All papers

The paper analyzes the impact of teacher rationales before distilling medical reasoning capabilities into compact language models.

The paper presents an empirical study evaluating syntactic language models at scales up to 1B parameters across linearizations and masks.

The paper introduces a benchmark to evaluate whether language-model guardrails properly distinguish authorized from unauthorized actions across identical operations.

The paper introduces a 30B tool-use model and an execution-grounded dataset to scale agentic tool supervision with verifiable environments.

The paper evaluates the impact of sarcasm on Turkish hate speech detection models using a targeted external test dataset.

The paper studies low-rank adaptation in Whisper to evaluate whether transcription gains require group-specific adapters rather than pooled fine-tuning.

Presents an open-source research workbench and portable packaging format to preserve relationships between computations, artifacts, and assessments.

Presents a unified encoder that jointly models type definitions and text for cross-domain entity, relation, and event extraction.

Shows that language agents often ignore identified cross-domain dependencies, causing unintended disruptive actions in network management.

Converts half-duplex speech models into streaming full-duplex conversational models capable of simultaneous listening and speaking.

Demonstrates that the scoring format chosen for an evaluation model strongly restricts its expressive capacity and ranking reliability.

Uses language-model perceptual descriptions of complex chemical mixtures to predict perceived odor similarity from molecular composition.

The paper describes a modular software backend architecture for extracting energy domain information and answering grounded questions from web sources.

The paper introduces decision models that balance fast single-pass predictions with adaptive reasoning for uncertain cases using reinforcement learning.

The paper develops an approach to map AI semantic blind spots and evaluates the quality of distinct human ideas.

The authors propose an evolutionary framework that optimizes instance coverage and updates language model weights to discover complementary optimization heuristics.

The paper presents a taxonomy that locates agent workflow failures across specific workflow stages and agent functional capabilities.

The authors examine why language model agents fail by analyzing discrepancies between internal representations of state values and actual action choices.

The paper studies how much models contribute to multi-model teams, showing individual test scores do not correlate well with collaborative performance.

The paper analyzes the failure modes of agent memory reuse and evaluates certified memory update strategies when facts change.

The paper investigates how indexed instruction wording affects the performance of lexical rankers for robot task retrieval in manipulation benchmarks.

The paper introduces an evaluation method that separates question answerability detection from answer span extraction in radiology report question answering.

The paper introduces a dataset and benchmark to evaluate vision-language models on recognizing and answering questions about Philippine cultural heritage sites.

The paper proposes a rubric-based reinforcement learning method combining selective teacher distillation with an adaptive hand-off to standard optimization.

Introduces DeltaTTG, a method using predicted remaining token count as a progress signal to allocate advantages in multi-turn reinforcement learning.

Introduces LatentRiskBench to evaluate safety risks in enterprise agents where action safety depends on latent workflow context.

This paper presents a compact decision model designed to screen malicious agent tool skills with high recall directly from specification files.

This paper provides four auditing tests to determine whether failure-attribution judges actually perform multi-step checks or merely rely on spurious surface cues.

This paper presents an optimization method that trains models to decide when to invoke reasoning tokens, reducing unnecessary thinking overhead.

This survey analyzes how verifier noise affects reinforcement learning with verifiable rewards across 58 studies and summarizes practical mitigation techniques.

Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages.

Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication.

Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models.

HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding.

Virtual Head Attention uses two linear operations to compress key-value cache overhead without increasing training compute or sacrificing output quality.

The authors show that medical vision-language models frequently abandon correct clinical answers when prompted with fabricated medical documents.

The paper proposes predictability-aware supervised fine-tuning to resolve training and inference order mismatches in diffusion language models.

LatentRemedy uses task-conditioned latent tokens to allow safe key-value cache reuse across conflicting tasks in multi-agent systems.

This paper presents an intermediate-representation method to audit whether specific private data was used in LoRA fine-tuning of language models.

This paper shows chain-of-thought monitoring can fail even when reasoning is necessary, proposing a multi-level hierarchy to measure monitorability.

This paper presents an interpretability protocol to monitor latent reasoning dynamics and stability across multiple looped language model families.

This paper shows data ownership determines whether agents should broadcast or relay table queries, and proposes ownership-aware routing.

This paper studies how different framings of user claims induce sycophancy in large language models on factual questions.

This paper introduces a self-distillation method that searches over repeated layer execution orders to create stronger teacher guidance.

TextualRL updates agent context prompts using groupwise trajectory comparisons rather than isolated feedback runs to improve policy adaptation without parameter tuning.

Runoff eliminates option-order bias in multiple-choice model predictions by scoring items independently before running balanced candidate comparisons.

The authors demonstrate that encoding malicious instructions in agent skills bypasses scanners and causes models to execute harmful actions upon decoding.

JevGPT combines fast local evidence assessments with deliberative planning to improve autonomous research agents operating under fixed retrieval budgets.

SimInter simulates conversational recommendation users by conducting synthetic interviews between language models to enrich implicit behavioral profiles.

The authors evaluate how 8-bit and 4-bit quantization methods affect model accuracy across Bavarian, Standard German, and English benchmarks.

The authors propose a probe-based method to measure and calibrate systematic scoring biases in black-box LLM evaluation judges.

The authors trace internal representations across prompt optimization techniques to show how discrete prompts, prompt tuning, and prefixes alter model processing.

The authors reproduce and compare five language model factuality alignment methods on a standardized benchmark across identical base models and evaluation protocols.

The authors introduce a 2.5-billion-parameter bilingual Tamil-English language model trained from scratch with tool-calling capabilities.

The authors analyze readout scaling in sparse autoencoders and propose canonicalized coordinates to improve the recovery of underlying generative factors.

The authors identify a derivation gap where LLM judges reject correct tool arguments lacking direct textual references, and propose a dedicated reference model.

The paper analyzes the geometric distribution of language model output logits, identifying a consistent log-linear relationship that emerges during pretraining.

The paper presents an extended behavior tree framework that automates modular robot task planning and runtime replanning for off-site construction fabrication.

The paper analyzes implicit morals in news coverage of artificial intelligence across four global regions, identifying distinct regional variations in narrative caution.

The paper examines projector training in multimodal models and proposes a training-free initialization method using summary statistics instead of paired image-caption data.