All papers

The paper introduces an attention conversion method enabling multi-head latent attention on modern architectures with reduced memory footprints.

The paper formulates online language model persona personalization as a dueling bandit problem using sequential pairwise user preferences.

The paper introduces a 174-hour speech benchmark and evaluates baseline models on Ewe automatic speech recognition.

The paper tests whether role-playing language models leak private or unobserved character information across simulated interactive scenes.

The paper introduces a framework combining guidelines and agentic linking to ground medical concepts across diverse text types and ontologies.

The paper introduces a lightweight guidance model to estimate parallel decoding inconsistency and accelerate diffusion language models.

The paper evaluates five language models on extracting complete interval networks of relative temporal events from narratives.

The paper studies compute allocation across multi-agent coding teams to identify Pareto-optimal trade-offs between cost and performance.

The paper trains language models to report internal states through supervised fine-tuning and on-policy learning.

The paper investigates lexical representations in transformer models by decomposing representations into item, class, and interaction components.

The paper proposes an edge-cloud collaboration framework that selectively delegates complex agent steps from local models to cloud models.

The paper introduces an annotated dataset and retrieval approach for Tagalog homograph diacritic restoration.

The paper studies long chain-of-thought distillation and argues improvements stem from structured reasoning patterns rather than factual correctness.

The paper presents a runtime framework that tracks evolving repository state and dependencies to assist long-horizon coding agents.

The paper introduces a diagnostic protocol to evaluate how speech recognition errors affect structured tool calling in voice agents.

The paper proposes disaggregating teacher and student models across separate nodes with a fused loss to accelerate knowledge distillation.

The paper proposes an unlearning method that removes sensitive facts from model responses while preserving surrounding non-sensitive factual context.

The paper compares discriminative and generative language models for rhetorical structure parsing and elementary discourse unit segmentation.

The paper presents a retrieval-aware denoising method for conversational user profiles to improve personalized dense retrieval.

The paper investigates methods for encoding context into model weights so that language models properly apply the temporary context during subsequent inference.

The paper evaluates how language models correct user misconceptions across five languages, showing indirect pushback often fails to fix factual errors.

The paper analyzes geometric properties of text embeddings to predict how many nearest neighbors an individual example loses under dimensionality reduction.

The paper compares direct prompting, model-distilled supervision, and larger reader datasets for predicting user text highlighting behavior on web pages.

The paper proposes ReMaRk, a reinforcement learning method that trains recommender models using all available reader replies simultaneously to reduce variance.

The paper introduces HopToM, a benchmark testing whether higher-order theory-of-mind tasks in language models genuinely evaluate nested reasoning rather than shortcuts.

The paper presents GENDIS, a multilingual dataset of YouTube comments from India, Kenya, and Nigeria annotated for gender-related discourse and stance.

The paper introduces BRIDGE, a runtime intervention framework that detects and repairs broken information handoffs across dependency graphs in multi-agent workflows.

The paper tests activation steering as a text watermarking technique, finding that standard logit-based watermarking provides more robust detection against paraphrasing.

The paper proposes a training approach that uses learned chunk routing to speed up long-context diffusion language model processing.

The paper introduces Swiss4Arg, a multilingual argument mining dataset based on official Swiss voting brochures across four national languages.

The paper proposes Batch-Paced, a method that dynamically allocates decoding compute across batched requests to improve discrete diffusion language model inference.

The paper introduces Agent-Reflex, a training framework designed to teach language agents how to reflect and recover from noisy environment feedback.

The paper examines machine grading on undergraduate essays across three universities, finding models compress score distributions and vary across institutions.

The paper introduces BanglaTOD, a tri-lingual benchmark for task-oriented dialogue understanding covering Bangla, English, and romanized Banglish across nine service domains.

The paper introduces an optimization method that jointly refines communication instructions between communicating agents to improve multi-agent collaboration.

The paper probes language model hidden states across generated stories to test whether models anticipate narrative events multiple steps ahead.

The paper evaluates typed decision models against standard language model judges to measure how emotional tone and biases affect their evaluation accuracy.

The paper presents SparPR, a runtime key-value cache recovery mechanism that reduces memory bottlenecks during model inference on consumer hardware.

Evaluates AI text detectors under attacks applied to both human and machine text, revealing high false-positive rates on edited human writing.

Proposes dynamic activation steering using preceding denoising predictions to reduce hallucinations in diffusion language models.

Introduces a benchmark evaluating coding agent misbehaviors across code generation, file management, and system environment integrity.

Surveys methods for compressing large language model benchmarks to reduce evaluation costs while maintaining accurate score estimates.

Presents a multilingual framework and dataset spanning four languages to evaluate clinical pain representations across exams and clinical notes.

Introduces a human-verified benchmark to evaluate how accurately reward models judge language model outputs on legal tasks.

Develops a simulation framework using persona-guided models to optimize comparative doctor summary presentations for medical patients.

Introduces a low-latency guardrail model that routes incoming agent actions through editable policy trees for safety auditing.

Probes layer representations in multilingual language models to study whether internal color geometry aligns with human perception or language syntax.

Demonstrates an attack method that alters embedded multi-bit text watermarks without knowing the secret generation key.

Equips diffusion language models with multiple residual streams and adaptive routing to separate visible context from masked token generation.

Shows linear prompt-injection monitors overfit to conversation features and reduces their false alarms by adding matched harmless negative examples.

Proposes estimating both contextual and unigram surprisal from the same language model to reduce estimation mismatch in reading time studies.

Studies agent-to-agent negotiations over money, finding that the receiving agent's characteristics primarily dictate the final transfer amounts.

Presents a randomized experiment framework to test and curate self-written agent playbooks, preventing task failure caused by unverified lessons.

Identifies passage copying as a failure mode in long-form retrieval systems and introduces sentence-level credit assignment to encourage genuine grounding.

Proposes a teacher-guided reinforcement learning method using bidirectional constraints to help small language models learn effective tool use.

Evaluates the safety of ChatGPT health advice using real user conversation logs across four Global South countries.

Introduces a benchmark evaluating how reliably software coding agents extract and digitize complex unstructured clinical trial protocol documents.

Demonstrates that knowledge distillation transfers safety dispositions and behavioral tendencies from teacher models to student models on unrelated tasks.

Analyzes retrieval divergence across languages in multilingual dense embedders and introduces disagreement-guided steering to align shared-index search results.

Evaluates fast, system-one decision models and language models across fifteen medical datasets, testing their accuracy, format robustness, and abstention awareness.