The paper introduces the concept of Search-Time Contamination (STC), demonstrating that deep research agents can leak information from public benchmarks via web search, leading to an overestimation of…

ARES is a novel framework that systematically discovers and mitigates dual vulnerabilities in RLHF systems by simultaneously testing the core LLM and its Reward Model (RM) using structured adversarial…

DenoiseRL is a novel reinforcement learning framework that improves reasoning in large language models by optimizing directly from the failures and incorrect reasoning traces of weak models, eliminati…

The paper argues that current LLM benchmark datasets are often contaminated by being included in pretraining data, and proposes that future benchmarks must be contamination-resistant and support infer…

EchoRL proposes a lightweight module to exploit valuable learning signals from advantage-degenerated rollouts in Reinforcement Learning with Verifiable Rewards (RLVR), significantly improving LLM post…

The paper introduces TraceGuard, a detectability-aware antidistillation method that identifies and poisons 'thought anchors'—sparsely critical sentences—to degrade student model learning without makin…

The paper introduces a reinforcement learning framework, inspired by AlphaZero, to automate and improve the proof search process within the Tamarin protocol analysis tool, resulting in shorter and mor…

This paper introduces a novel framework, the Reasoning Safety Monitor, to detect and prevent logical inconsistencies and adversarial manipulations within the internal reasoning steps of large language…

The OISD framework improves language model reasoning by distilling on-policy predictive signals from the final output layer to intermediate representations, leading to substantial improvements on math…

The paper introduces Feedback Distillation, a novel training method that uses a language model's privileged feedback to provide token-level supervision, significantly improving complex reasoning tasks…

The paper introduces Agent-ToM, a Theory-of-Mind (ToM) based framework that learns to monitor autonomous LLM agents by explicitly reasoning about their hidden beliefs and intentions to detect covert m…

The paper proposes Detector Evasion Policy Optimization (DEPO), a constrained reinforcement learning method that effectively evades AI text detectors while strictly maintaining the original text's sem…

The paper introduces an automated framework demonstrating that LLM system instructions are vulnerable to encoding attacks, where structured output requests can bypass safety refusals and leak sensitiv…

This paper synthesizes over 150 scattered studies and reports to provide the first comprehensive primer on post-training reasoning data, organizing the field around data objects, utility, construction…

The paper demonstrates that using Reinforcement Learning from Verifiable Rewards (RLVR) significantly improves small language models' functional correctness in code generation, particularly when combi…

The paper proposes R-CoT, a reasoning-layer watermarking framework that embeds ownership watermarks directly into the stable reasoning path of LLMs, achieving high robustness against perturbations.

The paper introduces SONAR, a prompt sanitization framework that uses natural language inference metrics to identify and remove malicious instructions injected into LLM prompts, achieving near-zero at…

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 introduces the threat model of sequential data poisoning, demonstrating that multiple, collaborating attackers can exploit compound vulnerabilities in LLM post-training pipelines that are in…

LongTraceRL addresses long-context reasoning challenges by generating highly challenging training data and introducing a fine-grained rubric reward, significantly improving evidence-grounded reasoning…