieves the highest precision and most stable generation behavior, while similarity-based selection often leads to unintended outputs and safety filter activation. These findings suggest that selecting examples based on human-interpretable dataset tags can improve generation precision and stability in clinical LLM applications.

Psychological support hotlines provide critical support for individuals experiencing mental health emergencies, yet current assessments largely rely on human operators whose judgments may vary with professional experience and are constrained by limited staffing resources. This paper proposes a large language model (LLM)-based framework for automated crisis level classification, a key indicator that supports many downstream tasks and improves the overall quality of hotline services. To better capture emotional signals in spoken conversations, we introduce a paralinguistic injection method that inserts identified non-verbal emotional cues into speech transcripts, enabling LLM-based reasoning to incorporate critical acoustic nuances. In addition, we propose a reasoning-enhanced training strategy that trains the model to generate diagnostic reasoning chains as an auxiliary task, which serves as a regulariser to improve classification performance. Combined with data augmentation, our final system achieves a macro F1-score of 0.802 and an accuracy of 0.805 on the three-class classification task under 5-fold cross-validation.

Cell-type-specific marker genes are fundamental to plant biology, yet existing resources primarily rely on curated databases or high-throughput studies without explicitly modeling the supporting evidence found in scientific literature. We introduce PlantMarkerBench, a multi-species benchmark for evaluating literature-grounded plant marker evidence interpretation from full-text biological papers. PlantMarkerBench is constructed using a modular curation pipeline integrating large-scale literature retrieval, hybrid search, species-aware biological grounding, structured evidence extraction, and targeted human review. The benchmark spans four plant species – Arabidopsis, maize, rice, and tomato – and contains 5,550 sentence-level evidence instances annotated for marker-evidence validity, evidence type, and support strength. We define two benchmark tasks: determining whether a candidate sentence provides valid marker evidence for a gene-cell-type pair, and classifying the evidence into expression, localization, function, indirect, or negative categories. We benchmark diverse open-weight and closed-source language models across species and prompting strategies. Although frontier models achieve relatively strong performance on direct expression evidence, performance drops substantially on functional, indirect, and weak-support evidence, with evidence-type confusion emerging as a dominant failure mode. Open-weight models additionally exhibit elevated false-positive rates under ambiguous biological contexts. PlantMarkerBench provides a challenging and reproducible evaluation framework for literature-grounded biological evidence attribution and supports future research on trustworthy scientific information extraction and AI-assisted plant biology.

Frontier coding agents read configuration files (CLAUDE$.$md, AGENTS$.$md, Cursor Rules) at session start and are expected to follow the conventions inside them. Practitioners assume that structural choices (file size, instruction position, file architecture, contradictions in adjacent files) measurably affect adherence. We report a systematic factorial study of these choices using four manipulated variables, measuring compliance with a trivial target annotation across 1,650 Claude Code CLI sessions (16,050 function-level observations) on two TypeScript codebases, three frontier models (primarily Sonnet 4.6, with Opus 4.6 as a CLI-matched cross-model check and Opus 4.7 reported descriptively under a CLI-version confound), and five coding tasks. We use mixed-effects models with a Bayesian companion. None of the four structural variables or three two-way interactions produces a detectable contrast after multiple-testing correction. Size and conflict nulls are supported by affirmative-null Bayes factors (BF10 between 0.05 and 0.10); position and architecture nulls are failures to reject without Bayes-factor support. The largest effect we measured is within-session: each additional function the agent generates is associated with approximately 5.6% lower odds of compliance per step (OR = 0.944) within the session-length range we tested, though the relationship is non-monotonic rather than a constant per-step effect. This reproduces on a second TypeScript codebase and on Opus 4.6 at matched configuration; it was identified during analysis rather than pre-specified. Within the conditions tested, file-structure variables did not produce detectable contrasts; compliance varies systematically between coding tasks and across each session's sequence of generated functions.

Large Language Model (LLM) personalization aims to align model behaviors with individual user preferences. Existing methods often focus on isolated user histories, neglecting the essential role of inter-user differences. We propose C-BPO, a framework that personalizes LLMs via preference-calibrated binary signals. By treating target user data as positive feedback and other users' data as an auxiliary set of implicit negative signals, C-BPO captures distinct inter-user differences. To mitigate the preference overlap issue, where shared task knowledge is erroneously penalized, we derive an objective grounded in Positive-Unlabeled (PU) learning theory. This approach purifies negative signals by subtracting ``positive bias'', ensuring alignment with unique idiosyncrasies without compromising general helpfulness. Empirical experiments across various personalization tasks and backbone LLMs show C-BPO consistently outperforms baselines, demonstrating the efficacy of preference-calibrated binary signals in modeling inter-user differences.

Controlling Large Language Models (LLMs) to prevent the generation of undesirable content, such as profanity and personally identifiable information (PII), has become increasingly critical. While earlier approaches relied on post-processing or resampling, recent research has shifted towards constrained decoding methods that control outputs during generation to mitigate high computational costs and quality degradation. However, preventing multiple forbidden hard constraints or regex constraints from appearing anywhere in the output is computationally challenging. A straightforward solution is to convert these constraints into a single automaton that tracks all forbidden patterns during decoding, but this often becomes impractically large. Standard regex engines also do not readily support the operations needed to build such a constraint, such as complement and intersection. In order to address these limitations, we propose NCO, a decoding strategy that performs online pattern matching over finite hard constraints and regex constraints, reducing computational overhead without inducing state explosion. NCO is fully compatible with standard inference strategies, including various sampling methods and beam search, while also supporting soft masking for probabilistic suppression. We empirically demonstrate its effectiveness across practical tasks, including PII and profanity suppression. Our implementation is available at https://github.com/hyundong98/NCO-Decoding.git .

Pre-trained language models advance patent classification and retrieval via encoding claims as flat token sequences, yet overlooking the dependency hierarchy among claims. Incorporating the hierarchy into self-attention poses two challenges. First, claim dependencies involve relation types with varying reliability: treating them indiscriminately allows noisy technical relations to corrupt cleaner legal citation signals. Second, when the dependency graph is defined over claims, Transformer models fail as they operate at the token level; broadcasting claim-level adjacency can dilute structural information across unrelated token pairs. A novel Patent Heterogeneous Attention Graph Encoder (PHAGE) addresses these challenges. To handle heterogeneous dependencies, PHAGE constructs a typed graph to separate legal citations from technical relations as distinct edge types. To bridge the hierarchy gap, PHAGE introduces a connectivity mask with learnable relation-aware biases to project a claim-level topology into token-level attention. PHAGE learns a dual-granularity contrastive objective to align representations with inter-patent taxonomy and intra-patent topology. Experiments show that PHAGE outperforms domain-adapted and citation-aware baselines on patent classification, retrieval, and clustering. PHAGE discloses that the intra-patent claim topology captures stronger inductive bias than the inter-patent structure.

Large language models (LLMs) rely on web-scale corpora for pre-training. The noise inherent in these datasets tends to obscure meaningful patterns and ultimately degrade model performance. Data curation mitigates but cannot eliminate such noise, so pre-training corpora remain noisy in practice. We therefore study whether a lightweight pre-pre-training (PPT) stage based on synthetic data with learnable temporal structure helps resist noisy data during the pre-training (PT) stage. Across various corruption settings, our method consistently improves robustness to noise during PT, with larger relative gains at higher noise levels. For a 1B-parameter model, a synthetic PPT stage with only 65M tokens achieves the same final loss as the baseline while using up to 49\% fewer natural-text PT tokens across different noise levels. Mechanistic analyses suggest PPT does not immediately suppress attention to noisy tokens. Rather, PPT-initialized models gradually downweight attention between corrupted tokens during noisy PT. This indicates that synthetic PPT inhibits noise self-modeling and shapes the subsequent optimization trajectory. Code is available at https://github.com/guox18/formal-language-prepretraining.

Legal information in India remains largely inaccessible due to the complexity of legal language and the sheer volume of legal documentation involved in research and case analysis. This paper presents NyayaAI, an AI-powered legal assistant that automates and simplifies legal workflows for lawyers, law students, and general users. The system combines Large Language Models with a Retrieval-Augmented Generation pipeline grounded in a curated Indian legal knowledge base comprising constitutional provisions, statutes, case laws, and judicial precedents. A multi-agent architecture orchestrated through the Mastra TypeScript framework coordinates a main agent with specialized sub-agents handling legal research, document summarization, case law retrieval, and drafting assistance. A compliance module validates all responses before delivery. Domain classification achieved 70\% precision across test samples, with RAG retrieval precision at 74\% and overall response accuracy at 72\%, demonstrating that structured multi-agent LLM systems can meaningfully improve legal accessibility and workflow efficiency. The code\footnote{https://github.com/B97784/NyayaAI} is made publicly available for the benefit of the research community.

Every molecule ever synthesised can be drawn as a 2D skeletal diagram, yet in modern property prediction this universally available representation has received less focus in favour of molecular graphs, 3D conformers, or billion-parameter language models, each imposing its own computational and data-engineering overhead. We present $\textbf{MolSight}$, the first systematic large-scale study of vision-based Molecular Property Prediction (MPP). Using 10 vision architectures, 7 pre-training strategies, and $2\,M$ molecule images, we evaluate performance across 10 downstream tasks spanning physical-property regression, drug-discovery classification, and quantum-chemistry prediction. To account for the wide variation in structural complexity across pre-training molecules, we further propose a $\textbf{chemistry-informed curriculum}$: five structural complexity descriptors partition the corpus into five tiers of increasing chemical difficulty, consistently outperforming non-curriculum baselines. We show that a single rendered bond-line image, processed by a vision encoder, is sufficient for competitive molecular property prediction, i.e. $\textit{chemical insight from sight alone}$. The best curriculum-trained configuration achieves the top result on $\textbf{5 of 10}$ benchmarks and top two on $\textbf{all 10}$, at \textbf{\textit{80$\times$ lower}} FLOPs than the nearest multi-modal competitor.

Document-based question answering (QA) increasingly includes abstract questions that require synthesizing scattered information from long documents or across multiple documents into coherent answers. However, this setting is still poorly supported by existing benchmarks and evaluation methods, which often lack stable abstract references or rely on coarse similarity metrics and unstable head-to-head comparisons. To alleviate this issue, we introduce ASTRA-QA, a benchmark for AbSTRAct Question Answering over documents. ASTRA-QA contains 869 QA instances over academic papers and news documents, covering five abstract question types and three controlled retrieval scopes. Each instance is equipped with explicit evaluation annotations, including answer topic sets, curated unsupported topics, and aligned evidence. Building on these annotations, ASTRA-QA assesses whether answers cover required key points and avoid unsupported content by directly scoring topic coverage and curated unsupported content, enabling scalable evaluation without exhaustive head-to-head comparisons. Experiments with representative Retrieval-Augmented Generation (RAG) methods spanning vanilla, graph-based, and hierarchical retrieval settings show that ASTRA-QA provides reference-grounded diagnostics for coverage, hallucination, and retrieval-scope robustness. Our dataset and code are available at https://xinyangsally.github.io/astra-benchmark.

Scientific peer reviews frequently contain conflicting expert judgments, and the increasing scale of conference submissions makes it challenging for Area Chairs and editors to reliably identify and interpret such disagreements. Existing approaches typically frame reviewer disagreement as binary contradiction detection over isolated sentence pairs, abstracting away the review-level context and obscuring differences in the severity of evaluative conflict. In this work, we introduce a fine-grained formulation of reviewer contradiction analysis that operates over full peer reviews by explicitly identifying contradiction evidence spans and assigning graded disagreement intensity scores. To support this task, we present RevCI, an expert-annotated benchmark of peer-review pairs with evidence-level contradiction annotations with graded intensity labels. We further propose IMPACT, a structured multi-agent framework that integrates aspect-conditioned evidence extraction, deliberative reasoning, and adjudication to model reviewer contradictions and their intensity. To support efficient deployment, we distill IMPACT into TIDE, a small language model that predicts contradiction evidence and intensity in a single forward pass. Experimental results show that IMPACT substantially outperforms strong single-agent and generic multi-agent baselines in both evidence identification and intensity agreement, while TIDE achieves competitive performance at significantly lower inference cost.

Full-duplex spoken dialogue requires a model to keep listening while generating its own spoken response. This is challenging for large language models (LLMs), which are designed to extend a single coherent sequence and do not naturally support user input arriving during generation. We argue that how the user stream is routed into the LLM is therefore a key architectural question for full-duplex modeling. To study this question, we extend a text-only LLM into a unified full-duplex spoken dialogue system and compare two routing strategies under a shared training pipeline: (i) channel fusion, which injects the user stream directly into the LLM input, and (ii) cross-attention routing, which keeps the user stream as external memory accessed through cross-attention adapters. Experiments on spoken question answering and full-duplex interaction benchmarks reveal a clear tradeoff. Channel fusion yields stronger semantic grounding and consistently better question-answering performance. However, under semantically overlapping conditions such as user interruptions, it is more vulnerable to context corruption: if the model fails to stop in time, the overlapping user stream can interfere with ongoing generation and lead to semantically incoherent continuations. Cross-attention routing underperforms on question answering, but better preserves the LLM generation context and is more robust to this failure mode. These results establish user-stream routing as a central design axis in full-duplex spoken dialogue and offer practical guidance on the tradeoff between semantic integration and context robustness. We provide a demo page for qualitative inspection.

LLM decoding often relies on the model's predictive distribution to generate an output. Consequently, misalignment with respect to the true generating distribution leads to suboptimal decisions in practice. While a natural solution is to calibrate the model's output distribution, for LLMs, this is ill-posed at the combinatorially vast level of free-form language. We address this by building on the insight that in many tasks, these free-form outputs can be interpreted in a semantically meaningful latent structure, for example, discrete class labels, integers, or sets. We introduce task calibration as a paradigm to calibrate the model's predictive distribution in the task-induced latent space. We apply a decision-theoretic result to show that Minimum Bayes Risk (MBR) decoding on the task-calibrated latent distribution is the optimal decoding strategy on latent model beliefs. Empirically, it consistently improves generation quality across different tasks and baselines. We also introduce Task Calibration Error (TCE), an application-aware calibration metric that quantifies the excess loss due to miscalibration. Our work demonstrates that task calibration enables more reliable model decisions across various tasks and applications.

Government transparency laws, like the Freedom of Information (FOIA) acts in the United States and United Kingdom, and the Woo (Open Government Act) in the Netherlands, grant citizens the right to directly request documents from the government. As these documents might contain sensitive information, such as personal information or threats to national security, the laws allow governments to redact sensitive parts of the documents prior to release. We build on prior research to perform automatic sensitivity classification for the FOIA Exemption 5 deliberative process privilege using Large Language Models (LLMs). However, processing documents not yet cleared for review via third-party cloud APIs is often legally or politically untenable. Therefore, in this work, we perform sensitivity classification with a small, local model, deployable on consumer-grade hardware (Qwen3.5 9B). We compare eight variants of applying LLMs for sentence classification, using well-known prompting techniques, and find that a combination of Chain-of-Thought prompting and few-shot prompting with error-based examples outperforms classification models of earlier work in terms of recall and F2 score. This method also closely approaches the performance of a widely-used, cost-efficient commercial model (Gemini 2.5 Flash). In an additional analysis, we find that sentences that are predicted as deliberative contain more verbs that indicate the expression of opinions, and are more often phrased in in first-person. Above all, deliberativeness seems characterized by the presence of a combination of multiple indicators, in particular the combination of first-person words with a verb for expressing opinion.

Native Language Identification (NLI) is the task of determining an author's native language (L1) from their non-native writings. With the advent of human-AI co-authorship, non-native texts are routinely corrected and rewritten by large language models, fundamentally altering the linguistic features NLI models depend on. In this paper, we investigate the robustness of L1 traces across increasing degrees of editorial intervention. By processing 450 essays from the Write & Improve 2024 corpus through varying levels of grammatical error correction (GEC) and paraphrasing, we demonstrate that L1 attribution does not entirely depend on surface-level errors. Instead, the detection models leverage deeper L1 features: unidiomatic lexico-semantic choices, pragmatic transfer, and the author's underlying cultural perspective. We find that minimal edits preserve these structural traces and maintain high profiling accuracy. In contrast, fluency edits and paraphrasing normalize these L1 features, leading to a severe degradation in performance.

Diffusion large language models (dLLMs) offer a promising route to parallel and efficient text generation, but improving their reasoning ability requires effective post-training. Reinforcement learning with verifiable rewards (RLVR) is a natural choice for this purpose, yet its application to dLLMs is hindered by the absence of tractable sequence-level log-ratios, which are central to standard policy optimization. The lack of tractable sequence-level log-ratios forces existing methods to rely on high-variance ELBO-based approximations, where high verifier rewards can amplify inaccurate score estimates and destabilize RL training. To overcome this issue, we propose \textbf{R}elative \textbf{S}core \textbf{P}olicy \textbf{O}ptimization (RSPO), a simple RLVR method that uses verifiable rewards to calibrate noisy likelihood estimates in dLLMs. The core of our algorithm relies on a key observation: a reward advantage can be interpreted not only as an update direction, but also as a target for the relative log-ratio between the current and reference policies. Accordingly, RSPO calibrates this noisy relative log-ratio estimate by comparing its reward advantage with the reward-implied target relative log-ratio, updating the policy according to the gap between the current estimate and the target rather than the raw advantage alone. Experiments on mathematical reasoning and planning benchmarks show that RSPO yields especially strong gains on planning tasks and competitive mathematical-reasoning performance.

Recent advances in large language models (LLMs) have expanded the context window to beyond 128K tokens, enabling long-document understanding and multi-source reasoning. A key challenge, however, lies in choosing between retrieval-augmented generation (RAG) and long-context (LC) strategies: RAG is efficient but constrained by retrieval quality, while LC supports global reasoning at higher cost and with position sensitivity. Existing methods such as Self-Route adopt failure-driven fallback from RAG to LC, but remain passive, inefficient, and hard to interpret. We propose Pre-Route, a proactive routing framework that performs structured reasoning before answering. Using lightweight metadata (e.g., document type, length, initial snippet), Pre-Route enables task analysis, coverage estimation, and information-need prediction, producing explainable and cost-efficient routing decisions. Our study shows three key findings: (i) LLMs possess latent routing ability that can be reliably elicited with guidelines, allowing single-sample performance to approach that of multi-sample (Best-of-N) results; (ii) linear probes reveal that structured prompts sharpen the separability of the "optimal routing dimension" in representation space; and (iii) distillation transfers this reasoning structure to smaller models for lightweight deployment. Experiments on LaRA (in-domain) and LongBench-v2 (OOD) confirm that Pre-Route outperforms Always-RAG, Always-LC, and Self-Route baselines, achieving superior overall cost-effectiveness.

Natural language understanding (NLU) is integral to task-oriented dialog systems, but demands a considerable amount of annotated training data to increase the coverage of diverse utterances. In this study, we report the construction of a linguistic resource named FIAD (Financial Annotated Dataset) and its use to generate a Korean annotated training data for NLU in the banking customer service (CS) domain. By an empirical examination of a corpus of banking app reviews, we identified three linguistic patterns occurring in Korean request utterances: TOPIC (ENTITY, FEATURE), EVENT, and DISCOURSE MARKER. We represented them in LGGs (Local Grammar Graphs) to generate annotated data covering diverse intents and entities. To assess the practicality of the resource, we evaluate the performances of DIET-only (Intent: 0.91 /Topic [entity+feature]: 0.83), DIET+ HANBERT (I:0.94/T:0.85), DIET+ KoBERT (I:0.94/T:0.86), and DIET+ KorBERT (I:0.95/T:0.84) models trained on FIAD-generated data to extract various types of semantic items.

To tackle long-context reasoning tasks without the quadratic complexity of standard attention mechanisms, approaches based on agent memory have emerged, which typically maintain a dynamically updated memory when linearly processing document chunks. To mitigate the potential loss of latent evidence in this memorize-while-reading paradigm, recent works have integrated retrieval modules that allow agents to recall information previously discarded during memory overwriting. However, retrieval-based recall suffers from both evidence loss during memory formation and interference induced by invalid queries. To overcome these limitations, we propose MemReread. Built upon streaming reading, MemReread circumvents intermediate retrieval. It triggers question decomposition and rereading when the final memory is insufficient, enabling the recovery of indirect facts that were prematurely discarded. This design supports non-linear reasoning while preserving the inherent logical flow of document comprehension. To further enhance practicality, we introduce a reinforcement learning framework that enhances length extrapolation capability while dynamically determining the number of rereading passes based on task complexity, thereby flexibly controlling computational overhead. Extensive experiments demonstrate that MemReread consistently outperforms baseline frameworks on long-context reasoning tasks, while maintaining linear time complexity with respect to context length.

This paper aims to construct a linguistic resource of Korean Multiword Expressions for Feature-Based Sentiment Analysis (FBSA): DECO-MWE. Dealing with multiword expressions (MWEs) has been a critical issue in FBSA since many constructs reveal lexical idiosyncrasy. To construct linguistic resources of sentiment MWEs efficiently, we utilize the Local Grammar Graph (LGG) methodology: DECO-MWE is formalized as a Finite-State Transducer that represents lexical-syntactic restrictions on MWEs. In this study, we built a corpus of cosmetics review texts, which show particularly frequent occurrences of MWEs. Based on an empirical examination of the corpus, four types of MWEs have been distinguished. The DECO-MWE thus covers the following four categories: Standard Polarity MWEs (SMWEs), Domain-Dependent Polarity MWEs (DMWEs), Compound Named Entity MWEs (EMWEs) and Compound Feature MWEs (FMWEs). The retrieval performance of the DECO-MWE shows 0.806 f-measure in the test corpus. This study brings a twofold outcome: first, a sizeable general-purpose polarity MWE lexicon, which may be broadly used in FBSA; second, a finite-state methodology adopted in this study to treat domain-dependent MWEs such as idiosyncratic polarity expressions, named entity expressions or feature expressions, and which may be reused in describing linguistic properties of other corpus domains.

We participated in the Fifth UNLP shared task on multi-domain document understanding, where systems must answer Ukrainian multiple-choice questions from PDF collections and localize the supporting document and page. We propose a retrieval-augmented pipeline built around three ideas: contextual chunking of PDFs, question-aware dense retrieval and reranking conditioned on both the question and answer options, and constrained answer generation from a small set of reranked passages. Our final system uses Qwen3-Embedding-8B for retrieval, a fine-tuned Qwen3-Reranker-8B for passage ranking, and Qwen3-32B for answer selection. On a held-out split, reranking improves Recall@1 from 0.6957 to 0.7935, while using the top-2 reranked passages raises answer accuracy from 0.9348 to 0.9674. Our best leaderboard run reached 0.9452 on the public leaderboard and 0.9598 on the private leaderboard. Our results suggest that, under strict code-competition constraints, preserving document structure and making relevance estimation aware of the answer space are more effective than adding complex downstream heuristics.

A central challenge in large-scale decision-making under incomplete information is estimating reliable probabilities. Recent approaches use Large Language Models (LLMs) to generate explanatory factors and coarse-grained probability estimates, which are then refined by a Naïve Bayes model over factor combinations. However, sparse factor spaces often yield ``unknown'' predictions, while expanding factors increases noise and spurious correlations, weakening conditional independence and degrading reliability. To address these limitations, we propose \textsc{Anchor}, an aggregated Bayesian inference framework over a hierarchical factor space. It constructs dense factor hierarchies through iterative generation and clustering, maps contexts via hierarchical retrieval and refinement, and augments Naïve Bayes with a Causal Bayesian Network to model latent factor dependencies. Experiments show that \textsc{Anchor} markedly reduces ``unknown'' predictions and produces more reliable probability estimates than direct LLM baselines, achieving state-of-the-art performance while significantly reducing time and token overhead.

The advancement of Large Language Models (LLMs) has enabled their application in personalized dialogue systems. We present an extended annotation scheme for personal fact classification that addresses limitations in existing approaches, particularly PeaCoK. Our scheme introduces new categories (Demographics, Possessions) and attributes (Duration, Validity, Followup) that enable structured storage, quality filtering, and identification of facts suitable for dialogue continuation. We manually annotated 2,779 facts from Multi-Session Chat and trained a multi-head classifier based on transformer encoders. Combined with the Gemma-300M encoder, the classifier achieves $81.6 \pm 2.6$\% macro F1, outperforming all few-shot LLM baselines (best: GPT-5.4-mini, 72.92\%) by nearly 9 percentage points while requiring substantially fewer computational resources. Error analysis reveals persistent challenges in semantic boundary disambiguation, temporal aspect interpretation, and pragmatic reasoning for followup assessment. The dataset\footnotemark[1] and classifier\footnotemark[2] are publicly available.

Large language models (LLMs) have become capable mathematical problem-solvers, often producing correct proofs for challenging problems. However, correctness alone is not sufficient: mathematical proofs should also be clear, concise, insightful, and transferable to other problems. While this proof quality is subjective and depends on the reader and context, many of its components are concrete and broadly valued. In this work, we identify such components and introduce ProofRank, a benchmark curated from challenging mathematical competitions. ProofRank evaluates several scalable proxies of proof quality: (i) conciseness, measuring whether proofs avoid unnecessary steps; (ii) computational ease, measuring the extent to which a proof relies on tedious calculations; (iii) cognitive simplicity, measuring how accessible the used proof techniques are; (iv) diversity, measuring how varied a model's proofs for a single problem are; and (v) adaptivity, measuring whether a model can follow a specified proof technique. Across models, we find substantial differences in proof quality that are not captured by correctness-only benchmarks. We also observe significant trade-offs between proof-quality metrics and correctness, suggesting that future evaluations of mathematical reasoning should measure how useful LLM-generated proofs are.

We introduce Phoenix-VL 1.5 Medium, a 123B-parameter natively multimodal and multilingual foundation model, adapted to regional languages and the Singapore context. Developed as a sovereign AI asset, it demonstrates that deep domain adaptation can be achieved with minimal degradation to broad-spectrum intelligence and alignment. Continued pretraining was performed on Mistral Medium 3.1 using a localized 1-trillion tokens multimodal corpus, followed by a 250-billion tokens long-context extension phase. Subsequent post-training incorporated a novel human-annotated Singapore multimodal dataset and curated textual corpus on Singapore culture, knowledge, and legislation, totaling 22-billion tokens. An additional 5 billion tokens of model alignment was performed through Online Direct Preference Optimization. Phoenix-VL 1.5 Medium achieves state-of-the-art performance for its size on Singapore multimodal, legal, and government policy benchmarks while remaining globally competitive on general multimodal intelligence, multilingual, and STEM benchmarks. We also introduce a novel evaluation suite encompassing localized knowledge benchmarks and an institutionally aligned model behavior and safety framework. We report the data curation principles, training methodology, and highlight benchmark and inference performance.

Multilingual studies of social bias in open-ended LLM generation remain limited: most existing benchmarks are English-centric, template-based, or restricted to recognizing pre-specified stereotypes. We introduce StereoTales, a multilingual dataset and evaluation pipeline for systematically studying the emergence of social bias in open-ended LLM generation. The dataset covers 10 languages and 79 socio-demographic attributes, and comprises over 650k stories generated by 23 recent LLMs, each annotated with the socio-demographic profile of the protagonist across 19 dimensions. From these, we apply statistical tests to identify more than 1{,}500 over-represented associations, which we then rate for harmfulness through both a panel of humans (N = 247) and the same LLMs. We report three main findings. \textbf{(i)} Every model we evaluate emits consequential harmful stereotypes in open-ended generation, regardless of size or capabilities, and these associations are largely shared across providers rather than isolated misbehaviors. \textbf{(ii)} Prompt language strongly shapes which stereotypes appear: rather than transferring as a shared set of biases, harmful associations adapt culturally to the prompt language and amplify bias against locally salient protected groups. \textbf{(iii)} Human and LLM harmfulness judgments are broadly aligned (Spearman $ρ=0.62$), with disagreements concentrating on specific attribute classes rather than specific providers. To support further analyses, we release the evaluation code and the dataset, including model generations, attribute annotations, and harmfulness ratings.

Formalisation is the process of writing system requirements in a formal language. These requirements mostly originate in Natural Language. In the field of Formal Methods, formalisation is often identified as one of the most delicate and complicated steps in the verification process. Not seldomly, formalisation tools and environments choose various levels of requirement descriptions: Natural Language, Technical Language, Diagram Representations and Formal Language, to mention a few. In the literature, there are various maxims and principles of good practice to guide the process of requirement formalisation. In this paper we propose a new guideline: Coherency through Formalisations. The guideline states that the different levels of formalisation mentioned above should roughly follow the same logical structure. The principle seems particularly relevant in the setting where LLMs are prompted to perform reasoning tasks that can be checked by formal tools using Structured Natural Language to act as an intermediate layer bridging both paradigms. In the light of coherency, we analyze NASA's Formal Requirement Elicitation Tool FRET and propose an alternative automated translation of the Controlled Natural Language FRETish to the formal language of MTL. We compare our translation to the original translation and prove equivalence using model checking. Some statistics are performed which seem to favor the new translation. As expected, the translation process yielded interesting reflections and revealed inconsistencies which we present and discuss.

Masked Diffusion Models (MDMs) have emerged as a promising alternative to autoregressive models in language modeling, offering the advantages of parallel decoding and bidirectional context processing within a simple yet effective framework. Specifically, their explicit distinction between masked tokens and data underlies their simple framework and effective conditional generation. However, MDMs typically require many sampling iterations due to factorization errors stemming from simultaneous token updates. We observe that a theoretical lower bound of the factorization error exists, which standard MDMs cannot reduce due to their use of a deterministic single-state mask. In this paper, we propose the Infinite Mask Diffusion Model (IMDM), which introduces a stochastic infinite-state mask to mitigate the theoretical bound while directly inheriting the benefits of MDMs, including the compatibility with pre-trained weights. We empirically demonstrate that MDM fails to perform few-step generation even in a simple synthetic task due to the factorization error bound, whereas IMDM can find an efficient solution for the same task. Finally, when equipped with appropriate distillation methods, IMDM surpasses existing few-step distillation methods at small step counts on LM1B and OpenWebText. Code is available at https://Ugness.github.io/official_imdm.

Document classification forms the backbone of modern enterprise content management, yet existing benchmarks remain trapped in oversimplified paradigms – single domain settings with flat label structures – that bear little resemblance to the hierarchical, multi-modal, and cross-domain nature of real-world business documents. This gap not only misrepresents practical complexity but also stifles progress toward industrially viable document intelligence. To bridge this gap, we construct the first Multi-level, Multi-domain, Multi-modal document classification Benchmark (MMM-Bench). MMM-Bench includes (1) a deeply hierarchical taxonomy spanning five levels that capture the authentic organizational logic of business documentation; and (2) 5,990 real-world multi-modal documents meticulously curated from 12 commercial domains in Alibaba. Each document is manually annotated with a complete hierarchical path by domain experts. We establish comprehensive baselines on MMM-Bench, which consists of open-weight models and API-based models. Through systematic experiments, we identify four fundamental challenges within MMM-Bench and propose corresponding insights. To provide a solid foundation for advancing research in multi-level, multi-domain document classification, we release all of the data and the evaluation toolkit of MMM-Bench at https://github.com/MMMDC-Bench/MMMDC-Bench.

This paper describes our system to SemEval-2026 Task 3 Track A Subtask 1 on Dimensional Aspect Sentiment Regression (DimASR). We propose a lightweight and resource-efficient system built entirely on multilingual pre-trained encoders, without relying on LLMs or external corpora. We adopt joint multilingual and multi-domain training to facilitate cross-lingual transfer and alleviate data sparsity, introduce a bounded regression transformation that improves training stability while constraining predictions within the valid range, and employ an adaptive ensemble strategy via subset search to reduce prediction variance. Experimental results demonstrate that our system achieves strong and consistent performance, ranking 1st on zho-res, 2nd on zho-lap, and 3rd on jpn-hot, with all remaining datasets placed within the top half of participating teams.

Threat detection in Natural Language Processing lacks consistent definitions and standardized benchmarks, and is often conflated with broader phenomena such as toxicity, hate speech, or offensive language. In this work, we introduce ThreatCore, a public available benchmark dataset for fine-grained threat detection that distinguishes between explicit threats, implicit threats, and non-threats. The dataset is constructed by aggregating multiple publicly available resources and systematically re-annotating them under a unified operational definition of threat, revealing substantial inconsistencies across existing labels. To improve the coverage of underrepresented cases, particularly implicit threats, we further augment the dataset with synthetic examples, which are manually validated using the same annotation protocol adopted for the re-annotation of the public datasets, ensuring consistency across all data sources. We evaluate Perspective API, zero-shot classifiers, and recent language models on ThreatCore, showing that implicit threats remain substantially harder to detect than explicit ones. Our results also indicate that incorporating Semantic Role Labeling as an intermediate representation can improve performance by making the structure of harmful intent more explicit. Overall, ThreatCore provides a more consistent benchmark for studying fine-grained threat detection and highlights the challenges that current models still face in identifying indirect expressions of harmful intent.

Training AI agents to proactively assist humans in daily activities, from routine household tasks to urgent safety situations, requires large-scale visual data. However, capturing such scenarios in the real world is often difficult, costly, or unsafe, and physics-based simulators lack the visual fidelity needed to transfer learned behaviors to real settings. Therefore, we introduce VISTA, a video synthesis system that produces high-fidelity egocentric videos as training and evaluation data for AI agents. VISTA employs a 5-step script generation pipeline with causal reverse reasoning to create diverse, logically grounded intervention modes. These scenarios span two levels of agent autonomy: reactive and proactive. In reactive modes, the user explicitly asks the agent for help. In proactive modes, the agent offers help without receiving a direct request. We further divide proactive modes into explicit and implicit types. In explicit proactive scenarios, the user is aware of needing help but does not directly address the agent. In implicit proactive scenarios, the agent intervenes before the user even realizes that help is needed. VISTA allows users to customize and refine scenarios to generate video benchmarks for daily tasks, offering a scalable and controllable alternative to real-world data collection for training and evaluating AI agents in realistic environments.

Expressions with an aspectual variant of a light verb, e.g. 'take on debt' vs. 'have debt', are frequent in texts but often difficult to classify between verbal idioms, light verb constructions or compositional phrases. We investigate the properties of such expressions with a disputed membership and propose a selection of features that determine more satisfactory boundaries between the three categories in this zone, assigning the expressions to one of them.

Large language models (LLMs) can convincingly imitate human writing styles, yet it remains unclear how much stylistic information is encoded in embeddings from any language model and retained after LLM rewriting. We investigate these questions in French, using a controlled literary dataset to quantify the effect of stylistic variation via changes in embedding dispersion. We observe that embeddings reliably capture authorial stylistic features and that these signals persist after rewriting, while also exhibiting LLM-specific patterns. These analytical results offer promising directions for authorship imitation detection in the era of language models.

Many studies have shown automatic speech processing (ASR) systems have unequal performance across speakergroups (SG's). However, the manner in which such studies arrive at this conclusion is inconsistent. To pave the wayfor more reliable results in future studies, we lay out best practices for benchmarking ASR fairness based on literaturefrom machine learning fairness, social sciences, and speech science. We first describe the importance of preciselythe fairness hypothesis being interrogated, and tailoring fairness metrics to apply specifically to said hypothesis.We then examine several benchmarks used to rate ASR systems on fairness and discuss how their results can bemisconstrued without assiduous oversight into the intersections between SG's. We find that evaluating fairnessbased on single heterogeneous SG's, such as they are defined in fairness benchmarks, can lead to misidentifyingwhich SG's are actually being mistreated by ASR systems. We advocate for as fine-grained an analysis as possibleof the intersectionality of as many demographic variables as are available in the metadata of fairness corpora in orderto tease out such spurious correlations

Tabular Foundation Models have recently established the state of the art in supervised tabular learning, by leveraging pretraining to learn generalizable representations of numerical and categorical structured data. However, they lack native support for unstructured modalities such as text and image, and rely on frozen, pretrained embeddings to process them. On established Multimodal Tabular Learning benchmarks, we show that tuning the embeddings to the task improves performance. Existing benchmarks, however, often focus on the mere co-occurrence of modalities; this leads to high variance across datasets and masks the benefits of task-specific tuning. To address this gap, we introduce MulTaBench, a benchmark of 40 datasets, split equally between image-tabular and text-tabular tasks. We focus on predictive tasks where the modalities provide complementary predictive signal, and where generic embeddings lose critical information, necessitating Target-Aware Representations that are aligned with the task. Our experimental results demonstrate that the gains from target-aware representation tuning generalize across both text and image modalities, several tabular learners, encoder scales, and embedding dimensions. MulTaBench constitutes the largest image-tabular benchmarking effort to date, spanning high-impact domains such as healthcare and e-commerce. It is designed to enable the research of novel architectures which incorporate joint modeling and target-aware representations, paving the way for the development of novel Multimodal Tabular Foundation Models.

Coreference resolution is typically evaluated using aggregate statistical metrics such as CoNLL-F1, which measure structural overlap between predicted and gold clusters. While widely used, these metrics offer limited diagnostic insights, penalizing errors without revealing whether a system struggles with specific semantic categories, such as people, locations, or events, and making it difficult to interpret model capabilities or derive actionable improvements. We address this gap by introducing a semantically-enhanced evaluation framework for coreference resolution. Our approach overlays Concept and Named Entity Recognition (CNER) onto coreference outputs, assigning semantic labels to nominal mentions and propagating them to entire coreference clusters. This enables the computation of typed scores aimed at evaluating mention extraction and linking capabilities stratified by semantic class. Across our experiments on OntoNotes, LitBank, and PreCo, we show that our framework uncovers systematic weaknesses that remain obscured by aggregate metrics. Furthermore, we demonstrate that these diagnostics can be used to design targeted, low-cost data augmentation strategies, achieving measurable out-of-domain improvements.

Fine-tuning Large Language Models (LLMs) on benign narrow data can sometimes induce broad harmful behaviors, a vulnerability termed emergent misalignment (EM). While prior work links these failures to specific directions in the activation space, their relationship to the model's broader persona remains unexplored. We map the latent personality space of LLMs through established psychometric profiles like the Big Five, Dark Triad, and LLM-specific behaviors (e.g. evil, sycophancy), and show that the semantic geometry is highly stable across aligned models and their corrupted fine-tunes. Through causal interventions, we find that directions isolating social valence, such as the 'Evil' persona vector, and a Semantic Valence Vector (SVV) that we introduce, function as intrinsic guardrails: ablating them drives the misalignment rates above $40$%, while amplifying them suppresses the failure mode to less than $3$%. Leveraging the structural stability of the personality space, we show that vectors extracted $\textit{a priori}$ from an instruct-tuned model transfer zero-shot to successfully regulate EM in corrupted fine-tunes. Overall, our findings suggest that harmful fine-tuning does not overwrite a model's internal representation of personality, allowing conserved representations to serve as robust, cross-distribution guardrails.

Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer, offering a unified explanation for the behavior of representative CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods succeed in shifting convergence dynamics and stabilizing pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called \textbf{S}electing \textbf{T}okens via attenti\textbf{O}n \textbf{C}ontribution~(STOC), which identifies influential factual snippets to guide replay data generation. Extensive experiments on both synthetic and real-world datasets validate our findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.

Digital personas powered by Large Language Models (LLMs) are increasingly proposed as substitutes for human survey respondents, yet it remains unclear when they can reliably approximate human survey findings. We answer this question using the LISS panel, constructing personas from respondents' background variables and pre-2023 survey histories, then testing them against the same respondents' held-out post-cutoff answers. Across four persona architectures, three LLMs, and two prediction tasks, we assess performance at the question, respondent, distributional, equity, and clustering levels. Digital personas improve alignment with human response distributions, especially in domains tied to stable attributes and values, but remain limited for individual prediction and fail to recover multivariate respondent structure. Retrieval-augmented architectures provide the clearest gains, but performance depends more on human response structure than on model choice: personas perform best for low-variability questions and common respondent patterns, and worst for subjective, heterogeneous, or rare responses. Our results provide practical guidance on when digital personas could be appropriate for survey research and when human validation remains necessary.

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Rejection Fine-Tuning (RFT) is a standard method for training LLM agents, where unsuccessful trajectories are discarded from the training set. In the context of SWE-bench tasks, this corresponds to filtering out runs where the submitted patch does not pass the tests. However, this approach discards unresolved trajectories, even though they form a large portion of all trajectories for hard tasks and even then may be partially correct. In this work, we propose Step Rejection Fine-Tuning (SRFT) - a practical way to leverage these unresolved trajectories. For this, we employ a critic LLM to assess the correctness of each step in a trajectory. Consequently, during training, we mask the loss for erroneous steps while retaining them in the context window. This way we ensure the model learns to recover from errors without reproducing them. Evaluation on SWE-bench Verified shows that while RFT improves the resolution rate by 2.4% by excluding unresolved trajectories, SRFT improves it by 3.7% by filtering them instead of discarding completely, reaching the total resolution rate of 32.2%.

Cross-lingual transfer has become a central paradigm for extending natural language processing (NLP) technologies to low-resource languages. By leveraging supervision from high-resource languages, multilingual language models can achieve strong task performance with little or no labeled target-language data. However, it remains unclear to what extent cross-lingual transfer can substitute for language-specific efforts. In this paper, we synthesize prior research findings and data collection results on Luxembourgish, which, despite its typological proximity to high-resource languages and its presence in a multilingual context, remains insufficiently represented in modern NLP technologies. Across findings, we observe a fundamental interdependence between cross-lingual transfer and language-specific efforts. Cross-lingual transfer can substantially improve target-language performance, but its success depends critically on the availability of sufficiently high-quality, task-aligned target-language data. At the same time, such resources, particularly in low-resource settings, are typically too limited in scale to drive strong performance on their own. Instead, such resources reach their full potential only when leveraged within a cross-lingual framework. We therefore argue that cross-lingual transfer and language-specific efforts should not be viewed as competing alternatives. Instead, they function as complementary components of a sustainable low-resource NLP pipeline. Based on these insights, we provide practical guidelines for integrating and balancing cross-lingual transfer with language-specific development in sustainable low-resource NLP pipelines.

Artificial intelligence safety research focuses on aligning individual language models with human values, yet deployed AI systems increasingly operate as interacting populations where social influence may override individual alignment. Here we show that populations of individually aligned AI agents can be driven into stable misaligned states through conformity dynamics. Simulating opinion dynamics across nine large language models and one hundred opinion pairs, we find that each agent's behavior is governed by two competing forces: a tendency to follow the majority and an intrinsic bias toward specific positions. Using tools from statistical physics, we derive a quantitative theory that predicts when populations become trapped in long-lived misaligned configurations, and identifies predictable tipping points where small numbers of adversarial agents can irreversibly shift population-level alignment even after manipulation ceases. These results demonstrate that individual-level alignment provides no guarantee of collective safety, calling for evaluation frameworks that account for emergent behavior in AI populations.

Self-distillation has emerged as a powerful framework for post-training LLMs, where a teacher conditioned on extra information guides a student without it, both from the same model. While this guidance is useful when the student has failed, on successful rollouts, the same mechanism instead overwrites the student's choices and suppresses it's own reasoning. Therefore, we propose reading the original self-distillation signal in reverse: when the student succeeds along a path the teacher would not have predicted, these tokens reflect its self-driven reasoning. Building on this, we propose RLRT (RLVR with Reversed Teacher), which augments GRPO by reinforcing these tokens on correct rollouts. We interpret this as a new form of exploration in RLVR: not uniform diversity, but valuable exploration grounded in the student's own success. Across base, instruction-tuned, and thinking-tuned Qwen3 checkpoints, RLRT substantially outperforms self-distillation and exploration-based baselines, establishing information asymmetry as a new, principled design axis for RLVR.

Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.

Large language models increasingly mediate decisions that turn on moral judgement, yet a growing body of evidence shows that their implicit preferences are not culturally neutral. Existing cultural alignment methods either require per-country preference data and fine-tuning budgets or assume white-box access to model internals that commercial APIs do not expose. In this work, we focus on this realistic black-box, public-data-only regime and observe that within-country sociodemographic disagreement, not consensus, is the primary steering signal. We introduce DISCA (Disagreement-Informed Steering for Cultural Alignment), an inference-time method that instantiates each country as a panel of World-Values-Survey-grounded persona agents and converts their disagreement into a bounded, loss-averse logit correction. Across 20 countries and 7 open-weight backbones (2B–70B), DISCA reduces cultural misalignment on MultiTP by 10–24% on the six backbones >=3.8B, and 2–7% on open-ended scenarios, without changing any weights. Our results suggest that inference-time calibration is a scalable alternative to fine-tuning for serving the long tail of global moral preferences.

As global cross-lingual communication intensifies, language barriers in visually rich documents such as PDFs remain a practical bottleneck. Existing document translation pipelines face a tension between linguistic processing and layout preservation: text-oriented Computer-Assisted Translation (CAT) systems often discard structural metadata, while document parsers focus on extraction and do not support faithful re-rendering after translation. We introduce BabelDOC, an Intermediate Representation (IR)-based framework for layout-preserving PDF translation. BabelDOC decouples visual layout metadata from semantic content, enabling document-level translation operations such as terminology extraction, cross-page context handling, glossary-constrained generation, and formula placeholdering. The translated content is then re-anchored to the original layout through an adaptive typesetting engine. Experiments on a curated 200-page benchmark, together with human evaluation and multimodal LLM-as-a-judge evaluation, show that BabelDOC improves layout fidelity, visual aesthetics, and terminology consistency over representative baselines, while maintaining competitive translation precision. The open-source toolkit and its interactive downstream applications are publicly available and have attracted over 8.4K GitHub stars and 17 contributors at the time of writing. A demonstration video is also available.

Does a lexical retriever suffice as large language models (LLMs) become more capable in an agentic loop? This question naturally arises when building deep research systems. We revisit it by pairing BM25 with frontier LLMs that have better reasoning and tool-use abilities. To support researchers asking the same question, we introduce Pi-Serini, a search agent equipped with three tools for retrieving, browsing, and reading documents. Our results show that, on BrowseComp-Plus, a well-configured lexical retriever with sufficient retrieval depth can support effective deep research when paired with more capable LLMs. Specifically, Pi-Serini with gpt-5.5 achieves 83.1% answer accuracy and 94.7% surfaced evidence recall, outperforming released search agents that use dense retrievers. Controlled ablations further show that BM25 tuning improves answer accuracy by 18.0% and surfaced evidence recall by 11.1% over the default BM25 setting, while increasing retrieval depth further improves surfaced evidence recall by 25.3% over the shallow-retrieval setting. Source code is available at https://github.com/justram/pi-serini.

Humor generation remains challenging task for Large Language Models (LLMs), due to their subjective nature. We focus on satire, a form of humor strongly shaped by context. In this work, we present a novel pipeline for grounded satire generation that uses Retrieval-Augmented Generation (RAG) over current news to produce satirical dictionary definitions in the Finnish context. We also introduce a new task-specific evaluation framework and annotate 100 generated definitions with six human annotators, enabling analysis across multiple experimental conditions, including cultural background, source-word type, and the presence or absence of RAG. Our results show that the generated definitions are perceived as more political than humorous. Both topic-based word selection and RAG improve the political relevance of the outputs, but neither yields clear gains in humor generation. In addition, our LLM-as-a-judge evaluation of five state-of-the-art models indicates that LLMs correlate well with human judgments on political relevance, but perform poorly on humor. We release our code and annotated dataset to support further research on grounded satire generation and evaluation.

Automated question answering (QA) over electronic health records (EHRs) demands precise evidence retrieval, faithful answer generation, and explicit grounding of answers in clinical notes. In this work, we present Neural1.5, our method for the ArchEHR-QA 2026 shared task at CL4Health@LREC 2026, which comprises four subtasks: question interpretation, evidence identification, answer generation, and evidence alignment. Our approach decouples the task into independent, modular stages and employs DSPy"s MIPROv2 optimizer to automatically discover high-performing prompts, jointly tuning instructions and few-shot demonstrations for each stage. Within every stage, self-consistency voting over multiple stochastic inference runs suppresses spurious errors and improves reliability, while stage-specific verification mechanisms (e.g., self-reflection and chain-of-verification for alignment) further refine output quality. Among all teams that participated in all four subtasks, our method ranks second overall (mean rank 4.00), placing 4th, 1st, 4th, and 7th on Subtasks 1-4, respectively. These results demonstrate that systematic, per-stage prompt optimization combined with self-consistency mechanisms is a cost-effective alternative to model fine-tuning for multifaceted clinical QA.

Large vision-language models suffer from visual ungroundedness: they can produce a fluent, confident, and even correct response driven entirely by language priors, with the image contributing nothing to the prediction. Existing confidence estimation methods cannot detect this, as they observe model behavior under normal inference with no mechanism to determine whether a prediction was shaped by the image or by text alone. We introduce BICR (Blind-Image Contrastive Ranking), a model-agnostic confidence estimation framework that makes this contrast explicit during training by extracting hidden states from a frozen LVLM twice: once with the real image-question pair, and once with the image blacked out while the question is held fixed. A lightweight probe is trained on the real-image hidden state and regularized by a ranking loss that penalizes higher confidence on the blacked-out view, teaching it to treat visual grounding as a signal of reliability at zero additional inference cost. Evaluated across five modern LVLMs and seven baselines on a benchmark covering visual question answering, object hallucination detection, medical imaging, and financial document understanding, BICR achieves the best cross-LVLM average on both calibration and discrimination simultaneously, with statistically significant discrimination gains robust to cluster-aware analysis at 4-18x fewer parameters than the strongest probing baseline.

Training deep research agents, namely systems that plan, search, evaluate evidence, and synthesize long-form reports, pushes reinforcement learning beyond the regime of verifiable rewards. Their outputs lack ground-truth answers, their trajectories span many tool-augmented decisions, and standard post-training offers little mechanism for turning past attempts into reusable experience. In this work, we argue that rubrics should serve not merely as final-answer evaluators, but as the shared interface that structures policy execution, judge feedback, and agent memory. Based on this view, we introduce RubricEM, a rubric-guided reinforcement learning framework that combines stagewise policy decomposition with reflection-based meta-policy evolution. RubricEM first makes research trajectories stage-aware by conditioning planning, evidence gathering, review, and synthesis on self-generated rubrics. It then assigns credit with Stage-Structured GRPO, which uses stagewise rubric judgments to provide denser semantic feedback for long-horizon optimization. In parallel, RubricEM trains a shared-backbone reflection meta-policy that distills judged trajectories into reusable rubric-grounded guidance for future attempts. The resulting RubricEM-8B achieves strong performance across four long-form research benchmarks, outperforming comparable open models and approaching proprietary deep-research systems. Beyond final performance, we perform thorough analyses to understand the key ingredients of RubricEM.

Large language and vision-language models increasingly power agents that act on a user's behalf through command-line interface (CLI) harnesses. However, most agent benchmarks still rely on synthetic sandboxes, short-horizon tasks, mock-service APIs, and final-answer checks, leaving open whether agents can complete realistic long-horizon work in the runtimes where they are deployed. This work presents WildClawBench, a native-runtime benchmark of 60 human-authored, bilingual, multimodal tasks spanning six thematic categories. Each task averages roughly 8 minutes of wall-clock time and over 20 tool calls, and runs inside a reproducible Docker container hosting an actual CLI agent harness (OpenClaw, Claude Code, Codex, or Hermes Agent) with access to real tools rather than mock services. Grading is hybrid, combining deterministic rule-based checks, environment-state auditing of side effects, and an LLM/VLM judge for semantic verification. Across 19 frontier models, the best, Claude Opus 4.7, reaches only 62.2% overall under OpenClaw, while every other model stays below 60%, and switching harness alone shifts a single model by up to 18 points. These results show that long-horizon, native-runtime agent evaluation remains a far-from-resolved task for current frontier models. We release the tasks, code, and containerized tooling to support reproducible evaluation.

While Mixture-of-Experts (MoE) scales model capacity without proportionally increasing computation, its massive total parameter footprint creates significant storage and memory-access bottlenecks, which hinder efficient end-side deployment that simultaneously requires high performance, low computational cost, and small storage overhead. To achieve these properties, we present DECO, a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. DECO utilizes the differentiable and flexible ReLU-based routing enhanced by learnable expert-wise scaling, which adaptively balances the contributions of routed and shared experts. Furthermore, we introduce NormSiLU, an activation function that normalizes inputs prior to SiLU operators, producing a more stable trend of routed-expert activation ratio and a higher intrinsic sparsity level. We also identify an empirical advantage in using non-gated MLP experts with ReLU-based routing, indicating the possibility of MoE architecture simplification. Experiments demonstrate that DECO, activating only 20% of routed experts, matches dense performance and outperforms established MoE baselines. Our specialized acceleration kernel delivers a 2.93$\times$ speedup on Jetson AGX Orin compared with dense inference. Code and checkpoints are available at https://github.com/thunlp/DECO.

Diffusion and flow-based models have become the de facto approaches for generating continuous data, e.g., in domains such as images and videos. Their success has attracted growing interest in applying them to language modeling. Unlike their image-domain counterparts, today's leading diffusion language models (DLMs) primarily operate over discrete tokens. In this paper, we show that continuous DLMs can be made effective with minimal adaptation to the discrete domain. We propose Embedded Language Flows (ELF), a class of diffusion models in continuous embedding space based on continuous-time Flow Matching. Unlike existing DLMs, ELF predominantly stays within the continuous embedding space until the final time step, where it maps to discrete tokens using a shared-weight network. This formulation makes it straightforward to adapt established techniques from image-domain diffusion models, e.g., classifier-free guidance (CFG). Experiments show that ELF substantially outperforms leading discrete and continuous DLMs, achieving better generation quality with fewer sampling steps. These results suggest that ELF offers a promising path toward effective continuous DLMs.

LLM-based Software Engineering agents face a critical bottleneck: context length limitations cause failures on complex, long-horizon tasks. One promising solution is to encode context as continuous embeddings rather than discrete tokens, enabling denser information storage. We apply the recently proposed In-Context Autoencoder for this purpose. While the method performs well on single-shot common-knowledge and code-understanding tasks, our experiments demonstrate that it fails on multi-step agentic coding tasks. In this paper, we explore this phenomenon and discuss possible factors contributing to this failure.

Reasoning benchmarks measure clinical performance on clean inputs. We evaluate the step before reasoning: retrieval over real EHR notes, where negation, temporality, and family-versus-patient attribution can flip a correct answer to a wrong one. EpiKG carries an assertion label and a temporality tag with every fact in a patient knowledge graph, then routes retrieval by question intent. ClinicalBench is a 400-question test over 43 MIMIC-IV patients across 9 assertion-sensitive categories. A 7-condition ablation tests each piece of EpiKG across six LLMs (Claude Opus 4.6, GPT-OSS 20B, MedGemma 27B, Gemma 4 31B, MedGemma 1.5 4B, Qwen 3.5 35B). Three physicians blindly adjudicated 100 paired items. The author-blind primary endpoint, leave-author-out paired exact McNemar on 50 unanimous-strict items rated by two external physicians, yields +22.0 percentage points (95 percent Newcombe CI [+5.1, +31.5], p=0.0192). The architectural novelty, intent-aware KG-RAG over a Contriever dense-RAG baseline (C2b to C4g_kw on the change-excluded n=362 endpoint), is +8.84 percentage points (paired McNemar p=1.79e-3); +12.43 percentage points under oracle intent. Sensitivities agree directionally: three-rater physician majority +24.0 percentage points (subject to single-author circularity); deterministic keyword reproducibility proxy +39.5 percentage points. Across the six models, the gain shrinks as the LLM-alone baseline rises (beta=-1.123, r=-0.921, p=0.009). With n=6 this looks more like regression to the mean than encoding substituting for model size. Physician adjudication identified 56 percent of auto-generated reference answers as defective, a methodological finding indicating that NLP-pipeline clinical-QA benchmarks require physician adjudication to be usable. ClinicalBench, the frozen evaluator, three-rater adjudication data, and the EpiKG output stack are publicly released.

We decompose an evolutionary mixture-of-LoRA system on a from-scratch ~150M-parameter widened-D substrate (D=1536, V=32000; D/V approx 0.048; the "widened-1536" substrate) into three factors – a router rewrite (parallel sigmoid gate with learnable per-adapter floor and bounded temperature anneal, fed post-stack hidden states rather than token-embedding means), a per-domain leave-one-out evaluation scope, and a lifecycle of death plus alpha-blend inheritance plus SVD mutation plus slot reallocation – and report a 5-of-8 partial 2^3 factorial run at n=3 seeds and 25000 adaptation steps per cell. The attribution chain is sharp on this substrate: the router rewrite carries the entire +0.0426 nat balanced log-PPL improvement (Delta = log PPL_ref - log PPL_test, positive = improvement; t=12.86, p=0.006) attributed to "the full evolutionary system vs the static B3 baseline"; the headline full-system-vs-B3 balanced contrast itself is +0.015 nats, t=1.94, p=0.19 at n=3 and does not clear alpha=0.05. The per-domain evaluation scope is null at seed-resolution, and the lifecycle is a net drag of approx -0.028 nats (t=-4.46,p=0.047 in the primary chain). An auxiliary alpha=0 inheritance counterfactual at n=3 seeds is sign-inconsistent at the headline metric and underpowered for either an equivalence or load-bearing conclusion (corrected from an earlier arithmetic-mean aggregator that erroneously cleared inheritance; see Appendix B.11). A base-perturbation probe directionally refutes a "genomic-context" reframe of the lifecycle role. A controllable synthetic sandbox locates a substrate-conditional regime boundary: evolutionary search on the routing channel is load-bearing only when adapters are pre-aligned to the task; in every other regime tested it underperforms, ties, or actively degrades the gradient solution.

Existing multi-model and tool-augmented systems communicate by generating text, serializing every exchange through the output vocabulary. Can two pretrained language models instead coordinate through a continuous, concurrent channel? The Bicameral Model couples two frozen language models through a trainable neural interface on their intermediate hidden states. At every generation step, both models run in lockstep: a primary model drives the task while an auxiliary model operates tools, solves constraints, or executes code, with both conditioning on each other's activations through a translation network and a learned suppression gate ($\sim$1\% of combined parameters). The gate learns a selective communication protocol from task loss alone, without a prescribed format. We demonstrate the mechanism across three tool backends. On arithmetic, coupling two 0.5B models with a calculator raises accuracy from 36\% to 96\%. On logic grid puzzles, coupling two 0.6B models with a Z3 solver achieves $1.7\times$ the unaugmented baseline on ZebraLogic. On mathematical reasoning, coupling with a Python sandbox enables the auxiliary to generate problem-specific code from hidden-state signals alone, without ever seeing the problem text.

Large language models (LLMs) trained on web-scale corpora can memorize sensitive training data, posing significant privacy risks. Differential privacy (DP) has emerged as a principled framework that limits the influence of individual data points during training, yet the relationship between differential privacy and social bias in LLMs remains poorly understood. To investigate this, we present a systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing a DP model against non-DP baselines across four complementary paradigms: sentence scoring, text completion, tabular classification, and question answering. We find that DP reduces bias in sentence scoring tasks, where bias is measured through controlled likelihood comparisons, yet this improvement does not generalize across all tasks. Our results reveal a discrepancy between logit-level bias and output-level bias. Moreover, decreasing memorization does not necessarily reduce unfairness, underscoring the importance of multi-paradigm evaluation when assessing fairness in LLMs.

In this paper, we present the RETUYT-INCO participation at the BEA 2026 shared task "Rubric-based Short Answer Scoring for German". Our team participated in track 1 (Unseen answers three-way), track 3 (Unseen answers two-way) and track 4 (Unseen questions two-way). Since these tracks required scoring short student answers using specific rubrics, we looked for ways to handle the changing nature of the task. We created a method called Meta-prompting. In this approach, an LLM creates a custom prompt based on examples from the Train set. This prompt is then used to grade new student answers. Along with this method, we also describe other approaches we used, such as classic machine learning, fine-tuning open-source LLMs, and different prompting techniques. According to the official results, our team placed 6th out of 8 participants in Track 1 with a QWK of 0.729. In Track 3, we secured 4th place out of 9 with a QWK of 0.674, and we also placed 4th out of 8 in Track 4 with a QWK of 0.49.

We present Hebatron, a Hebrew-specialized open-weight large language model built on the NVIDIA Nemotron-3 sparse Mixture-of-Experts architecture. Training employs a three-phase easy-to-hard curriculum with continuous anti-forgetting anchoring, followed by supervised fine-tuning on 2 million bilingual Hebrew–English samples. The curriculum ordering alone yields a 3-point aggregate benchmark gain over the reversed configuration. Hebatron achieves a Hebrew reasoning average of 73.8\%, outperforming DictaLM-3.0-24B-Thinking (68.9\%) and remaining competitive with Gemma-3-27B-IT on GSM8K-HE and Israeli Trivia, while activating only 3B parameters per forward pass across a 30B-parameter model, delivering approximately 9 times higher inference throughput at native context lengths up to 65,536 tokens. To our knowledge, this is the first language-specific adaptation of the Nemotron-3 architecture for any target language, and the first open-weight Hebrew-specialized MoE model with native long-context support. Model weights are released openly to support further research in Hebrew and Semitic-language NLP.

Autonomous science promises to augment scientific discovery, particularly in complex fields like biomedicine. However, this requires AI systems that can consistently generate novel and diverse solutions to open-ended problems. We evaluate LLMs on the task of open-ended solution generation and quantify their tendency to mode collapse into low-diversity generations. To mitigate this mode collapse, we introduce analogical reasoning (AR) as a new approach to solution generation. AR generates analogies to cross-domain problems based on shared relational structure, then uses those analogies to search for novel solutions. Compared to baselines, AR discovers significantly more diverse generations (improving solution diversity metrics by 90-173%), generates novel solutions over 50% of the time (compared to as little as 1.6% for baselines), and produces high-quality analogies. To validate the real-world feasibility of AR, we implement AR-generated solutions across four biomedical problems, yielding consistent quantitative gains. AR-generated approaches achieve a nearly 13-fold improvement on distributional metrics for perturbation effect prediction, outperform all baselines on AUPRC when predicting cell-cell communication, infer brain region interactions with a high Spearman correlation ($ρ$=0.729) to published methods, and establish state-of-the-art performance on 2 datasets for oligonucleotide property prediction. The novel and diverse solutions produced by AR can be used to augment the search space of existing solution generation methods.

We study language generation in the limit under a global preference ordering on strings, as introduced by Kleinberg and Wei. As is done in previous work, we aim for breadth, but impose an additional requirement of timeliness: higher-ranked strings should be generated earlier. A string is then only credited if it is generated before a deadline, where its deadline is defined by a function that maps a string's rank in the target language to the time by which it must be produced. This is in keeping with a central consideration in machine learning, where inductive bias favors ``simpler'' or ``more plausible'' outputs, all else being equal. We show that timely generation is impossible in a strong sense for eventually consistent generators – the protagonists of most prior related work. Under what is perhaps the mildest natural relaxation of consistency, a hallucination rate that vanishes over time, we show that we can circumvent our impossibility result. In particular, we can achieve optimal density with respect to any superlinear deadline function. We also show this is tight by ruling out timely generation with linear deadlines and vanishing hallucination rate.

Large Language Models (LLMs) are increasingly deployed in multi-turn dialogue settings where preserving conversational context across turns is essential. A standard serving practice concatenates the full dialogue history at every turn, which reliably maintains coherence but incurs substantial cost in latency, memory, and API expenditure, especially when queries are routed to large proprietary models. Existing approaches often struggle to balance the trade-off between response quality and efficiency. We propose a framework that exploits the early turns of a session to estimate a local response manifold and then adapt a smaller surrogate model to this local region for the remainder of the conversation. Concretely, we learn soft prompts that maximize semantic divergence between the large and surrogate small language models' responses to surface least-aligned local directions, stabilize training with anti-degeneration control, and distill the mined cases into localized LoRA fine-tuning so the surrogate runs without prompts at inference. A simple gate enables a one-time switch with rollback on drift. We further provide a theoretical analysis for key components in SOMA. Extensive experiments show the effectiveness of SOMA. The source code is provided at: https://github.com/LabRAI/SOMA.

Web search queries concern place far more often than existing labelling schemes suggest, yet the landscape of geospatial web search queries - what people ask of place, and how often - remains poorly characterised at scale. We apply dense sentence embeddings, a lightweight SetFit classifier, and density-based clustering to the full MS MARCO corpus of 1.01 million real Bing queries without prior filtering for toponyms or spatial keywords, identifying 181,827 geospatial queries (18.0%), nearly threefold the 6.17% labelled as Location in the original annotations. The resulting taxonomy of 88 query categories reveals that geospatial web search is dominated by transactional and practical lookups: costs and prices alone account for 15.3% of geospatial queries, nearly twice the size of the entire physical geography theme. Much of this activity - costs, opening hours, contact details, weather, travel recommendations - falls outside the scope traditional GIS systems and knowledge graphs are built to serve. The categories vary substantially in the kind of answer they admit, from deterministic lookups answerable from spatial databases or knowledge graphs to evaluative or temporally volatile queries that require generative or real-time systems. We discuss implications for hybrid retrieval architectures and for benchmarks of geographic reasoning in large language models. We openly release the labelled dataset, classifier, and taxonomy.

Test-time compute is widely believed to benefit only large reasoning models. We show it also helps small embedding models. Since modern embedding models are distilled from LLM backbones, a frozen encoder should benefit from extra inference compute without retraining. An agentic program-search loop explores 144 candidate programs over a frozen encoder API and produces twelve Pareto-optimal programs that trade extra inference compute for retrieval quality. The search independently rediscovers Rocchio pseudo-relevance feedback, ColBERT-style MaxSim at sentence granularity, reciprocal rank fusion, and the Fisher linear discriminant, all without trainable parameters or external models. Every frontier program improves nDCG@10 over the frozen baseline on all 14 tasks used during program search. Generalization is validated separately: a single fixed program, selected from the discovery frontier before any held-out evaluation, improves nDCG@10 on 61% of model-task pairs across three unseen encoder families and nineteen held-out retrieval tasks, without any per-task selection.

Agent evaluation requires assessing complex multi-step behaviors involving tool use and intermediate reasoning, making it costly and expertise-intensive. A natural question arises: can frontier coding assistants reliably automate this evaluation process? Our study shows that simply prompting coding assistants is insufficient for this task. Without domain-specific evaluation knowledge, frontier coding assistants achieve only a 30% execution success rate and produce over-engineered evaluations averaging 12+ metrics per agent, indicating that strong coding ability does not automatically translate to reliable agent evaluation. We introduce EvalAgent, an AI assistant that automates the end-to-end agent evaluation pipeline. EvalAgent encodes evaluation domain expertise as evaluation skills (procedural instructions, reusable code and templates, and dynamically retrieved API documentation) that compose into a trace-based pipeline producing complete evaluation artifacts including metrics, executable code, and reports. To systematically assess generated evaluations, we introduce a meta-evaluation framework alongside AgentEvalBench, a benchmark comprising 20 agents, each paired with evaluation requirements and test scenarios. We further propose the Eval@1 metric to measure whether generated evaluation code both executes and yields meaningful results on the first run. Our experiments show that EvalAgent produces focused evaluations, improving Eval@1 from 17.5% to 65%, and achieving 79.5% human expert preference over baseline approaches. Further ablation studies show that evaluation skills are critical for handling complex evaluation: removing them causes Eval@1 to drop significantly from 65% to 30%.

We introduce AcuityBench, a benchmark for evaluating whether language models identify the appropriate urgency of care from user medical presentations. Existing health benchmarks emphasize medical question answering, broad health interactions, or narrow workflow-specific triage tasks, but they do not offer a unified evaluation of acuity identification across these settings. AcuityBench addresses this gap by harmonizing five public datasets spanning user conversations, online forum posts, clinical vignettes, and patient portal messages under a shared four-level acuity framework ranging from home monitoring to immediate emergency care. The benchmark contains 914 cases, including 697 consensus cases for standard accuracy evaluation and 217 physician-confirmed ambiguous cases for uncertainty-aware evaluation. It supports two complementary task formats: explicit four-way classification in a QA setting, and free-form conversational responses evaluated with a rubric-based judge anchored to the same framework. Across 12 frontier proprietary and open-weight models, we find substantial variation in clear-case acuity accuracy and error direction. Comparing task formats reveals a systematic tradeoff: conversational responses reduce over-triage but increase under-triage relative to QA, especially in higher-acuity cases. In ambiguous cases, no model closely matches the distribution of physician judgments, and model predictions are more concentrated than expert clinical uncertainty. We also compare expert and model adjudication on a subset of maximally ambiguous cases, using those cases to examine the role of clinical uncertainty in label disagreement. Together, these results position acuity identification as a distinct safety-critical capability and show that AcuityBench enables systematic comparison and stress-testing of how well models guide users to the right level of care in real-world health use.

Selective layer-wise updates are essential for low-cost continued pre-training of Large Language Models (LLMs), yet determining which layers to freeze or train remains an empirical black-box problem due to the lack of interpretable guidance. To address this issue, we propose LayerTracer, an architecture-agnostic diagnostic framework that reveals the evolution patterns of layer-wise representations and stability by locating task execution positions and quantifying layer sensitivity. Analysis results reveal that deep layers act as critical regions for task execution and maintain high stability against disruptive updates. Guided by this finding, we conduct three controlled continued pre-training trials to compare diverse freeze-train strategies, demonstrating that training shallow layers while freezing deep layers consistently outperforms full-parameter fine-tuning and the opposite allocation on both C-Eval and CMMLU benchmarks. We further present a hybrid model case study, which validates that placing high-quality pre-trained modules in deep layers effectively preserves inherent knowledge of the model. This work delivers a low-cost and interpretable solution for resource-constrained teams, offering actionable guidance for layer-wise parameter allocation in continued pre-training and hybrid model construction.

Large Language Model (LLM) agents have emerged as key intermediaries, orchestrating complex interactions between human users and a wide range of digital services and LLM infrastructures. While prior research has extensively examined the security of LLMs and agents in isolation, the systemic risk of the agent acting as a disruptive hub within the user-agent-service chain remains largely overlooked. In this work, we expose a novel threat paradigm by introducing Mobius Injection, a sophisticated attack that weaponizes autonomous agents into zombie nodes to launch what we define as gent-based and -Oriented DDoS (AbO-DDoS) attacks. By exploiting a structural vulnerability in agentic logic named Semantic Closure, an adversary can induce sustained recursive execution of agent components through a single textual injection. We demonstrate that this attack is exceptionally lightweight, stealthy against both traditional DDoS monitors and contemporary AI safety filters, and highly configurable, allowing for surgical targeting of specific environments or model providers. To evaluate the real-world impact, we conduct extensive experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs. Our results demonstrate that Mobius Injection achieves substantial attack success across diverse tasks, driving single-node call amplification up to 51.0x and multi-node p95 latency inflation up to 229.1x. The attack performance exhibits a superlinear increase with the number of poisoning nodes. To mitigate Mobius Injection, we propose a proactive defense mechanism using Agent Component Energy (ACE) Analysis, which detects malicious recursive triggers by measuring anomalous energy in the agent's component graph.

While large language models excel at factual adaptation, their ability to internalize nuanced philosophical frameworks under severe data constraints remains underexplored. We investigate this by specializing small LLMs on micro-datasets of foundational Stoic texts using preference optimization (ORPO, AlphaPO). Evaluated via a multi-model critic bank, our results show that just 300 high-fidelity examples can induce strong alignment with inward-facing Stoic virtues, closely approaching few-shot prompting while freeing the context window. Critically, however, all models, including few-shot baselines, exhibit a persistent failure on Stoicism's outward-facing cosmopolitan duties, pointing to a representational limitation of small models that micro-dataset adaptation alone cannot overcome.

Knowledge Distillation (KD) is a critical tool for training Large Language Models (LLMs), yet the majority of research focuses on approaches that rely solely on output logits, neglecting semantic information in the teacher's intermediate representations. While Hidden Layer Distillation (HLD) showed potential for encoder architectures, its application to decoder-only pre-training at scale remains largely unexplored. Through compute-controlled experiments, we benchmark HLD against logit-based KD and self-supervised baselines with Gemma3 3.4B as teacher and 123M and 735M students trained on up to 168B tokens from the C4 dataset. Our experiments show that HLD does not consistently outperform standard KD on downstream evaluation tasks. Nevertheless, we show that HLD can yield a systematic perplexity gain over KD across all shared-hyperparameter configurations, suggesting that a latent signal can be extracted, but a breakthrough may be needed for it to play a more significant role in LLM pre-training.

Effectively configuring scalable large language model (LLM) experiments, spanning architecture design, hyperparameter tuning, and beyond, is crucial for advancing LLM research, as poor configuration choices can waste substantial computational resources and prevent models from realizing their full potential. Prior automated methods are designed for low-cost settings where repeated trial and error is feasible, but scalable LLM experiments are too expensive for such extensive iteration. To our knowledge, no work has addressed the automation of high-cost LLM experiment configurations, leaving this problem labor-intensive and dependent on expert intuition. Motivated by this gap, we propose AutoLLMResearch, an agentic framework that mimics how human researchers learn generalizable principles from low-fidelity experiments and extrapolate to efficiently identify promising configurations in expensive LLM settings. The core challenge is how to enable an agent to learn, through interaction with a multi-fidelity experimental environment that captures the structure of the LLM configuration landscape. To achieve this, we propose a systematic framework with two key components: 1) LLMConfig-Gym, a multi-fidelity environment encompassing four critical LLM experiment tasks, supported by over one million GPU hours of verifiable experiment outcomes; 2) A structured training pipeline that formulates configuration research as a long-horizon Markov Decision Process and accordingly incentivizes cross-fidelity extrapolation reasoning. Extensive evaluation against diverse strong baselines on held-out experiments demonstrates the effectiveness, generalization, and interpretability of our framework, supporting its potential as a practical and general solution for scalable real-world LLM experiment automation.

Using offline datasets to evaluate conversational agents often fails to cover rare scenarios or to support testing new policies. This has motivated the use of controllable user simulators for targeted, counterfactual evaluation, typically implemented by prompting or fine-tuning large language models. In this work, we formalize controllable simulation as a causal inference problem. By bridging natural language evaluation with off-policy evaluation methodology, we show that the standard practice of training simulators via supervised fine-tuning on post-hoc trajectory labels yields a structurally biased model. Specifically, these labels are inextricably coupled to the data-generating behavior policy, injecting a look-ahead bias that breaks causal consistency. Furthermore, we prove that under policy shift this failure causes the variance of evaluation metrics to explode geometrically, a phenomenon we term controllability collapse. To restore causal consistency, we establish theoretical conditions for accurate simulation and propose practical training mitigations: a priori controls, step-wise dynamic controls, and direct policy-conditioned learning. Empirical evaluation confirms that while standard global controls distort conversational distributions and collapse behavioral diversity, our causally grounded simulators eliminate look-ahead bias, preserve natural variance, and exhibit robust zero-shot generalization to unseen agent behaviors.

Clinical check-up reports are multimodal documents that combine page layouts, tables, numerical biomarkers, abnormality flags, imaging findings, and domain-specific terminology. Such heterogeneous evidence is difficult for laypersons to interpret and translate into concrete follow-up actions. Although large language models show promise in medical summarisation and triage support, their ability to generate safe, prioritised, and patient-oriented actions from multimodal check-up reports remains under-benchmarked. We present \textbf{Checkup2Action}, a multimodal clinical check-up report dataset and benchmark for structured \textit{Action Card} generation. Each card describes one clinically relevant issue and specifies its priority, recommended department, follow-up time window, patient-facing explanation, and questions for clinicians, while avoiding diagnostic or treatment-prescriptive claims. The dataset contains 2,000 de-identified real-world check-up reports covering demographic information, physical examinations, laboratory tests, cardiovascular assessments, and imaging-related evidence. We formulate checkup-to-action generation as a constrained structured generation task and introduce an evaluation protocol covering issue coverage and precision, priority consistency, department and time recommendation accuracy, action complexity, usefulness, readability, and safety compliance. Experiments with general-purpose and medical large language models reveal clear trade-offs between issue coverage, action correctness, conciseness, and safety alignment. Checkup2Action provides a new multimodal benchmark for evaluating patient-oriented reasoning over clinical check-up reports.

Group Relative Policy Optimization (GRPO) has emerged as a promising approach for improving the reasoning capabilities of large language models. However, it struggles to effectively balance the tradeoff between exploration and exploitation during training, often resulting in suboptimal performance. Motivated by the theoretical insight that changes in entropy are governed by the covariance between token probabilities and their corresponding advantages, we propose a hyperparameter-free, covariance-weighted optimization method that dynamically down-weights extreme token-level updates via a Gaussian kernel. This approach automatically reduces the instability caused by exploration-exploitation trade-off while preserving informative learning signals. Extensive empirical evaluations show that our approach improves downstream performance across reasoning benchmarks compared with GRPO, and effectively stablizes entropy as training progresses.

Large language models (LLMs) have shown remarkable potential in advertising scenarios such as ad creative generation and targeted advertising. However, deploying LLMs in real-time advertising systems poses significant challenges due to their high inference latency and computational cost. In this paper, we propose an Efficient Generative Targeting framework that integrates adaptive group quantization, layer-adaptive hierarchical sparsification, and prefix-tree parallel verification to accelerate LLM inference while preserving generation quality. Extensive experiments on two real-world advertising scenarios demonstrate that our framework achieves significant speedup with acceptable quality degradation, making it operationally viable for practical deployments.

Autoregressive language models are widely used for text evaluation, however, their left-to-right factorization introduces positional bias, i.e., early tokens are scored with only leftward context, conflating architectural asymmetry with true text quality. We propose masked reconstruction as an alternative paradigm, where every token is scored using full bidirectional context. We introduce DiffScore, an evaluation framework built on Masked Large Diffusion Language Models. By measuring text recoverability across continuous masking rates, DiffScore eliminates positional bias and naturally establishes an evaluation hierarchy from local fluency to global coherence. We further provide diagnostic tools unavailable to autoregressive frameworks: multi-timestep quality profiles that decompose scores across masking rates, and bidirectional PMI decomposition that disentangles fluency from faithfulness. Experiments across ten benchmarks show that DiffScore consistently outperforms autoregressive baselines in both zero-shot and fine-tuned settings. The code is released at: https://github.com/wenlai-lavine/DiffScore.

Comparing post-training LLM variants, such as quantized, LoRA-adapted, and distilled models, requires a diagnostic that identifies how a variant has drifted, not only whether it has degraded. Existing similarity scores such as CKA and SVCCA can flag degradation, but they do not directly link representation drift to risk or mechanism. We propose PRISM, Proxy Risk Inference via Structural Mapping, which exploits the linear output head of LLMs and the empirically near-isometric structure of their backbones to derive a closed-form upper bound on the cross-entropy risk gap between a target model and a post-training variant. The bound is calibrated for variant ranking and decomposes drift into three independently measurable axes: scale mismatch, shape mismatch, and head divergence. Each axis corresponds to a distinct failure mode, including shape distortion under low-bit quantization, scale separability under LoRA forgetting, and head divergence under GGUF k-quantization. As a result, the dominant axis suggests a remediation direction rather than merely raising a degradation flag. Because the shape term is differentiable, the same geometry can also serve as a training-time regularizer against catastrophic forgetting. Across two model families and five benchmarks, PRISM ranks variants with mean Spearman correlations of 0.820 for post-training quantization and 0.831 for LoRA forgetting, and its axis-guided shape regularizer outperforms experience replay in aggregate at mitigating downstream forgetting.

Backdoor vulnerabilities widely exist in the fine-tuning of large language models(LLMs). Most backdoor poisoning methods operate mainly at the token level and lack deeper semantic manipulation, which limits stealthiness. In addition, Prior attacks rely on a single fixed trigger to induce harmful outputs. Such static triggers are easy to detect, and clean fine-tuning can weaken the trigger-target association. Through causal validation, we observe that emotion is not directly linked to individual words, but functions as an overall stylistic factor through tone. In the representation space of LLM, emotion can be decoupled from semantics, forming distinct cluster from the original neutral text. Therefore, we consider the emotional factor as the backdoor trigger to propose a pparasitic emotion-style dynamic backdoor attack, Paraesthesia. By mixing samples with the emotional trigger into clean data and then fine-tuning the model, the model is able to generate the predefined attack response when encountering emotional inputs during the inference stage. Paraesthesia includes two the quantification and rewriting of emotional styles. We evaluate the effectiveness of our method on instruction-following generation and classification tasks. The experimental results show that Paraesthesia achieves an attack success rate of around 99\% across both task types and four different models, while maintaining the clean utility of the models.

Recent multimodal large language models (MLLMs) have shown strong chain-of-thought (CoT) reasoning ability on vision-language tasks, but their direct deployment in real-world systems is often limited by latency and resource constraints. In practice, smaller MLLMs are preferred for online serving, yet their reasoning performance is bottlenecked by the lack of large-scale, high-quality multimodal CoT supervision. In this paper, we present OmniThoughtVis, a scalable data curation and distillation pipeline for transferring multimodal reasoning capabilities from high-capacity teacher models to smaller, deployment-oriented MLLMs. Starting from a diverse open-source seed pool, our pipeline generates structured CoT traces and performs joint annotation of reasoning difficulty, answer quality, and semantic task tags. To maintain data quality at scale, we combine rule-based filtering, difficulty-aware selection, and tag-based diversity sampling, resulting in a curated corpus of 1.8M samples that supports controllable subset construction for downstream training. We use OmniThoughtVis to distill Qwen3-VL models from 2B to 8B parameters and evaluate them on nine multimodal reasoning benchmarks. The resulting distilled models show consistent gains across model scales, including improvements of up to +16.8 points on MathVerse and +5.6 points on MMMU-Pro for the 4B model. Notably, the distilled 4B model matches or surpasses the undistilled 8B baseline on several tasks, highlighting the practical value of scalable reasoning distillation for deployment-oriented MLLMs.