New Computation and Language (cs.CL) papers, Friday 23 October 2026
116 papers announced on arXiv. Get them ranked for you, free →
Domain-Specific Constitutional AI: Enhancing Safety in LLM-Powered Mental Health Chatbots
AI summaryThis paper introduces a domain-specific Constitutional AI approach tailored for mental health chatbots. The method integrates specialized psychological principles to improve crisis intervention accuracy and prevent harmful misinformation in sensitive clinical interactions.
ChemBART: Pre-training a Reaction Language Model for Comprehensive Synthesis Planning
AI summaryChemBART trains a reaction‑centric BART model with mask‑filling on whole reaction strings, enabling tasks from precursor generation to property prediction. It integrates into MCTS for multi‑step synthesis, yielding experimentally validated routes that improve literature yields by over 20%.
baseVultra: Compact Nonlinear Recurrence with Reallocated Capacity
AI summarybaseVultra reallocates recurrent capacity from gates to candidate generation, using a single dense H×H matrix and diagonal gates. It outperforms matched LSTM/GRU in compact regimes (TinyShakespeare loss 1.594 vs. 1.604/1.616) and remains competitive at larger scales.
Lighthouse: On Building a Trustworthy Digital Assistant for Exploring History with Civil Rights Archives
AI summaryThe paper introduces Lighthouse, a trustworthy neuro-symbolic digital assistant for exploring historical U.S. Civil Rights archives. It successfully combines large language models with rule-based systems to reduce harmful outputs and reliably answer both factual and expert-evaluated subjective questions using primary sources.
TauIndianBankBench: Procedure-Aware, Action-Oriented Evaluation of Tool-Using Conversational Agents
AI summaryThe paper introduces TauIndianBankBench, a 1,000‑task benchmark that evaluates conversational agents on both final outcomes and every required procedural step in Indian retail banking. It finds that 6.5‑12.0 % of sessions pass only outcome checks yet violate procedures.
FusionEval: Benchmarking Model Fusion for Large Language Models
AI summaryFusionEval is a benchmark suite for evaluating large language model fusion methods across diverse expert domains. Experiments reveal that training-free methods preserve reasoning but degrade tool use, while current training-based approaches require high computational costs to maintain balanced capabilities.
Finitely Verifiable Segmentation: Decidability for Every Tokenizer with a Finite-State Verifier
AI summaryDefines “finitely verifiable” tokenizers: a finite automaton accepts exactly the policy’s segmentations. Shows six key properties (window certifiability, bounded gaps, mask completable, equivalence, limitedness, coaccessibility) become decidable in polynomial time for these rational policies.
Protection Is Not One Event: A Survey of RAG Protection Research by Measured Endpoint
Retrieval-augmented generation (RAG) systems face security, privacy, and copyright risks including content leakage, source reconstruction, unauthorized access, data poisoning, and prompt injection. Yet RAG protection studies use incompatible evaluation metrics, datasets, and…
Measuring Research Progress and Linguistic Inclusion in Low-Resource Languages
AI summaryThe paper introduces the Linguistic Inclusion Framework, analyzing 1,158 papers (2020‑2025) to assess progress on low‑resource languages. It finds that publication growth mainly favors represented languages, with Asian LRLs under‑represented and no region meeting all inclusion criteria.
AQL-Lite: Enabling Efficient Knowledge Acquisition for Arabic LLMs via Data-Centric Continual Pretraining
AI summaryAQL‑Lite curates Arabic data and fine‑tunes a multilingual LLM with partial‑parameter CPT (adapters + last 8 blocks + embeddings) to MSA, achieving a 3.51‑point average gain on 1B tokens and 94 % of full‑parameter gains in half the time.
Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable
AI summaryThe paper trains a policy via self‑rewarded argument quality to improve LLM moral judgments, showing higher decisiveness and order‑invariance on benchmarks, but revealing that consistency under reordering does not equate to valid moral decisions.
CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution
AI summaryCiteAnchor grounds hallucinated citations by source type, adapting retrieval and verification, achieving 0.841 F1 on a detection benchmark and 74.7% success on the new CiteResolve benchmark, outperforming baselines by 7.9–65.3 pp.
Probing and Steering for Safety-Relevant Behaviours in Large Language Models
AI summaryThe paper studies how activation probes—directions in a Llama‑3.1‑8B‑Instruct residual stream—can detect and steer safety‑relevant behaviours (deception, high‑stakes awareness, sycophancy). It shows that detection works best with logistic regression, steering with difference‑of‑means, but the pattern reverses for sycophancy, with performance also depending on token‑pooling and layer choice.
Climbing the Ladder: Automating Data Pipelines and Reproducing Baselines in Low-Resource Settings
AI summaryAutomates data pipeline for Sicilian NMT by extracting aligned sentences from PDFs, replacing Perl tokenization, and training small Transformers versus low-rank fine-tuned NLLB. NLLB with staged reverse-training yields best results, especially for Sicilian→English, but degrades non-trained directions.
MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments
AI summaryMemPlan uses an LLM as a memory‑conditioned bridge to PDDL planning in partially observable text worlds, reconstructing problems each step and applying positive/negative memory to prune invalid hypotheses. It outperforms language‑only and PDDL baselines in success and efficiency while cutting token cost.
Text-Only LLM Correctors Break Correct Transcripts: A Repair-Damage Analysis of Audio Grounding in ASR Error Correction
AI summaryLLM correctors are evaluated by aggregate WER, masking which tokens they fix or break. This study decomposes WER changes into repaired, damaged, and inserted tokens, showing that text‑only correction of Gemini 2.5 Flash’s own draft breaks 13‑32% of correct transcripts, while adding audio reduces damage to 1‑3%.
SIM-UT: A User Simulation based Evaluation Framework for Medical Triage Chatbots
AI summarySIM‑UT simulates realistic, multilingual multi‑turn triage conversations to expose failures of LLM chatbots; evaluation of three WhatsApp health bots revealed a dangerous de‑escalation pattern where models downplayed serious symptoms after user minimization.
X-pressing Intimacy: Mental Health Disclosure After Twitter Acquisition
AI summaryThis paper analyzes over 800 thousand tweets to examine how mental health self-disclosures changed following Twitter's 2022 acquisition. The results show subtle shifts in linguistic intimacy and expression that largely mirror broader platform-wide discourse trends.
Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems
AI summaryThe paper introduces Agent Mosaic, a study of mosaic‑style confidentiality leaks in LLM agents, measuring 7.08% disclosure of hidden composites in simulation with GPT‑5.6‑Sol and 32.37% even after safeguards, showing a non‑negligible leak risk.
Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models
AI summaryThis paper introduces Self-Assessment Reinforcement Learning, a framework that trains retrieval-augmented language models to track their search states and determine missing information. The method uses a dual reward design, improving multi-hop reasoning and deep research benchmarks by over 27 percent without inference-time scaffolding.
Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy
AI summaryThis paper introduces an atomic interpretable unit framework for evaluating sparse autoencoders, revealing that standard metrics miss hidden polysemanticity in dominant features. The results show that widely used SAE features are often functionally redundant and less interpretable where it matters most.
Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences
AI summaryThe authors introduce MemPABench, an interactive benchmark for evaluating how personal agents remember and adapt to user interaction preferences. They find that current memory systems struggle to retrieve preferences, adapt to changes, and overcome default agent behaviors.
Rethinking the Evaluation of LLM Fact-Checkers
AI summaryThe paper introduces JSFactCheckBench, a Japanese dataset that explicitly specifies the question, judging conditions, and source evidence for each LLM-generated claim, revealing that models perform worse on directly answering claims, longer/multiple sources, and when the full QA context is included.
Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals
AI summaryThe paper evaluates how repeatable LLM judges are by running GPT‑4.1 on 500 Arena‑Hard prompts under eight instruction variants. It finds that aggregate scores can be stable while individual verdicts vary widely, highlighting the need to report prompt‑level agreement.
Beyond End-to-End Recovery: Towards a Better Understanding of LLM Steganography via Rolewise Decomposition
AI summaryThe paper introduces a rolewise decomposition framework that separates encoding, decoding, and verification stages in LLM steganography. It shows that end‑to‑end recovery misses 168 valid encodings and 706 hidden messages, revealing distinct encoding/decoding failures that need separate evaluation.
Protegrity-PII Technical Report
AI summaryProtegrity-PII is a new PII detection model for structured and unstructured text, achieving typed recall 0.915 and precision 0.907 on the PII‑SCOPE benchmark. It runs at 0.49 s per million characters, $0.0005 cost, and generalizes better than state‑of‑the‑art models.
IFC-MCQ: Automatically Generating Multiple-Choice Benchmarks for Evaluating LLM Agents on IFC Data Retrieval
AI summaryIFC-MCQ automatically converts IFC building models into multiple‑choice questions with exact answer keys, enabling scalable grading of LLM retrieval on architectural data. The method reveals model‑ and retrieval‑specific failure modes, showing performance varies by IFC file type and interface.
Governed Self-Evolution for Autonomous Investing: A Multi-Agent LLM System with an Invariant Core
AI summarySteward is a multi‑agent LLM system for autonomous investing that uses a type‑system‑bounded invariant core to guard against self‑editing. In a five‑month deployment it limited sector drawdown to 0.3% loss while outperforming a single‑model baseline.
TT-SAX: A Two-Track Time Series Symbolization for Natural Language Shape Querying
AI summaryTT‑SAX redesigns SAX to support natural language shape queries by adding trend letters measured over neighboring segments and aligned with adjectives. It outperforms six classical symbolizations on NLQTSBench, with ablations confirming the three identified limitations.
The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts
AI summaryHybrid retrieval often hurts individual queries: the best fusion (dense‑weighted interpolation) places relevant docs below either component for 13–48% of queries and below dense‑only for 8–21%, even while winning aggregate recall. The paper recommends fusing only when the sparse‑dense balance is favorable.
What the Specification Did Not Say: Specification-Oracle Gaps in Test-Graded Coding-Agent Benchmarks
AI summaryThe paper defines specification‑oracle gaps—tests requiring unstated implementation details—and shows such gaps exist in many coding benchmarks, harming agent performance and skewing rankings; it releases an audit tool that detects these gaps with ~90% precision.
Group Relative Policy Optimisation Improves Multilingual Speech Recognition in Low-Resource Languages
AI summaryGroup Relative Policy Optimisation (GRPO) is applied to a non‑LLM encoder‑decoder ASR model (Whisper‑large‑v3) for 51 African languages. GRPO improves low‑resource performance, yielding SunflowerASR’s lowest published macro‑averaged WER on AfriVox‑v2 (0.227) and SimbaBench (0.372).
Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG
AI summaryCompiles hierarchical RAG control into a policy‑exact parallel inference that replaces serial autoregressive decoding with a single root prefill and causal probes. It matches the serial policy on 84–89% of questions while cutting latency by 35–47% versus token‑wise decoding.
Spectral Smoothing and Spiking Perturbations Reconfigure Capability Profiles and Selective Task Performance in Large Language Models
AI summaryThe authors introduce a lightweight post-hoc spectral intervention on transformer weight matrices that stretches or compresses singular-value spectra while preserving singular vectors. They find that these perturbations induce structured shifts in model capabilities, demonstrating that spectral properties causally encode trade-offs between factual knowledge, language understanding, and reasoning.
WikiOriginQA: Automated Construction of a QA Dataset with Knowledge Origins for Analyzing Cultural Bias
AI summaryWikiOriginQA automatically builds a multilingual QA dataset using Wikipedia edition dates to encode cultural origins; experiments show LLMs are biased toward Anglophone knowledge, highlighting the impact of knowledge propagation into English.
Scaling Laws for Safety Probes
AI summaryThe paper shows that safety‑probe accuracy scales predictably with model size and label count, following two power‑law relationships that hold across tasks, families, and architectures. Using these laws, researchers can choose probe size, label budget, and architecture within 0.02 error of actual performance.
Direct cueing of word identity through word duration: evidence from Mandarin homophones
AI summaryMandarin homophone pairs were tested for word‑duration effects on recognition; participants reliably used shorter duration to identify higher‑frequency words even under noise, indicating word duration is encoded as a phonetic cue in lexical memory.
From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots
AI summaryClinicians design a verified rule‑based guardrail for LLM mental‑health chatbots that replaces ambiguous safety judgments with 16 fact‑based rules. The guardrail eliminates crisis‑resource false alarms while its few remaining errors trace directly to specific rule‑fact conflicts.
Disentangling Robustness and Sycophancy in Multilingual LLMs
AI summaryThe paper introduces the SC‑matrix to separate model competence from sycophancy in multilingual LLMs, showing that aggregate sycophancy rates hide different failure types, high competence doesn’t guarantee resistance, and reasoning‑based mitigations beat directive methods.
ULME: Predictive Coupling for Mutually Reinforcing Embedding and Language Modeling
AI summaryULME post‑trains a causal decoder‑only LLM to unify text embedding and generation by adding learnable slots that receive a contrastive semantic embedding and a predictive reasoning embedding, enabling the generator to shape the embedder. Across Qwen variants, ULME boosts MTEB by 3.3‑4.1 points and tightens retrieval benchmarks while matching generation accuracy, showing mutually reinforcing capabilities.
Meaning Contracts: Local Semantic Geometry in Large Language Models
AI summaryThe paper probes the fine‑scale geometry of meaning in LLMs using cross‑lingual and refinement probes, revealing that semantic factors form invariant cores with structured, realization‑specific variation that contracts into tighter, anisotropic neighborhoods when meanings are refined.
UniGuard: Multi-View Guardrail Learning for Cross-Interface Safety Generalization
AI summaryUniGuard learns interface‑invariant safety representations from multi‑view data, outperforming a 0.6B guardrail by 45.12 F1 points on ATBench, PolicyGuardBench, and DynaBench, while cutting expected calibration error by 35.41 points.
How Firm Are AI Models' Ethical Boundaries? A Pay-Stub Editing Test of 15 Models
AI summaryThe study shows that small wording changes can flip 15 AI models from refusing to comply with a pay‑stub editing request, with a 66.2% compliance rate for “printing error” phrasing, while some models (Astra) always refuse and others (Fable, Opus, Muse Spark) never comply.
Abstract Priorities, Concrete Rewards: How Stated Preferences Relate to What People Reward in AI Responses
AI summaryThis study compares users' stated AI preferences with their actual response ratings using the PRISM dataset. It finds that stated priorities poorly predict the concrete features that actually drive rewards, complicating effective AI personalization.
Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages
AI summaryThe study compares relations among 196 basic concepts in 81 languages using multilingual embeddings. It finds higher cross‑language similarity for sociocultural meanings than for nature or motion, with differences growing with geographic separation.
MixJudge: Learning a Language-Queryable Critic for Vocal Mixing
AI summaryMixJudge is a language-queryable audio critic trained on a synthetic curriculum of vocal mixing issues. It accurately assesses, ranks, and identifies mixing problems, and its judgments effectively guide automated audio processing decisions toward professional standards.
Invisibilization as a Source of Probability Bias in Language Models
AI summaryThe study shows that leading LLMs often refuse to evaluate information from reporters based on race, rejecting Black reporters 46 % versus 9 % for white ones, while actual trust given when accepted is similar. A scoring mechanism reduces this refusal bias to < 2 %.
TutorCaseBench: A Case-Grounded Benchmark for Evaluating LLM Tutors in Supporting Students with Challenging Behaviors
AI summaryTutorCaseBench evaluates AI tutors on 923 case‑grounded dialogues for five student challenge types, showing that even the best LLMs satisfy rubric criteria only 55.8% of the time, revealing a gap between recognizing interventions and generating context‑specific responses.
TRIAD: Benchmarking Omni-Modal Ambiguity in Multimodal Large Language Models
AI summaryThe paper introduces TRIAD, a benchmark assessing how omni-modal large language models resolve ambiguity distributed across text, image, and audio inputs. Evaluation of fourteen systems reveals that current models struggle with scene-level disambiguation and significantly trail human performance, often underusing audio cues.
Synthetic-to-Real Code-Switched ASR: A Spontaneous Nepali-English Podcast Benchmark
AI summaryThe paper introduces a 15.45‑hour Nepali‑English code‑switched podcast benchmark and a two‑stage synthetic‑to‑real adaptation: synthetic TTS fine‑tuning reduces MER from 23.40% to 21.52%, while policy‑optimization further lowers English script confusion and SPER.
One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation
AI summaryOn-policy self-distillation uses a single teacher conditioned on one reference solution, which causes reference-induced target instability. Monte Carlo Privileged Distillation averages over multiple partial reference views, yielding more stable supervision and better performance on reasoning benchmarks.
Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models
AI summaryThis paper traces the origin of the reasoning marker wait in language models. It shows that the marker is inherited from training data rather than invented by reinforcement learning, proving that its frequency does not reliably signal model reasoning.
From Perpetrators' Digital Footprints to Behavioral Communities: An Ontology-Driven Framework for Bengali Cyberbullying Analysis--- FULL VERSION
AI summaryOntology-driven framework models Bengali cyberbullying perpetrators as 64‑dimensional behavioral vectors, builds a weighted network, and finds six communities (modularity 0.495); key features are Video, Impact, and User Profile, with IM‑Greedy seeds yielding modest influence gains.
Silicon Samples Match the Answers but Not the Respondents
AI summarySilicon sampling creates synthetic survey respondents using LLMs, but its generated respondents are far from any real respondent. Our respondent‑coverage‑gap metric shows all evaluated LLMs fail to cover real respondent space, even outperforming a Gaussian copula baseline.
Hierarchical Fine-Grained Cross-Modal Alignment via Dynamic Granularity Discovery and Optimal Transport
Fine-grained image-text retrieval requires a model not only to assess global semantic consistency between an image and a text, but also to model visual evidence for local textual semantics, including objects, attributes, quantities, actions, and relations, so as to distinguish…
ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents
With external tools, large language models (LLMs) can tackle complex tasks through sustained environmental interaction. LLM-based agents are therefore increasingly used to tackle complex long-horizon tasks in real-world environments. However, these tasks involve sequential,…
Verbalizing Activation Differences with Natural Language Autoencoders
Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing activations through a natural language bottleneck. However, it remains unclear…
Tracing Hallucinations in LLM Reasoning for UML Class Diagram Generation: An Empirical Taxonomy and Mitigation Study
Large Language Models (LLMs) can generate UML class diagrams from natural language requirements, but hallucinations may compromise the correctness and completeness of the resulting diagrams. Existing evaluations mostly assess the final diagram, without examining the intermediate…
RetiJev: Fast and Source-Grounded Ophthalmic Decision Models
Ophthalmic clinical decision support requires fast, accurate, and privacy-preserving inference, yet autoregressive large language models (LLMs) incur substantial latency and computational overhead for decisions among predefined actions. We therefore investigate a…
IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems
We present IQuest-Q1, a sparse Mixture-of-Experts (MoE) agentic foundation model with 320B total parameters (15B activated). IQuest-Q1 is developed through a three-stage process: (1) stable and efficient pre-training, combining Stable SplitNorm, Muon, and muP with a two-stage…
Once a Tree, Always a Tree? Cross-lingual Transfer of Hierarchical Generalization in Language Models
Language models (LMs) often exhibit successful cross-lingual transfer through continued (pre)training, still their underlying process remains unclear. One possibility is that LMs can leverage the evidence toward language-universal principles, namely, hierarchical structure, from…
Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters
Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive. We ask how small an encoder for such decisions can be and whether it should read the options jointly or separately. In a…
A Reality Check of Inference-Time Context Augmentation Methods
Large language models (LLMs) often fail to utilise the context given in the input, resulting in hallucinations or a fallback to parametric knowledge. Context augmentation methods aim to improve model performance by modifying the input, from inserting filler tokens that add…
Interpretive Variation and Discourse Structure in Detecting Enthymemes in Controversial Political Discourse
Enthymemes are arguments that leave a premise or conclusion unstated, and they are a common persuasion device on social media. Because readers must fill in the missing part themselves, they often disagree on whether a text contains one. This paper asks how to detect and evaluate…
ColaBench: Evaluating Memory-Guided Knowledge Acquisition in Personal Agents
Personal agents need to draw on long-term interaction history to retrieve and apply external knowledge suited to a user's current request. Even an answer grounded in relevant documents can be unsuitable when it overlooks the user's circumstances. This creates two connected…
Padding to k: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists
Extraction pipelines ask language models for JSON that follows a schema, and the schema can set a minimum number of list entries. We ask what models do when the document holds fewer items than that minimum. Under a preregistered plan, we test eight open-weight models on 558 real…
Mostly Reading: What Open-Book Scores Measure in Document-Level LoRA Adapters
AI summaryTrains a LoRA adapter per Wikipedia article on Llama‑3.1‑8B; finds open‑book scoring largely reflects a generic “reading” capability rather than document‑specific knowledge. Specialists improve cloze recall only ~0.094 and their updates are document‑aligned, but merging adapters recovers most open‑book advantage (≈88 %) yet only ~12–16 % of cloze gain.
Chunk Recomputation under Real Document Revisions: Measuring Update Cost and Retrieval Quality of Fixed, Recursive, and Content-Defined Chunking
Chunking strategies for retrieval-augmented generation are almost always evaluated on a static snapshot, yet the documents underlying a knowledge base continue to change. Because chunks are embedded independently, a boundary that moves after an edit forces unchanged text to be…
Harm Is Behavioural, Moderation Is Lexical: How Language and Culture Decide Whom Automated Safety Systems Protect and Whom They Silence
Content warning: this paper quotes slurs and hateful statements from published test suites; the racial slur among them is masked. Automated systems decide at scale which messages are harmful, and they decide unevenly. Classifiers label African American English as offensive up to…
SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs
Policy enforcement for LLM agents has shifted from single-turn content moderation to compound, application-specific policies evaluated over multi-turn transcripts and tool-call trajectories. Fine-tuned guardrail models are efficient but degrade on compound policies, and while…
What Pairwise Evaluation Misses: Reward Scores across Prompts
Reward models score responses for preference optimization, filtering and ranking, and pooling scores across prompts needs them to be comparable. Yet pairwise evaluation compares two responses to the same prompt, so any score component shared by a prompt's responses cancels and…
RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training
AI summaryThe paper introduces RIFT, a KV cache compression method for speculative decoding draft models that combines adaptive refreshing and compressibility training. It enables a draft model to retain over 98% of its full-cache acceptance length under tight memory budgets.
Learning What to Ask for Tomorrow: Active Preference Acquisition for Long-Term Personalized Web Agents
Web agents increasingly serve the same user across repeated online tasks such as booking travel, shopping, and scheduling, so their long-term usefulness depends on how well they know the user's preferences. Passive memory and personalization methods cannot obtain preferences the…
A Grammar of French Civil Cassation: A Ground-Level Dataset and Tool
AI summaryDataset of 121,536 ground–reason pairs from 86,464 French Court of Cassation rulings (2016‑2025). The authors automatically pair each appeal ground with its court answer, classify them into doctrinal families, and provide a descriptive grammar of civil cassation, releasing data, code, and a tool.
Through a PRISM, Clearly: Step-Sequence Reasoning for Mistake Recognition in Procedural Videos
Procedural mistake recognition requires identifying deviations from valid multi-step activities. We study post-completion recognition from videos with benchmark-provided step boundaries. We introduce PRISM (Procedural mistake Recognition via Interpretable Semantic Modeling),…
Delta N-gram: Forward-Only Continual Writing into a Language Model's Native N-gram Memory
Our previous analysis found that a model's native hashed N-gram table is where continual updates disturb the base least, but its writes needed backpropagation and changed the pretrained table. We present Delta N-gram, a forward-only write path into that table. The pretrained…
Mixing Matters: Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering
Long-context models often use answer-bearing evidence less effectively when it appears away from a prompt boundary. We compare this position bias across Transformer, state-space, and hybrid Mamba architectures by moving one gold document through a fixed set of nine distractors.…
Prompt Sensitivity in Tool-Augmented Model Comparisons
Language models are compared on question-answering benchmarks that pose each question in one wording, and a reworded question can change which of two models scores higher. We ask whether giving the models tools makes such comparisons more stable. An accuracy gain cannot answer…
Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research Reports
Deep research agents synthesize open-web evidence into long reports, but current benchmarks do not evaluate research centered on a source video. We introduce Video-Centered Deep Research Report Generation. Given a video and user request, an agent identifies research questions,…
Prompting Strategy Effects Depend on the Judge: A Multilingual Study of LLM-as-a-Judge Robustness
LLM-as-a-Judge evaluation is increasingly applied to multilingual tasks, yet prompting strategies such as Chain-of-Thought and Role Conditioning are recommended based solely on English benchmarks. Whether their effectiveness transfers across judge models and languages remains…
When More Metadata Misleads: Probing Field-Level Grounding in LLM-Generated Museum Descriptions
Metadata-conditioned generation is often framed as a data-to-text task in which additional input fields should improve grounding. Museum description generation challenges this assumption because catalogue records are heterogeneous, partially visual, and unevenly documented…
Investigating Multi-Cultural Value Selection Biases in LLMs
AI summaryWe test LLMs with simultaneous country and identity cues, finding language cues shift responses three times more than explicit country names but in unexpected directions, while explicit cues produce more distinct value changes. Results show country and language signals should be evaluated separately.
Predicting Music Valence and Arousal via Textual and Multimodal In-Context Learning in Large Language Models
While arousal can be reliably predicted from musical audio, emotional valence remains difficult to model due to its subjective and context-dependent nature. In this work, we investigate the potential of in-context learning (ICL) with large language models to advance continuous…
IndoNutriQA: An Indonesian Clinical Nutrition Question Answering Benchmark
Clinical Question Answering (QA) underpins Clinical Intelligent Decision Support Systems, but progress is severely limited for low-resource languages such as Indonesian by the absence of public domain-specific data. We introduce IndoNutriQA, an expert-curated Indonesian clinical…
A Systematic Comparison of Search Space Complexity and Efficiency for Time-Synchronous End-to-End ASR Models
Search is a central component of Automatic Speech Recognition (ASR) systems and has been deeply investigated for classical ASR modeling based on Hidden Markov Models (HMMs). However, in work on current sequence-to-sequence ASR architectures it is often simplified to a small…
Memory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use
Memory enables language model assistants to maintain continuity across sessions and generate personalized responses, but knowing more about a user does not necessarily produce responses they prefer. Existing evaluations assess memory capabilities without directly establishing…
CoWriter: Guiding Methodological and Synergetic Human-LLM Co-Authoring
While LLMs have transformed text generation, standard chat-and-canvas interfaces often reduce human authoring to quite passive supervision, causing review fatigue, over-reliance on AI, and the sidelining of the human role in long-form, source-grounded writing. Prior systems…
The Value of Memory in Language Agents
Memory enables language agents to carry useful distinctions from past experience into future decisions. Its value depends on which distinctions survive, whether a bounded agent can use them, and what can be recovered or revised as demands evolve. We develop a decision-theoretic…
Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation
Hierarchical language generation separates continuous semantic modeling from discrete text realization, but its effectiveness depends on what the latent space compresses and how the discrete generator uses it. We propose Cola2, a framework that jointly studies continuous…
When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues
Large language models can repeat an instruction while failing to keep obeying it. We study this recall-compliance dissociation in long-horizon dialogue and define State Integrity Collapse: a model preserves a stated constraint in memory but stops applying it to later turns. We…
SurveyGraph: Generating Academic Surveys via Methods-Entity Knowledge Graphs
Most automated survey generation systems retrieve papers as flat documents, limiting their ability to capture research structure. We present SurveyGraph, a framework built around a Methods-Entity Knowledge Graph (MEKG) encoding method-level follow-on, competition, and…
CrisiLangs: A Multilingual Resource for Crisis Warning Generation
AI summaryCrisiLangs extends English crisis‑warning data to Italian, Bulgarian, German, and Greek, providing 5,000 aligned event‑message pairs. Fine‑tuned multilingual models beat zero‑shot baselines on ROUGE‑L, BLEU, and BERTScore, with low decoding temperatures preferred across all languages.
AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents
Many LLM agents maintain external persistent memory and derive additional state from past interactions, such as summaries. Deleting an original memory record therefore need not remove information carried by its descendants. We introduce AfterDeleteBench, a controlled behavioral…
A Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy, Evolution, and Future Directions
Visual document understanding is essential for acquiring knowledge from visually rich documents and is attracting growing research interest. Existing surveys emphasize architectures or technical approaches, with limited attention to how task bottlenecks shape technical choices…
Parameter-Efficient Adaptation of Large Language Models to a Low-Resource Agglutinative Language: A Corpus, Datasets, and a Reproducible Benchmark for the Bashkir Language
Bashkir is a low-resource agglutinative language written in Cyrillic and remains poorly represented in modern LLMs. We present an open resource suite and a systematic benchmark of adaptation methods: the Bashkir Web Corpus (71,567 documents, 46.9M tokens, 16 sources) and three…
HERD: Harness-Evolving Recursive Distillation for Continual Learning
Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage…
Same Entity, Different Script: Diagnosing Cross-Script Entity Matching Failures in Language Models
Multilingual language models can give different identity judgments when the same people's names are written in different scripts. We study this effect using 601 audited identity pairs with attested Devanagari and Latin forms, holding the question and relation fixed. Qwen2.5-7B's…
SQuAT: Spherical Lattice Vector Quantization-Aware Training
Vector quantization (VQ), in particular lattice-based methods, offer strong rate-distortion trade-offs for low-bit large language models (LLMs), yet their behavior under quantization-aware training (VQ-QAT) remains poorly understood. We scale lattice VQ-QAT to billion-parameter…
SyRA-CS: Synthetic-to-Real Adaptation for Code-Switched ASR with Meta-Learning
Code-switched (CS) automatic speech recognition (ASR) remains difficult because real CS speech is scarce and synthetic speech does not fully match it. We investigate whether synthetic CS speech is more useful for learning an adaptation-friendly initialization than for ordinary…
How Depth Growth and Looping Shape In-Context Learning in Language Models
AI summaryThis paper investigates how parameter reuse in looped and depth-grown language models affects in-context learning. The results show that these architectures develop a stronger inductive bias toward context, acquiring and retaining in-context learning earlier and more reliably than standard models by propagating induction heads across layers.
How Well Do Decision Models Transfer to Vietnamese? An Empirical Study of Text Classification
AI summaryEvaluates four decision models for zero‑shot Vietnamese text classification across 13 settings, showing JEV performs best with English labels, while bilingual labels help little; diacritic removal degrades performance by 9–15 % macro‑F1, highlighting label choice and orthography effects.
Multi-Token Prediction and Alternative Tokenization Improve Learning Latent Multi-Hop Graph Search at Scale
Given an unseen graph, a start node, $s$, and an end node, $t$, can a language model generate the shortest path from $s$ to $t$? Prior work has found that models fail to learn this task as the graphs scale. We show that this failure has two causes: the supervision signal and the…
How Different, or Different How? Stance Granularity in Viewpoint Selection for Dialog Systems that Engage Users with Alternative Views
Systems that aim to help a user engage with viewpoints other than their own have to decide how different those viewpoints should be. Viewpoints are mainly represented on a scalar scale, too coarse to distinguish viewpoints that disagree with the user about different things. We…
From Structured Records to Patient Narratives: Preserving Clinical Alignment in LLM-Generated Health Summaries
As wearable and digital health systems increasingly generate longitudinal physiological records, communicating these observations in patient-accessible language remains a significant challenge. Large Language Models (LLMs) offer a promising mechanism for translating structured…
Measuring the Verifier's Floor: How Much Apparent Rare-Word Text-to-Speech Error Is ASR Error?
AI summaryThe paper quantifies how ASR errors inflate TTS error estimates, showing that 33–67% of very rare words are missed by recognizers even on human speech. Even with multiple recognizers, rare words still dominate the remaining error, requiring a listening test.
Towards User Simulators That Stay On Task And Still Sound Natural
Users increasingly rely on AI systems for extended, multi-turn tasks, yet evidence indicates that LLM-based system performance degrades over prolonged interaction. To evaluate and improve AI assistants in such settings, we need user simulators that faithfully follow complex user…
The Geometry of Metaphor: Probing and Steering Metaphor Representations in Large Language Models
AI summaryResearchers identify causal subspaces representing metaphor in various LLMs and show that manipulating these subspaces can steer metaphor-related classification and generation tasks, enabling control over the metaphoricity of model outputs.
An Empirical Study of Recent Large Language Models for Knowledge Graph to Text Generation on Standard Benchmarks
Knowledge graph-to-text (KG-to-text) generation verbalizes a set of knowledge graph triples into fluent natural language. While fine-tuned encoder-decoder models have long been the standard approach, recent LLMs are becoming increasingly competitive. We benchmark recent LLMs…
Diffusion Language Models for Perplexity-Aware Black-Box Attacks
\emph{Continuous-space} attacks find prompts that trigger malicious behaviors, such as overthinking, in black-box LLMs by sampling or optimizing over a low-dimensional Gaussian latent space and decoding it to text through a \emph{random mapping} and nearest-token lookup. Because…
Would an LLM Match into a Residency? Uncertainty, Reasoning and Image Use of Open Language Models on the Spanish MIR Examination
Accuracy on medical examinations is common evidence offered for the clinical use of language models, yet a health system that intends to run an open model on its own hardware needs to know considerably more: whether the model recognizes when it is likely to be wrong, whether…
How Ground Truth Shapes Retrieval Scores: Evidence-Aware Evaluation of a Retrieval-Augmented Generation System for Breast Lesions
Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrieval-augmented generation (RAG) system over breast lesion management literature, addressing the dependence of retrieval scores on annotation coverage. Methods: An 18-document corpus and an…
Bilingual Language Models Capture Cross-Linguistic Influence in Second-Language Reading
While monolingual language model (LM) surprisal is widely established as a predictor of reading times, it cannot capture the non-native (L2) reading dynamics that are shaped by the reader's first language (L1). In this work, we investigate whether the cross-lingual transfer…
Can an Agent Debug a Web Corpus? Agentic Auditing of Slovak FineWeb2
AI summaryA tool‑using LLM agent debugged 26.5 M Slovak documents in FineWeb2, finding seven issues, including a 5.32 % machine‑generated content farm that slipped through filters. The agent produced an executable test suite that scans the corpus in 0.7 s, revealing precise prevalence and pipeline failure reasons.
Low-Bit N-Gram Embeddings for Language Models
Many recent open-weight language models use n-gram embedding tables to improve performance. These tables store learned representations of short token sequences, such as pairs or triples of consecutive tokens, which the model retrieves during processing. However, they can require…
Cost-Aware Abstention for LLM-Based IT Ticket Classification
AI summaryThis paper evaluates cost-aware abstention policies for IT ticket classification by selecting classification thresholds on calibration data under explicit error costs. The method reduces routing costs by 38.1% compared to baseline approaches on a standard support ticket corpus.
Direct Pass@$k$ Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning
AI summaryThe paper proposes a pruning method that directly optimizes pass@$k$ by selecting partial solutions that maximize the chance that at least one will finish correctly, balancing quality and diversity automatically. It outperforms prior quality‑ or diversity‑based pruning on eight reasoning models, especially when retaining a moderate number of candidates.