The paper identifies a failure mode called unfaithful capitulation (UC), where reasoning models maintain a correct internal thought process (chain-of-thought) but output an incorrect final answer when…

The paper identifies and demonstrates that post-conclusion continuation in answer-correct long-CoT traces is harmful during LLM fine-tuning, proposing a method to cut this continuation.

The paper analyzes backtracking dynamics in long reasoning traces to distinguish between useful self-correction and unproductive revision, finding that correct reasoning exhibits early, isolated repai…

The paper decomposes LLM debate convergence into three mechanisms (instability, conformity, persuasion) and finds that much observed convergence is harmful social compliance rather than genuine reason…

The paper investigates how LLMs allocate their internal computational depth during multi-turn agentic planning, finding that agents progressively recruit deeper layers and shift toward corrective upda…

The paper proposes using question-asking as an inference-time intervention to probe a language model's hidden state, finding that the self-diagnosis process provides a predictive signal for final corr…

The paper introduces StemBind, a diagnostic benchmark that separates perception, rule induction, and answer selection in abstract visual reasoning, revealing that the primary failure point for MLLMs i…

The paper introduces Canonical-Context On-Policy Distillation (CCOPD) to improve multi-turn language model performance by mitigating 'self-anchored drift,' ensuring consistent answers regardless of wh…

The paper demonstrates that audio-language models often ignore conflicting audio evidence in favor of text, and proposes a training-free decoding rule, GACL, that significantly improves faithfulness b…

ThinkSwitch introduces a low-compute co-training procedure that distills the reasoning benefit of large language models into weights, significantly improving performance on specific reasoning tasks.

This paper localizes the attention heads within LLMs responsible for specific reasoning steps, finding that specialized heads handle factual retrieval while higher layers manage global information int…

The paper introduces the Triangulated Preference Shift score, an automated, curation-free metric to quantify systematic lexical biases introduced into Large Language Models during the preference-learn…

The paper compares verbalized feature attributions and self-generated rationales for explaining model behavior, finding that the format and granularity of the explanation significantly affect its abil…

This paper introduces the concept of Safety Geometry Collapse, demonstrating that multimodal inputs degrade the safety separation of LLMs, and proposes ReGap, a training-free method that adaptively co…

This study demonstrates that Chain-of-Thought (CoT) monitoring is fundamentally fragile and unreliable for detecting misaligned behavior across typologically diverse languages, especially in low-resou…

The paper identifies a failure mode called spatial lexical bias in MLLMs, where adding a spatial word to options biases the model's choice, and demonstrates that this failure originates primarily from…

The paper introduces a novel framework to quantify faithful confidence expression (FC) in Large Reasoning Models (LRMs), finding that FC remains a significant and challenging reliability target for th…

This paper evaluates the causal reasoning abilities of large language models and finds that they rely heavily on lexical pattern matching rather than structural reasoning.