All papers

The paper evaluates how language exposure distribution and ordering affect the performance of multilingual models trained with limited data.

This paper introduces plug-in decision heads that read action choices directly from visual agent hidden states rather than generating text tokens.

This paper surveys 138 studies on sparse autoencoders in NLP, examining dictionary design, feature recoverability, and practical utility.

This paper introduces Gap Detection, an optimization method that identifies unoccupied semantic directions to reduce homogenization in open-ended model answers.

This paper presents RM-OPD, a framework that trains domain-specialized reward models via on-policy distillation into a generalist model.

This paper introduces TreeSpark, a framework training semi-autoregressive drafters to generate candidate trees aligned with speculative verification.

This paper analyzes career path prediction systems, reporting that explicit skill representations provide marginal improvement over text-only job descriptions.

This paper proposes a joint scaling law evaluating how adding experts in multilingual mixture-of-experts models affects language capacity and loss penalties.

This paper introduces AFFIRM, using affine fields on frozen speech models to repair spoken access to recently updated text knowledge.

This paper introduces Anuvani LID, a dedicated 115M parameter model for spoken language identification across 22 Indian languages.

This paper introduces an annotated part-of-speech tagging dataset and evaluates pretrained transformer encoders for the low-resource Saraiki language.

This paper proposes PUlSE, using internal layer representations to detect factual errors and trigger step-level retrieval during medical reasoning.

This paper proposes MMLA-RSI, a recursive self-improvement framework using reasoning-time policy training and authoritative memory adaptation.

This paper presents Anuvani-Diarization, evaluating zero-shot speaker diarization across 22 Indian languages with uncertainty-aware silver supervision.

This paper investigates whether gossip discourse naturally emerges in autonomous AI agent online communities compared to human discussion boards.

This paper measures semantic drift and stability across AI agent communities, finding higher drift and lower temporal continuity than human forums.

This paper proposes SCOUT, a cascaded two-track framework combining visual candidate filtering, offline semantic summaries, and query refinement for document retrieval.

This paper introduces a multi-hop reasoning benchmark to evaluate fine-grained, explainable hate speech analysis in Persian.

This paper introduces FOLD, a training-free speculative decoding approach treating distant tokens as position-free far memory for long-context generation.

This paper introduces COUNT, scaling language model memory without training parameters by storing precomputed n-gram counts directly in host memory.

This paper proposes Claim Cards to make agent interpretations explicit when verifying ambiguous natural-language scientific claims.

This paper introduces MolLean, a verifiable language and tool interface that grounds chemical reasoning in formal state transitions and explicit checks.

This paper introduces ChronoCulture to evaluate how well language models capture temporal changes in cultural attitudes across 72 countries.

This paper introduces AdaRSI, an on-demand recursive self-improvement framework for embodied agents that isolates and fine-tunes only persistent sub-step failures.

This paper evaluates multilingual bias audits across ten language models, showing that lower stereotype scores often reflect translation artifacts rather than fairer behavior.

This paper presents ReaDy, a speculative decoding method using hybrid linear attention recurrent states to provide position-specific context during drafting.

Documents safety alignment methods, negative preference tuning outcomes, and automated judge limitations for a Portuguese language model.

Compresses long-context key-value caches using test-time training on generic document prompts without task-specific labels.

Evaluates closed and open models on English and Bengali abstractive summarization using lexical metrics, factuality checks, and automated judges.

Introduces a welfare-aware objective to allocate limited human feedback across diverse individuals when training pluralistic alignment models.

Combines spectral graph coarsening with continuous relaxation to find high-quality solutions for compute-constrained MAX-CUT problems.

Argues that inferential accuracy fails to evaluate agent coordination safely because correlated errors can compound into costly team actions.

Introduces a prompt optimization framework that monitors item-level regressions and recoveries rather than relying solely on aggregate accuracy.

Presents a large graphical user interface grounding dataset and a combined distillation and reinforcement learning training method for agents.

Surveys dataset design, construction methods, and evaluation protocols for improving spatial reasoning in multimodal language models.

Analyzes how language models develop in-context learning by separating simple demonstration copying from multi-token rule composition across training checkpoints.

The paper proposes a training-free structured pruning method that trims multilayer perceptron layers while preserving attention mechanisms for reasoning tasks.

The paper presents a system using technical language processing to transform industrial maintenance work order records into an ontology.

The paper proposes confining feature distillation loss to language-routed adapters to balance performance in compact multilingual speech recognition models.

The paper introduces a benchmark assessing language models on operational planning from incomplete, complex, and contradictory military document sets.

The paper introduces a benchmark to evaluate whether frontier language model agents can execute open scientific research and produce follow-up papers.

The paper introduces an evergreen benchmark to evaluate the ability of language models to retrieve and format multi-element factual lists.

The paper presents a method to retrofit attention models with hierarchical test-time neural memory modules to improve long-context inference efficiency.

The paper analyzes whether syntactic attention heads in transformers play a direct causal role in syntactic language predictions.

The paper proposes adding an explicit copy-paste prediction head to language models to improve in-context retrieval and copying efficiency.

The paper presents a multimodal agent designed to trace the origin, spread, and potential safety risks of online memes.

The paper introduces a framework to audit cultural, geographic, and user-representation biases embedded within popular machine learning benchmarks.

The paper details the development, training data criteria, and performance of the open Japanese sovereign models, LLM-jp-4.

The paper introduces a test showing that steering vectors designed for competing ethical frameworks behave interchangeably in open-weight models.

The paper analyzes Best-of-N sampling as an extreme-value competition between tails of correct and incorrect responses to explain reward hacking.

The paper provides an analytical framework for how multimodal agents form, represent, and adapt memory over long-horizon tasks.

The paper proposes a fine-tuning method that trains language model agents using textual expert critiques alongside corrected actions.

The paper evaluates whether language models genuinely benefit from compressed memory by comparing them against memory-free baseline models.

The paper introduces a speech dataset spanning 675 underrepresented languages to support speech recognition and text-to-speech technologies.

The paper introduces a benchmark to evaluate how well language models recognize, infer, and coin newly emerging Korean terms.

The paper proposes a Bayesian cue-combination model to explain how vision-language models integrate conflicting visual and textual evidence.

The paper introduces a framework for recursively composing, verifying, and persistently reusing modular agent capabilities across diverse problem domains.

The paper introduces a benchmark to evaluate how well language models understand legal philosophy and apply it to legal reasoning.

The paper introduces a document-level data augmentation method using discourse structure parsing to preserve logical flow during text generation.

The paper proposes a reinforcement learning framework that decides when and what external knowledge to retrieve for machine translation.