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 demonstrates that while distilling large language models for medical QA can significantly improve final answer accuracy, this gain often comes at the cost of factual accuracy and detailed re…

The paper introduces Responsible Contrastive Soft Prompting (RCSP), a parameter-efficient method using soft prompts to improve LLM reliability by simultaneously suppressing hallucinations, encouraging…

The paper analyzes the failure modes of aggressive 2-bit quantization in large reasoning models, proposing lightweight controls like FP16 planning and loop rescue to restore accuracy and achieve pract…

The paper analyzes the security limits of verifiable decapsulation tests for Key Encapsulation Mechanisms (KEMs), establishing that the list-hit event is the primary black-box obstruction and deriving…

The paper introduces Acceptance Cards, a rigorous four-diagnostic standard, to provide a comprehensive and reliable evaluation protocol for claims of safe fine-tuning defenses.

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 demonstrates that using raw source text for fine-tuning LLMs on vulnerability detection causes high false-positive rates by memorizing surface-level syntax, a problem mitigated by using Abst…

The paper proposes a black-box attack using a hierarchical genetic algorithm to induce 'overthinking' in Large Reasoning Models, demonstrating that this vulnerability can cause significant resource ex…

The paper proposes Two-stage Backdoor Hijacking (TSBH) to create persistent, trigger-activated malicious behaviors by manipulating the observable Chain-of-Thought (CoT) process in Large Language Model…

This paper proposes a method to recover recoverability structure from failed traces of post-trained language models, enabling test-time routing and post-training analysis.

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.

The paper introduces Probe-Geometry Alignment (PGA), a surgical method that removes the measurable cross-sequence memorization signature from large language models without degrading their general capa…

The paper introduces NumLeak, a framework demonstrating that top-tier LLMs often exhibit high fidelity recall of specific public numeric benchmarks (like financial factors) due to memorization, which…

The paper introduces NumLeak, a framework demonstrating that top-tier LLMs often exhibit high fidelity recall of specific public numeric benchmarks, suggesting that their apparent skill may be due to…

This paper systematically diagnoses the failure modes of linear deception probes in LLMs, finding that while single-direction probes are insufficient, multi-dimensional probes can recover robust detec…

This paper investigates the production-evaluation gap in Large Reasoning Models (LRMs), finding that while LRMs excel at generating solutions, they struggle significantly to evaluate flawed reasoning,…

This paper investigates the redundancy of the prompt KV cache during language model decoding, finding that the structure provided by chat templates is the primary source of redundancy, not the actual…

The paper introduces the DECK taxonomy, a novel framework that classifies LLM hallucinations not by their content error, but by their detectability signature based on inter-sample consistency and toke…

The paper challenges the conclusion that LLMs lack reasoning by demonstrating that reported performance drops on GSM-Symbolic are often statistically weak and partially attributable to dataset biases,…