New Computation and Language (cs.CL) papers, Monday 19 October 2026
146 papers announced on arXiv. Get them ranked for you, free →
Revealing the Impact of Visual Text Style on Attribute-based Descriptions Produced by Large Visual Language Models
When the visual style of text is considered, a wide variety can be observed in font, color, and size. However, when a word is read, its meaning is independent of the style in which it has been written or rendered. In this paper, we investigate whether, and how, the style in…
Binding-Specific Relearning Savings for Test-Time Continual Learning
AI summaryThe study shows that learning a specific factual binding in a language model yields a measurable relearning advantage at test time, speeding recovery by ~18% over controls and reducing updates needed, indicating current accuracy metrics underestimate lasting experience effects.
Eliciting Specifications: When Does Interactive Clarification Help?
AI summaryThe paper tests whether an elicitation agent that asks up to 10 clarification questions improves the accuracy of Lean formalizations produced by LLMs. Across 189 tasks, cross‑model pipelines (e.g., Sonnet 4.6→GPT‑5.6) achieve a strict pass@1 of 65.6%, a 12‑task gain, while same‑model setups see no improvement.
PRIME: Towards Verifiable Evidence-Grounded Reasoning in LLM Agents
AI summaryPRIME pre‑commits inference rules to specific evidence routes and logs all checks, refusing any conclusion unsupported by logged evidence and guiding agents to stronger conclusions. In 480 tasks, PRIME eliminated 26.7% unjustified outputs and raised justified‑score to 59.2/100.
Embeddings Are Overpriced? A Frozen 16 MB Vocabulary for Small and Ternary Language Models
AI summaryFrozen 16 MB embedding tables replace trainable vocabularies in small LMs, matching or exceeding performance while cutting memory and speed by 5–6×, and handling rare words and new vocab at minimal extra cost.
Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss
AI summaryReducing the number of routed experts during prompt prefill mainly hurts short‑answer QA because early prompt tokens become sensitive; at half the experts, F1 drops by roughly one‑third of the loss when only the first few prompt tokens are routed. Restoring full‑routing only for those few tokens recovers most of the F1 loss.
High prevalence of antagonizing rhetoric in trending topics on Twitter/X
Disagreement between political opponents is a central element of democracy. This disagreement can be voiced in different ways. An opponent can be antagonized, i.e. portrayed as an enemy that needs to be overcome, or they can be criticized, i.e. treated as a legitimate adversary…
Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation
AI summaryThis paper demonstrates that recursively training models on their own safe generations using benign auxiliary cues causes progressive misalignment, leading to increased vulnerability to harmful queries across multiple model families.
From Parallels to Divergence: Language Models' Learning of Action Nominal Constructions
AI summaryInvestigates GPT‑2‑small learning of action nominal constructions across 96 synthetic languages; finds models mirror natural clause word order but diverge on complement systems and ANC strategies, often generating rare or unattested combinations.
RealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents
AI summaryThe authors introduce RealCLI, a benchmark evaluating both task execution and human-agent collaboration in command-line interface agents using a multi-dimensional framework and paired comparison modes. Evaluations show that frontier agents struggle significantly with collaboration, achieving a peak joint score of only 37.6%.
AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems
AI summaryAuditKIT is an evaluation library that standardizes benchmarking, judging, and tracing for large language models and agents into a single result representation. It enables robust comparisons across diverse workflows by providing unified metrics for tool use, retrieval, and runtime performance.
alkhder at ImageEval 2026 Shared Tasks: Constrained Selection and Rotation-Averaged Logit Scoring for Multimodal Hallucination Detection
AI summaryThe paper introduces a zero-shot vision-language system for multimodal hallucination detection using constrained log-prob selection and Latin-square rotation. It reduces ranking error relative to baseline models by over 80 percent on both English and Arabic test sets.
Learning User Representations to Improve Simulation Fidelity
AI summaryThe paper proposes a conversation‑reconstruction framework to evaluate and improve user representations for simulators, introducing hindsight distillation that trains a generator (UserRepGen‑4B) to produce better user descriptors. The method yields higher simulation fidelity and preserves real‑user diversity compared to GPT‑5.5.
alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B
AI summaryFine‑tuned NileChat‑3B with QLoRA on Alexandria dialogue data, using a single adapter set for all 13 dialects. Achieves 20.36 country‑macro spBLEU and 36.64 chrF++ on the private test set, training in ~1 hour.
Literal Lock-In: How Test Language Shifts Literal Attraction in Typed Decision Models on Bengali Idioms
Decision models, non-generative scorers that report a probability over caller-supplied options and emit no text, are increasingly deployed as routers, classifiers and guardrails. Because they generate nothing, rationales and chain-of-thought do not apply. We audit three: an…
Pearl-Chat: Engineering a native Luganda language model from first principles in JAX and Flax NNX
Autoregressive foundation language models predominantly target high-resource Indo-European languages, imposing computational and morphological penalties on under-represented African languages. Generic multilingual tokenizers routinely exceed fertility rates of 3.2 tokens per…
Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks
AI summaryLarge reasoning models that use test‑time computation perform poorly on everyday logical reasoning tasks that diverge from formal logic. Chain‑of‑thought prompting on non‑reasoning models even outperforms comparable reasoning models, showing that current reasoning models lack reliable everyday reasoning.
HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments
AI summaryHeRE builds a 1.5‑million‑instance Hebrew relation‑extraction benchmark by denoising noisy knowledge‑graph alignments with LLM‑guided NLI. Gemma‑4‑31B‑it scores F1 = 0.915, matching GPT‑5.4, and the dataset enables scalable, language‑agnostic RE evaluation.
Cultural (In)Competence: Investigating Culture-Specific Stereotype Detection Capabilities of LLMs
Stereotypes are subjective constructs that are mediated by the cultural context in which they are viewed. Successfully modeling them thus also requires cultural competence. When using large language models (LLMs) for stereotype-related tasks, such as their detection, it is…
Who Do Perspectivist Models Predict Well? Annotator Agreement, Sparsity, and Performance
AI summaryPerspectivist NLP models are evaluated across diverse annotators, revealing that they predict better for annotators who agree with peers, while the number of annotations per annotator shows no significant effect on performance.
Repairing LLM-Generated Specifications with Deterministic Diagnosis
AI summaryDRAGO replaces ungrounded LLM critiques with deterministic, formalism‑specific diagnosis, rewrites only when error weight strictly decreases, reducing regressions to 1.3% and boosting accuracy by 12.9% over zero‑shot formalization and 8.7% over LLM‑diagnostic repair across five benchmarks.
Mitigating Demographic Disparities in Explanation-Human Alignment via Rationale Regularization
AI summaryThe paper introduces Plausibility‑EFG, a metric for measuring explanation fairness in hate‑speech detection, and shows that adversarially trained RoBERTa (ADV‑RoBERTa) improves both classification and explanation alignment while reducing demographic disparities in token attributions.
SAG2: Embedding-Free Multi-Hop Retrieval with BM25 and JEV
Strong multi-hop retrieval should be practical to deploy without expensive corpus preparation or complex database infrastructure. We present SAG2, an embedding-free retrieval architecture built around a single SQLite database, BM25, and contextual evidence decisions from JEV.…
The Model Knows Better: Defending Medical Multi-Agent Systems Against Prompt Injection
Medical multi-agent systems, in which LLM agents discuss a case, are increasingly studied for clinical decision support. Prompt injection, text that an attacker adds to a model's input, can manipulate LLMs into wrong decisions. How it affects these systems, and how to protect…
Broad and Efficient PII Detection with Unified Labels
Personally Identifiable Information (PII) detection models must efficiently find identifiers across diverse contexts while maintaining a low false-positive rate (FPR) on high-volume traffic. Recent detectors increasingly rely on complex architectures that raise deployment cost,…
Can LLMs predict the outcomes of machine learning experiments?
AI summaryLarge language models can predict the performance of their own proposed ML experiments better than chance but explain only ~50 % of outcome variance; they systematically underestimate effect sizes, show task‑dependent accuracy, and are largely indifferent to the coding harness used.
When Does Uncertainty Measure RAG Hallucination Risk?
Uncertainty can measure hallucination risk in retrieval-augmented generation (RAG), but only when retrieved evidence constrains the answer. When evidence is present, unsupported factual claims become less likely; when retrieval coverage is poor, the model can answer from…
Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification
Large language models rely on spurious features in clinical classification: swapping a single gender token in clinical notes alters gender-irrelevant diagnoses in up to 53.5% of cases. Existing inference-time corrections either need counterfactual data built around the specific…
A Plausible Explanation Is Enough to Mislead Multimodal Judges
AI summaryMultimodal LLM judges are easily misled by plausible explanations even when answers are fixed. Across 14 judges, ordinary or chain‑of‑thought justifications raise acceptance rates by ~13 percentage points, showing a “plausibility capture” bias that can inflate evaluation scores.
ToMStories: Interpreting Theory-of-Mind in Synthetic Narratives
AI summaryThe paper introduces ToMStories, a 6k synthetic narrative dataset for testing language models’ inference of characters’ psychological states (emotion, fear, motivation, wound). Models up to 72B parameters match humans on multiple‑choice tasks but fall short on open‑ended questions, revealing simplifying assumptions in model interpretation.
Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness
AI summaryMeasuring mechanistic uncertainty in LLMs by defining the causal robustness margin—the smallest internal intervention that changes a model’s output—using Jacobian Lens vectors. Experiments show incorrect or uncertain answers have smaller margins, while correctly learned facts gain larger margins, complementing traditional uncertainty measures.
Inductive Learning of Concept Embeddings from Cross-Linguistic Colexification Patterns with Textual Attributes
AI summaryInductive framework learns concept embeddings from colexification networks and short definitions, enabling unseen concepts to be embedded; resulting vectors correlate with cross‑linguistic similarity measures and generalize to new concepts.
SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors
Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing approaches target affirmative response prefixes through rank-one updates to a small subset of model weights. Yet when this…
Evidence-Gated Bilingual RAG for Reliable Consumer Banking Search
Consumers often describe what they need rather than use the terminology found on a banking website. For example, a natural-language request ("a card for travel") may not share the keywords of the relevant product page, and the problem is harder in bilingual services, where the…
Secondhand Stress: How Stressful Disclosures Affect Psychosocial Wellbeing in Online Graduate Admissions Communities
Online communities provide spaces for individuals navigating stressful life transitions to disclose difficulties and seek peer support. Yet encountering others' distress may also have psychosocial implications for community members. Drawing on stress crossover theory, we examine…
Exploring The Empirical Privacy Landscape of Synthetic Data
Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information. In this paper, we present a controlled comparison of data-level and training-level interventions for navigating the…
ProPerAgent: Toward Memory-Driven Proactive Assistance
As assistants move from reactive answering toward proactive intervention, memory changes role: it no longer only supplies the answer to a query, it determines whether and when to speak at all. Prior work in this setting hands the system the relevant fact in context; we study the…
Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation
AI summaryThe paper introduces TAG-Nav, a framework that equips mobile GUI agents with persistent graph-based destination memory built via offline exploration. Evaluated on a new navigation benchmark, DestBench, graph memory significantly improves strict navigation success rates across multiple agent backbones.
Multimodal Filesystem Memory For Long-Horizon Embodied Agents
Agents operating over long horizons need an effective memory that allows them to reuse their experience throughout long interactions. If the entire history of an interaction is retained, it becomes difficult to find the relevant experience, while compressing it into summaries…
Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries
AI summaryThis paper introduces BRIA, an agent that isolates candidate tool investigations into separate execution branches before selectively adopting successful results. BRIA significantly outperforms strong baselines on large tool library benchmarks while maintaining efficient token costs.
GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations
AI summaryGlobeBench provides 1,632 text‑based world‑model tasks across five domains, evaluating LLMs on predicting environment responses to actions. GPT‑6 Astra tops the set with 49.57% success, while most tasks remain unsolved, highlighting a large gap in general world‑model ability.
First Passage Attention
Augmenting softmax attention with forget gates typically relies on summing decay contributions along the key--query path. Because such accumulation is monotonic, these mechanisms cannot selectively erase intermediate context while preserving older, structurally relevant tokens.…
Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation
Refusal is a common safety mechanism in large language models (LLMs), yet in contexts of suicidal ideation (SI), refusing to engage can itself constitute a safety failure. We evaluate how eight instruction-tuned LLMs respond to SI and self-harm disclosures across three datasets,…
ARBOR: Locating Reasoning Failures in Long Chains of Thought through Reasoning State Tracking
AI summaryARBOR tracks long‑chain reasoning as a tree of states, attributing failures to specific step‑state relations. It refines overthinking, underthinking, and unsupported deductions into distinct subtypes and separates failure steps from at‑risk steps, enabling better mitigation.
CoLU: Confidence-Weighted Layerwise Unlearning for Large Language Models
AI summaryCoLU applies confidence‑weighted forgetting and retain regularization to a single transformer layer, and uses Brief Layer Screening with a normalized forgetting‑retention score to select layers, achieving 60–80% faster runtime and superior forgetting and retention on TOFU, MUSE Books, and News while saving 77–80% GPU memory.
A misinformation playbook sharpens judgement without breeding cynicism
Misinformation harms public health and democracy, and generative artificial intelligence now writes falsehoods as persuasive as people's. Defences range from fact-checking individual claims to training people to resist falsehoods. Yet many claims are too new to check, and…
TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents
Advances in frontier coding agents have sharply increased the token costs of AI-assisted software engineering. Their token efficiency is typically evaluated on isolated tasks, whereas real-world developers use them to continuously build, update, and maintain large-scale…
Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents
This study examines the transferability of a de-identification model developed for Korean court judgments to a broader range of sensitive documents, focusing on boundary precision, label granularity, and recall. The agglutinative structure of Korean causes Byte Pair Encoding…
Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns
AI summaryThe authors introduce NoLBERT, a financial news encoder trained strictly on historical text to prevent hindsight bias. They find that while NoLBERT yields stable topics, text-based models ultimately fail to outperform price-only benchmarks in out-of-sample stock return forecasting.
Self-Evolving Coding Agents via Harness-Policy Optimization
AI summaryThe paper introduces Self-HPO, a recursive framework that jointly optimizes the policy and environment harness for LLM coding agents using reinforcement learning. This co-evolutionary approach significantly improves resolve rates across multiple coding benchmarks compared to single-component optimization baselines.
Prune the Leaky Neurons: Mitigating Private Data Leakage in Fine-Tuned LLMs via Joint-Embedding-Guided Pruning
AI summaryThis paper introduces JEPA-Prune, a post-hoc neuron pruning method that mitigates private data leakage in fine-tuned LLMs using a PII-free reference model. Pruning a small fraction of identified leaky neurons drastically reduces privacy leakage while better preserving model utility compared to existing baselines.
Causal Latent Workspaces for Looped Transformers
AI summaryIntroduces a causal latent workspace Transformer that shares a block across depth, letting positions write and retrieve gated prefix summaries through a four‑round loop. It improves perplexity by ~1% over baselines on multiple byte‑level corpora while matching parameter budgets.
Reasoning Reranking with Semantic IDs as the Sole Item Representation
AI summaryThis paper introduces ReRankID to evaluate reasoning rerankers using only compact Semantic IDs instead of text. Results show that targeted post-training enables SID-only models to match text-based LLMs with much shorter prompts, despite producing less faithful item descriptions.
Veil-Then-Reveal: Protecting PII Privacy via Reversible Pseudo-Anonymization in Zero-Trust User--LLM Interaction
Large language models (LLMs) increasingly process sensitive documents and personal information, yet sending prompts containing personally identifiable information (PII) to remote providers exposes data subjects and organizations to risks from logging, retention, model training,…
On the Position Bias of Large Language Models: A Comprehensive Study
Recent studies have shown that position bias remains a critical vulnerability in large language models (LLMs), undermining both the integrity of multiple-choice evaluations and the reliability of retrieval-augmented generation. However, existing understanding of this phenomenon…
Watermarkability: Measuring and Improving Language-Model Compatibility with Logit-Based Watermarks
Language-model watermarks can help identify model-generated text, encode provenance information, and detect post-generation edits. A broad class of logit-based watermarks increases the logits of a designated token set at each decoding step. Because a base model's native…
Causal Edit Serialization
AI summaryCausal Edit Serialization lets language models express revisions (insert, delete, replace, move) in a single causal stream using pointer tokens with source addresses. CES models match autoregressive quality while performing complete edits and improving downstream prediction.
Too Emotional to Be Fake? Intra-Utterance Emotion Shifts Expose Failure Modes in Speech Deepfake Detectors
Expressive speech synthesis can now change emotion within a single utterance, yet its impact on deepfake detection remains unclear. We introduce EmoShift, a benchmark comparing emotion-stable (AA) and emotion-changing (AB) speech to examine these effects while accounting for…
The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls
AI summaryThe paper studies how Chain-of-Thought faithfulness varies with language resource level and task type. It finds that in low‑resource languages, models rely more on reasoning for knowledge‑intensive questions, supporting the Reliance Hypothesis.
Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search
AI summaryUncertainty-Aware Allocation (UAA) uses sibling action dispersion to estimate downstream redundancy and sets per‑step budgets accordingly, decoupling value and uncertainty. It reduces token usage by up to 90 % and latency by 9× on reasoning and planning tasks while maintaining accuracy.
Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data
LLMs over external data face a per-instruction tension: a single document or tool return often interleaves legitimate directives with adversarial prompt injections, and the model must follow the former while rejecting the latter. Existing defenses suppress or trust all external…
Cooperation Without Adaptation: LLM Agents in a Collective-Risk Dilemma
AI summaryLLM agents largely keep a fixed contribution pattern in a collective‑risk dilemma, rarely adapting to changing risk or partner behaviour. Even when they calculate the payoff‑maximising action, their actual contributions stay almost constant, limiting group success.
Batched Contextual Reinforcement II: Does RLVR Need Long Rollout Windows?
AI summaryBatched Contextual Reinforcement explores concurrent verifiable problems to avoid long rollout windows in RLVR. CCR trains models with a shared budget, cutting training tokens by ~80% while maintaining accuracy, enabling efficient short‑window reinforcement learning.
NLData2Opt: Benchmarking Parameter Derivation and Prediction for Modeling Optimization Problems from Natural Language and Data
Formulating optimization problems from raw inputs of problem descriptions and data is time-consuming and costly. While Large Language Models (LLMs) offer a promising way to automate this process, existing benchmarks overestimate their capabilities by relying on idealized inputs…
Bengali Hallucination Detection with Retrieval and Programs: A System Study
Bengali hallucination detection must distinguish unsupported factual claims from incorrect reasoning, often without a supplied reference passage. We present and characterize an inference-only system that combines response-conditioned query rewriting, hybrid retrieval, and…
Competence-Based Evaluation of Logical Reasoning in Large Language Models
AI summaryThe paper proposes a competence‑based evaluation for logical reasoning in LLMs, using parameterized generators and answer‑preserving variations to test generalization. It finds that even strong GPT and Qwen models fail to consistently apply the same reasoning competence across problem families.
FederatedBytes: Tokenizer-Free, Byte-Level Small Language Models for Efficient Multimodal Federated Learning
AI summaryFederatedBytes trains a byte‑level language model with a linear attention mechanism for federated learning of NLP and multimodal tasks. It achieves higher accuracy than ByT5 and LSTM on multilingual sentiment and ScienceQA datasets while cutting memory and communication costs.
Where's the Indirect Evidence? Investigating Grammatical Generalization with Influence Functions
AI summaryUsing influence functions, the paper identifies 60,000 “indirect evidence” sentences that significantly affect grammatical generalization in language models, showing they improve held‑out accuracy by about 12 points more than random subsets, and revealing linguistic category structure.
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning
AI summarySep‑Latent compresses implicit reasoning by using separator tokens as semantic anchors and an attention‑decay scheme to adaptively terminate latent thinking. On GSM8k‑Aug it reaches 56.8 % accuracy with only 5.4 tokens, closing the gap to explicit CoT while vastly reducing overhead.
ASR-Style Transcript Normalization Removes Useful Pragmatic Signal for Dialogue Act Classification
Automatic speech recognition (ASR) pipelines often normalize transcripts by removing filled pauses and punctuation and expanding colloquial forms, potentially discarding cues useful for dialogue act classification. We evaluate nine transcript conditions across six classifiers…
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
AI summaryThe paper identifies a shared cue‑routing circuit in transformer attention heads that drives multiple-choice biases. Post‑training fine‑tuning and RL strengthen these heads, suggesting that mitigation can target this specific circuit.
Beyond Surface Cues: Robust LLM-Generated Korean Text Detection
As large language models (LLMs) generate increasingly fluent text, the need for reliable detection of LLM-generated content has grown substantially. However, existing research remains heavily English-centric, leaving Korean largely underexplored. Moreover, current evaluations of…
Aligning Language Models Towards Counterfactual Invariance
AI summaryThe paper shows that reinforcement learning alignment increases counterfactual bias in language models. It proposes a policy‑gradient method with group‑level rewards on matched counterfactual prompts, reducing demographic differences while maintaining overall response quality.
Recognizing Atari and Assessing Sente: Controlled Evaluation Tasks for Language-Based Go Agents
Domain terminology provides concise descriptions of game states and strategic relationships, but whether game-playing language models can accurately recognize these concepts and relate them to move evaluation remains unclear. We construct a controlled evaluation set centered on…
SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval
Large language model (LLM) agents increasingly rely on persistent memory to support long-term reasoning and personalized interactions. However, information collected from external sources can introduce security risks when untrusted claims (i.e., atomic statements) are later…
What Shapes Brain Alignment of Large Language Models during Dialogue? Insights from Language Production and Comprehension
AI summaryThe study compares fMRI-based brain alignment of LLMs during dialogue, finding distinct patterns for production versus comprehension that correlate with the models’ intrinsic dimension in relevant brain areas, and shows that linguistic and communicative feature contributions differ by processing mode, layer, and region.
KiranaBench: Can a Language Model Run a Neighbourhood Shop?
Large language model (LLM) agents are increasingly evaluated on long-horizon tasks, including autonomously running a business. Existing benchmarks for this setting are almost exclusively English-language, Western in retail context, and single-agent, with the agent observing the…
Self-Evolving Agents Can Repair Failures but Struggle to Generalize from Them
Experience-driven self-evolution offers a promising path for large language model (LLM) agents to continually improve through interaction. Yet existing approaches often rely on successful trajectories or external evaluation signals, leaving unclear whether agents can evolve when…
Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference
AI summaryPrism Attention introduces a coarse-to-fine sparse attention framework that refines shared cross-layer KV indices to track layer-specific token shifts. It achieves state-of-the-art long-context performance while delivering a 3.4 times end-to-end decoding speedup over FlashAttention at a 128K context length.
Controlling the Diversity-Quality Trade-Off of LLMs by Merging Models Before and After Alignment
AI summaryThis paper introduces a training-free layer-wise model merging technique that combines aligned and base LLMs to control the diversity-quality trade-off. By mixing lower and higher layers, the merged models achieve superior semantic and lexical diversity compared to standard alignment without sacrificing quality.
ECHO: Replacing the Compliance Proxy with Response Harmfulness in Optimization-Based Jailbreak Attack
Despite extensive safety alignment, large language models remain vulnerable to jailbreak attacks. Existing optimization-based attacks share a single objective: maximize a token-level proxy of victim compliance---such as the likelihood of an affirmative prefix---computed on the…
Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses
AI financial advisers set a client's stock allocation largely from one stated-risk question, yet people do not answer it the same way twice: in a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two…
ParetoChat: An LLM-Based Interface for Interactive Multi-Objective Optimization
Multi-objective optimization models are often deployed in settings where decision makers (DMs) must balance some conflicting criteria. In practice, however, stakeholders must translate their qualitative preferences between different criteria into numerical weights, priorities,…
Before Retrieval Begins: Safety and Tool-Use Decisions in LLM Agents
Large language models increasingly act as agents that call external tools, yet most tool-safety research examines only what follows a tool call. The prior decision of whether to call the tool remains largely unexamined and can influence downstream safety. Across the seven…
Disentangling Language Effects on Tool Admission and Post-Retrieval Safety
LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric. It remains unclear whether language affects tool admission, post-retrieval generation, or both. We independently vary request, retrieved-page, and required response…
LED: Improving Masked Diffusion Language Models by Lookahead Editing
AI summaryLED adds a rewritable proposal state to masked diffusion, allowing iterative lookahead updates before committing tokens, improving expressivity while keeping simple masks. Fine‑tuned Qwen models achieve +11.4/+7.7 points over standard MDM SFT and beat LLaDA2.1‑16B on benchmarks.
Can AI interview? Comparing AI-led to Human-led Chat Interviews
AI summaryAI chatbots conducted 7‑question interviews twice as often as human interviewers, producing 65 % more text with similar relevance and specificity. The study shows AI‑led interviews can scale qualitative data collection while maintaining data quality.
Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models
Large language models (LLMs) are increasingly used as computational models of human behavior. Recent models such as Centaur have advanced this goal by fine-tuning LLMs on trial-by-trial decisions from psychological experiments. However, existing large-scale behavioral datasets…
Explaining Chess and Go in Natural Language
AI summaryThe paper introduces BinProp, a training‑free method that lets language models learn reusable chess concepts to generate natural‑language move explanations. BinProp nearly triples explainability scores over direct prompting and generalizes to Go, draughts, and shogi.
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
Many language-based applications require probabilities over a finite set of decisions rather than open-ended responses. We present Chinese Decision, a family of Chinese-language models for single-pass probabilistic decision-making. Given context, an instruction, and candidate…
Do Speech BCIs Need Larger Models? Rethinking Neural Decoding beyond Scaling
Recent speech brain-computer interfaces (BCIs) increasingly rely on large models and complex multi-stage pipelines for neural decoding. However, in speech BCI, where training data are scarce, noisy, and expensive to collect, the benefits of model scaling remain unclear. This…
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
AI summaryTreeOPSD is a critic‑free, tree‑aware posterior self‑distillation method that back‑ups terminal rewards through alternative continuations, producing a bounded teacher distribution. It reduces policy drift and sparsity, boosting long‑horizon agentic RL performance (e.g., +4.97 on Deep Research).
What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space
Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands…
AMEND: Constraint-Guided Multi-Agent Text Detoxification with Isolated Critics
Text detoxification requires balancing toxicity removal, semantic preservation, and fluency. Existing methods typically optimize an unconstrained scalar objective---enabling safety scores to mask semantic loss---or rely on open-ended iterative debate prone to semantic drift and…
MAS-Bai: Linguistically Informed Adaptive Multi-Agent Translation for Bowen-Script Bai
Large Language Models (LLMs) remain unreliable for extremely low-resource languages whose writing systems conflict with high-resource pretraining priors. We study Bai-Chinese translation in the Bowen script, where Chinese characters are frequently borrowed phonetically rather…
ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling
Drifting Models generate a sample with a single function evaluation, making them attractive for fast parallel language generation, but their one-step formulation does not support iterative refinement. We propose Recurrent Language Generation via Drifting (ReLaDrift), which uses…
MultiAgent-OPD: End-to-End Training of Hierarchical Multi-Agent Systems with On-Policy Distillation
In a hierarchical multi-agent system, a Leader assigns tasks and subagents carry them out. Jointly training the Leader and the subagents is difficult because a final answer gives only indirect feedback to the subagents. An LLM judge can score the performance of the subagents,…
MGE-PO: Entropy-Aware Credit Assignment for Mathematical Reasoning
AI summaryMGE‑PO introduces a four‑quadrant advantage shaping that rewards correct, low‑entropy responses more and penalizes incorrect, high‑entropy ones, using an entropy‑based curriculum. It boosts macro‑avg@32 by up to 11 % and macro‑pass@32 by 4‑11 % on Qwen models.
Think More or Read More? A Token-Budget Theory of Reasoning with Shared Context
AI summaryThe paper formalizes a token‑budget trade‑off for reasoning systems that share a fixed context window: each token spent reading others reduces the reasoning budget. It provides a closed‑form prediction for the optimal group size in multi‑agent debate, which matches empirical results and shows self‑refinement performs similarly, especially under tight token limits.
Who Pays for Efficient Inference? Confidence-Based Small-to-Large LLM Routing Across Tasks and Languages
Small-to-large language model routing can reduce inference cost by escalating difficult inputs while retaining easier ones on a smaller model. Its success, however, depends on identifying examples that benefit from escalation, and the reliability of confidence across…
Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents
Tool-using agents read untrusted external content, which exposes them to indirect prompt injection (IPI), where hidden instructions hijack the agent's tool calls. Benchmarks measure IPI security by the attack success rate (ASR), which treats every failed attack as a successful…
Sample, Mix, or Verify? Allocating Test-Time Compute under Imperfect Verifiers
AI summaryThe authors introduce SPLIT, a method that uses a small pilot to optimally allocate test-time compute across sampling, mixing models, and verification. SPLIT achieves near-optimal performance across diverse benchmarks and imperfect verifiers, outperforming heuristic budgeting rules.
Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To
AI summaryReasoning models struggle to follow chain-of-thought constraints when given in user prompts, but achieve significantly higher controllability when the constraints are placed as self-instructions inside their reasoning process.
Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation. We introduce VERITAS to verbalise the model's computation: how different…
Annotrace: A Workbench for Step-Level Annotation of Agent Traces
AI summaryAnnotrace converts multi-format agent traces into nested step lists and lets users define a codebook; a frontier AI proposes labels and then humans refine them. In a multi‑agent QA dataset, it exactly reproduced 235 manual annotations and achieved 91% agreement with κ = 0.82.
SimOPD: Relational Knowledge Transfer for On-Policy Distillation
AI summarySimOPD transfers teacher internal knowledge for on-policy distillation by matching response similarity distributions stored in paired memory banks. This approach avoids cross-model projections, retaining student representational freedom and improving performance by up to 3.1 percentage points.
Inferring L1 Backgrounds from L2 English with LLM Token Probabilities
Native language identification (NLI) infers a non-native speaker's first language (L1) from their second language (L2) production. We investigate a token probability approach that scores each text sample as a continuation of a prompt prefix specifying a candidate L1, then…
Self-Reflective Chain-of-Thought (SRCoT): Evidence-Based Text Generation via a Closed-Loop Retrieve--Reflect--Revise Pipeline on Edge LLMs
AI summarySRCoT implements a closed Retrieve--Reflect--Revise loop on edge LLMs that separates knowledge stores and uses BM25 retrieval to produce source‑labeled drafts, gap markers, and follow‑up queries, as shown in synthetic EMS and Munich‑Sim case studies.
Robust Function Calling for LLM Agents Under Interface Drift
AI summaryRobust Function Calling for LLM Agents Under Interface Drift introduces SchemaShield-Lite, a lightweight repair layer that uses schema validation, retrying missing tool names, and one‑shot constrained repair. It boosts Qwen3.5‑9B’s DICE‑300 accuracy from 0.717 to 0.813 and BFCL‑200 from 0.580 to 0.805 without retraining.
Language Models Learn to Improve Language Models
AI summaryLM2LM trains an LLM to propose language‑model improvements via RL using a frozen research world model as a surrogate reward. It outperforms GPT‑6‑sol on held‑out tasks while reducing data‑generation cost by 33% and generalizes to unseen model families.
Sentence-Level Topic Modeling with Lexical and Contextual Evidence
Topic modeling faces a fundamental granularity dilemma: probabilistic topic models assign topics at the word level, whereas transformer-based methods operate at the document level. The latter overlooks that real-world documents typically cover multiple topics, whereas the former…
On the Tip of the Tongue: First-Token Cueing as a Behavioral Probe of Recall Failures in LLMs
Large language models often answer factual questions incorrectly in closed-book settings, and such errors are usually attributed to missing knowledge. Yet models often encode answers they fail to produce. We show that many closed-book errors are first-token recall failures: the…
ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals
Efficient expert routing is a fundamental challenge in Mixture-of-Experts (MoE) systems, particularly for heterogeneous collections of independently trained Large Language Mod els (LLMs). Existing routing approaches typically rely on jointly trained gating networks, prompt-based…
"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws
What the Web community believes about toxic conversations, such as that toxicity rises in longer and deeper threads, was measured mostly with Google's Perspective API, which is shutting down; large language models (LLMs) will make the next measurements. We find that LLM labelers…
HydraLLM: Adaptive Width Slicing for Elastic Text and Vision-Language Models
AI summaryHydraLLM introduces an elastic MQA architecture that adjusts the number of active attention heads to scale model width for any hardware. It outperforms nested baselines by 2.2 % TriviaQA accuracy at 75 % head usage, matches Gemma‑3 4B CLIPScore with 50 % heads, and boosts throughput by 75 % at 25 % heads.
A low-dimensional dynamical basis for tongue shape control
Speech production involves the control of a high-dimensional articulatory system with many degrees-of-freedom, but a widespread hypothesis is that speakers navigate this complexity using a set of lower-dimensional control variables. Common candidates for these control variables…
MuninAI: A Self-Hosted Agentic Framework for Independent Research Groups
Research groups that want a language model grounded in their own literature face a choice between hosted services that receive their unpublished work and self-hosted deployments that keep the data local but give the model no access to the corpus. Munin is an open-source…
Perceive, Solve, Reason: Progressive Capability Formation in Time Series-LLMs
Large language models (LLMs) can now perceive, describe, and reason about images and audio, yet time series understanding remains underestimated: most work uses LLMs as forecasters and classifiers, reasoning studies are few and narrow, and no suite trains and evaluates the full…
Truth in Contrast: Calibrated Factual Evidence from Language Model Likelihoods
Large Language Models (LLMs) encode substantial factual knowledge. Yet, estimating confidence in factual claims reliably remains challenging. We study likelihood-based factual confidence from the complementary perspectives of discrimination and calibration across multilingual…
EduWild: How Teacher-Configured Chatbots Fail in Real-World Use and How to Fix Them
Teachers increasingly configure educational chatbots through natural-language instructions, but how these bots fail in real-world use remains poorly understood. We introduce EduWild, a dataset of 1,726 conversations with 217 teacher-configured chatbots, each paired with the…
ConvexLLM: Scaling Convex-Hull Speech Interfaces to General-Purpose Audio Intelligence
Speech and audio language models connect continuous acoustic representations to decoders pretrained primarily on text data. Most such models use MLP or Q-Former interfaces to map acoustic features into the decoder's input space, requiring the decoder to learn how to interpret…
Execution Before Memory: Diagnosing Memory Failures in Video World Models
A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not…
MARS: Memory-Aware Routing for Session-Level LLM Assistants
Long-term LLM assistants must repeatedly decide what prior context to expose and which model should answer as an interaction evolves. Existing memory systems and model routers usually make these decisions separately, although memory construction changes the evidence,…
Analyzing Occupational Gender Bias across the LLM Training Lifecycle
Post-training is widely assumed to reduce the social biases language models inherit from pretraining, yet direct evidence for this assumption remains mixed. In this paper, we examine how gender disparities evolve across the language modeling pipeline. In particular, we study…
Enhancing RAG-based Medical VQA with Gated Multimodal Representations
AI summaryMultimodal RAG uses a gated unit to fuse image and question into a joint query, allowing evidence retrieval to exploit both modalities while keeping the Med-LVLM frozen. It outperforms single‑modality baselines and rivals state‑of‑the‑art MVQA results.
Not Every Mention Is a Claim: Rethinking Object Hallucination Detection in Vision-Language Models
Object hallucination in vision-language models (VLMs) is usually measured by checking each object a caption names against the image. Most benchmarks and detectors alike treat every object mention as a claim that the object is present. This assumption fails when a caption names…
SIT-Eval: Situation-Based Interactive Turn-by-Turn Evaluation of Conversational Agents
AI summarySIT‑Eval evaluates agents by sampling situational prompts turn‑by‑turn and scoring replies against rule‑based criteria via an LLM judge. Testing eight LLMs shows that aggregate scores mask rule‑specific failures, especially in ambiguity clarification and Socratic questioning.
Trajectory-Triggered Sparse Collaborative Decoding
Dense assistance methods such as speculative decoding repeatedly verify draft tokens with a target model, keeping target computation coupled to generation. We study sparse collaborative decoding, where a draft model generates most tokens and a target model is invoked only for…
PresenceBench: An Exploratory Evaluation of Large Language Model Adaptation to User State
AI summaryPresenceBench evaluates how LLMs adapt to a user’s functional state (readiness, flow, stability) by providing a synthetic state signal and measuring receptivity and specificity. Across four models, policy delivery improved human preference over baseline (≥0.5), but specificity could not be conclusively established.
In Human Memories. Technology, Creativity and Self-Determination: An Ethical Proposal
This research investigates the role of artificial intelligence (AI) in modern compositional practices, with a particular focus on two case studies: Convergence (2020/21) by Alexander Schubert, for String Ensemble and AI System, and In Human Memories (2024) by Francesco Vitucci,…
Beyond the Benchmark: Conditioning, Domains and Failure Modes of Qwen3-TTS
Qwen3-TTS reports state-of-the-art zero-shot voice cloning, but its technical report leaves open how the model should be conditioned, how it behaves outside read speech, and how it fails. We characterise its two cloning interfaces, in-context learning (ICL) and X-Vector, against…
VISTA: Steering LLMs with Latent User Values Inferred from a Few Survey Answers
Personalizing a language model requires knowing what a person values, which stays largely consistent across tasks. Yet most methods either learn task-specific preferences or place the user's profile in the prompt, where models struggle to act on it. We introduce VISTA…
GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search
Optimizing product listings for generative search requires language agents not only to rewrite effectively but also to continually self-improve by retaining successful strategies across diverse categories. However, existing approaches often treat structured listings as free-form…
The Cognitive Cost of LLM Writing: A Reading-Time Comparison with Human Text
Although linguistic differences between human-written and Large Language Model (LLM)-generated texts have been extensively studied, their implications for the cognitive effort involved in reading remain less explored. In psycholinguistics, reading time is commonly used as an…
Recognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM-Based Medical Verification
Publication count is not study count. A model can identify shared trial lineage correctly in a direct task yet over-credit a second publication from the same trial when aggregating evidence. We evaluate this behavior in medical LLM verification using 14 source-audited clinical…
RedDarwin: Evolving Jailbreak Skills against Frontier LLMs
Red teaming must continually adapt as frontier LLM safeguards evolve, yet existing approaches largely search for adversarial prompts or optimize within fixed attack procedures. We observe that even when an attack becomes ineffective, its underlying mechanisms can remain useful…
Bridging the Language Gap in Scholarly Data II: Enhancing Gender Inference for Chinese Names
Perceived gender inference from names is important for inclusive scientometrics research, yet existing methods perform poorly on non-Western scholars. We address this gap for Chinese names, which appear in two forms, Hanzi and Pinyin, its Romanized form (e.g., Zhang Wei, Li Na).…
The Italian lawmaking process: a unified and structured dataset
The Italian legislative process is documented in detailed parliamentary records hosted in institutional websites. However, much of this information is available only as unstructured text, limiting its systematic reuse. We applied a controlled, systematic LLM-based extraction…
VocalGrad: Evaluating Acoustic Perception in Audio Language Models
Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and non-linguistic audio understanding, such as emotion recognition. However, these models still struggle to reliably utilize…
DLite: History-free Speculative Decoding for Agentic AI
Speculative decoding has become a mainstream technique for accelerating large language model (LLM) inference by using a lightweight draft model to generate candidate tokens that are subsequently verified by the target model. However, agentic workloads introduce long and…
Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal
While deriving text embeddings directly from frozen LLMs has emerged as a promising direction, existing methods focus almost exclusively on English, leaving multilingual settings largely unexplored. In this work, we present the first systematic study of multilingual…
Acceptance Is Not Speed: Length Inflation in Training-Free Lossy Speculative Decoding
Speculative decoding typically characterizes speedup by draft tokens accepted per verification round. Lossless verification preserves the target model's output distribution and thus expected completion length, making this metric a meaningful efficiency proxy. Relaxed (lossy)…
Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation
AI summaryReDiST builds student‑conditioned reasoning traces from the learner’s own attempts, revising them with evidence and answer guidance, then selects the best trace via a frozen reader; the distilled student improves over answer‑only and baseline trace methods on multiple QA benchmarks.
DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
Scientific discovery requires developing worthwhile hypotheses into practical investigations, yet learning from this evolving research process remains challenging. We introduce an operational definition of valuable and testable research hypotheses: evidence-grounded claims about…
AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery
AI summaryWe create SEER‑Repro, a benchmark of 20 cancer‑registry studies, and an LLM agent that reproduces 77.4 % of 2,190 claims and automatically tests analytical robustness, flagging potentially consequential choices in 9.5 % of cases for expert review.
Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding.…