All papers

Gates reinforcement learning policy-shift distillation using intermediate checkpoints to ensure student models only learn consistent teacher shifts.

Introduces a data augmentation method that breaks spurious contextual correlations causing emergent misalignment during domain finetuning.

Identifies a specific vision transformer attention head that spuriously attends to medical image corner text instead of lung pathology.

Proposes an availability framework to identify and resolve within-event timing mismatches in earnings-call financial NLP benchmarks.

Improves reinforcement learning credit assignment in multimodal reasoning using pairwise token contrast and adaptive trajectory confidence.

Frames diffusion language models as open quantum systems to enable text generation conditioned on superpositions of prompts and thoughts.

Shows across 200 models that proposed intrinsic tokenizer metrics do not reliably predict downstream language model performance.

Reviews and formalizes data selection, weighting, and scheduling policies across the pretraining and post-training lifecycle of language models.

Compares abstract-only versus full-text inputs and few-shot versus fine-tuning approaches for generating biomedical lay summaries.

Introduces a system that generates and compares alternative scene drafts to guide full dramatic script adaptations from prose.

Applies adversarial training and distillation to remove unwanted gender bias in audio-enabled language models.

Demonstrates that machine unlearning often fails to prevent targeted information recovery across alternative multilingual access paths.

Analyzes how specific nouns and semantic clusters relate to answer selection and user engagement in logistics question-answering forums.

Proposes a compressed representation using Gaussian trellises to quantize conditional memory lookup tables down to two bits.

Introduces a non-autoregressive transformer model that dynamically decides tag generation order at inference time for low-resource sequence labeling.

Distills multiple speech attributes from four teacher models into a single shared speech encoder.

Proposes a reinforcement learning credit-assignment method for document parsing models to align rewards with structured page elements.

Introduces a benchmark that structures multi-pull-request code changes into dependency graphs to evaluate long-horizon software engineering agents.

Evaluates whether internal hidden-state probes can identify failure modes and pinpoint failing steps in multi-step reasoning traces.

Measures demographic bias related to age, gender, and race based on first names across local language models.

Applies a dual language model consensus pipeline to extract clinical phenotypes from electronic health records for patients with Rett syndrome.

Studies how using UTF-16 byte order improves boundary prediction and representation efficiency in multilingual byte-level language models.

Introduces a compression-based patching strategy for byte-level language models that allocates compute where upcoming bytes are unlikely to repeat.

Specializes a translation model to a specific indigenous Nawatl variety using a curated dataset and finite-state morphological analysis.

Develops small multimodal models that output structured choices using masked language modeling to reduce inference latency in agent tasks.

Introduces a benchmark to evaluate whether text-to-SQL systems can detect flawed query premises and abstain from generating invalid SQL.

Develops a two-stage training approach combining evidence reasoning and reinforcement learning to improve pragmatic social reasoning in conversational models.

Introduces lexical reference frames to interpret intermediate hidden states in language models using the model's own remaining layers.

Proposes an event-centric framework to align audio and video features when generating detailed video descriptions with language models.

Analyzes the overuse and contextual appropriateness of a common corrective phrasing pattern in Mandarin language model text.

Presents an adversarial training approach designed to improve language model honesty across diverse benchmarks without degrading core capabilities.

Uses diffusion language models to perform multi-target few-shot infilling across shared code context guided by demonstration examples.

Evaluates the ability of language models to predict state changes in discrete spatial grids across sequential operations.

Presents an evaluation framework combining simulated mental health patients and expert-developed rubrics to assess conversational language models.

Proposes a visual verifier that decomposes complex prompts into sub-questions and assigns continuous scores to improve multimodal evaluation.

The paper identifies an internal activation direction in spoken language models to control and transfer conversational turn-taking styles.

The paper introduces an encoder architecture that processes input once to answer multiple structured classification queries simultaneously.

The paper presents a framework and dataset for learning to detect and repair errors during multi-agent system execution.

The paper presents an agent that expands its domain-specific language for exact answering without external rewards or judges.

The paper presents open-weight language models adapted for the Wolof language using instruction tuning, preference alignment, and model merging.

The paper proposes a method to translate reverse reasoning paths into verified forward chains of thought for distillation.

The paper introduces a confidence-calibrated controller to optimize inference compute and test-time expansion decisions in reasoning models.

The paper investigates performance vulnerabilities in self-improving agents when their co-evolved skill banks undergo post-deployment drift.

The paper analyzes how vision-language models process negation by comparing model attention mechanisms with human eye-tracking data.

The paper analyzes why reinforcement learning with majority-vote rewards provides effective training gradients despite noisy individual labels.

The paper proposes a non-elitist evolutionary optimization framework that retains and iteratively refines low-reward agent skills using execution feedback.

The paper presents a decision model that performs deeper reasoning by looping its layers instead of generating output tokens.

The paper investigates when language models learn capabilities by tracking unembedding matrix dynamics during pretraining using parameter-trajectory crosscoding.

Develops an open agent environment for evaluating how autonomously language agents can conduct post-training machine learning experiments.

Estimates language model ensemble member reliability from unlabeled agreement patterns using an expectation-maximization jury formulation.

Uses reinforcement learning to train an agent module that selects, sizes, and orders retrieved memories for downstream tasks.

Measures tool-agent performance failures under controlled harness limits such as completion caps and tool menu visibility.

Analyzes trade-offs between aggressive and conservative planning horizons in language models operating under partial observability.

Introduces a dataset benchmarking human-agent interaction and collaboration across long-horizon professional workflows.

Proposes a training-free structural lexical retriever to select relevant agent cards without overloading model context windows.

Examines failure modes in label-free evaluation where models reach high consistency across perturbations while generating identical incorrect answers.

Presents a framework that decouples speech generation and synthesis token spaces to speed up spoken dialogue model inference.

Develops a structured explainable reasoning architecture to extract sensitive self-disclosures from cybersecurity interviews using local small language models.

Analyzes working memory error patterns in language models to show that failures stem from readout interference rather than target retrieval.

Proposes an ensemble framework that weights judge confidence and masks disagreements when training models with rubrics as rewards.