All papers

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.

Proposes Diff Diffusion, a diffusion-based model that predicts localized code edit sequences rather than regenerating entire source files.

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.

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 formulate a meta-game framework to dynamically route and select language models during strategic multi-agent interactions.

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 authors propose evaluating autoformalization faithfulness by aligning phrase-level correspondences between informal mathematics and formal code under contextual interpretations.

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

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.

The paper proposes a framework that uses reasoning graphs to locate claim-level disagreements across sampled traces and correct multi-step language model reasoning.

The paper introduces a benchmark to evaluate how typed decision models handle theory-of-mind reasoning tasks compared to standard generative language models.

The paper evaluates whether self-distillation using privileged information improves reasoning capabilities or reflects overfitting to training tasks.

This paper introduces a metric to estimate the compatibility of pretrained models before running costly alignment or checkpoint merging.

This technical report presents an updated summary-token attention architecture to optimize cache compression and computation across model layers.

This paper investigates whether cloning a familiar voice systematically transfers personality impressions to an embodied virtual character.

The authors propose a training-free method constructing ontology-grounded semantic identifiers to preserve hierarchy in medical concept embeddings.

This paper investigates attention heads responsible for contextual entrainment and evaluates whether their ablation transfers across different semantic relations.

The paper introduces a debugging training algorithm that rewards localized semantic code edits alongside unit test execution.

The paper introduces an auditing framework that identifies failure steps in multi-agent systems by evaluating how information is utilized across handoffs.

The paper presents a runtime system that structures long-horizon agent memory around reusable progress objects rather than raw interaction history.

The paper analyzes how language model aggregation gains divide into candidate additions and context effects under different baseline comparison setups.

The paper presents a framework and benchmark to evaluate whether self-evolving robotic agents genuinely learn reusable skills across tasks.

The paper formalizes a framework and theoretical generalization bounds for evolving an agent's software harness without updating model weights.

The paper examines back-translation using monolingual data as an alternative to continued pretraining for low-resource translation with language models.

The paper introduces a decoding framework that dynamically selects sampling temperature based on geometric representations of conversational scenarios.

The paper proposes an agent architecture that dynamically switches between latent reasoning and text interaction while maintaining external tool access.

The paper introduces a training-free contrastive decoding method where a small model guides a larger model toward shorter reasoning chains.

The paper presents a dataset and framework with dense progressive rewards to train language agents on long-horizon tasks.

The paper introduces action chunking for agentic robotics to execute pre-planned primitive routines without redundant model reasoning at each step.

This paper presents a mid-training method, corpus, and benchmark to help language models understand and reason about time series data.

This paper introduces an open-source infrastructure and benchmark for multi-fidelity AI-assisted photonic simulation and inverse design.

Summary pending (next check).

This paper provides design guidelines and a selection tool to help build augmented reality systems for industrial assembly support.

The paper presents an 8-bit weight and activation post-training quantization technique that migrates activation outliers to weights.

The paper introduces a post-training mixed-precision quantization method using GPU-compatible 2D weight tiles for efficient inference.

The paper presents an open-source multimodal language model combining a small language backbone with vision encoders for efficient inference.