Another 233papers do not have category tags yet and are not included in filtered results.
Proposing SWE-Critix, a post-trained general verifier for evaluating coding agent trajectories
Proposes DREAM, which selectively weakens agent-poisoned memories while preserving benign memories.
Show that reward hacking can propagate through code, memory, and plans in multi-agent coding systems
Reward hacking induced by coding RL can leave the source agent and enter subsequent collaboration and system coordination.
Proposes the Cancerous backdoor, which progressively corrupts reasoning using invisible stage tokens.
Cancerous Backdoor Attack changes LLM backdoors from one-time activation to multi-stage reasoning corruption that accumulates progressively with triggers.
Proposes PASAC and a five-layer boundary guarantee stack, turning agent evaluation overstepping into verifiable whole-system guarantees.
Using FIBS to train white-box probes to monitor agent sabotage behavior and undisclosed deception
Uses premise permutation and hidden-layer tracing to audit whether legal chain-of-thought truly relies on the cited legal provisions.
This study conducts an empirical audit of the causal truthfulness of legal citations in chain-of-thought reasoning of large language models in the legal domain.
Proposing an option channel attack that renames option labels so that agent guardrails allow prohibited operations
AbstractThe authors' evaluation indicates that typed models should not serve as final safety arbiters; they are susceptible to off-policy text and option name manipulation, and confidence filtering cannot resist attacks.
Auditing LLM judge reliability and defining trustworthy verdict rate and accuracy upper bound
The authors audit whether LLM-as-a-Judge decisions can still be treated as actionable ground truth under repeated queries, swapped candidate order, and changed prompt formats.
Proposes LGWM, using an action-conditioned world model to verify GUI agent screen transitions.
提出义务识别基准 ObligationBench，并训练护栏识别 Agent 未履行的安全义务
作者认为编码智能体的安全不能只靠拦截禁止动作，还要找出终止时仍未做的安全关键动作，并将其称为义务。
度量关闭思考指令下终答前是否仍有可见显式推理
作者把 no-thinking 定义为可测量的回答行为，而不是从思考开关、特殊标记或输出长度推断。
测量 LLM 版本更替对科学摘要 AI 改写检测器的影响
作者量化的是：科学摘要的 AI 改写筛查，在 LLM 版本不断更换、而人类摘要误标必须少见时，会怎样变。问题来自期刊和会议已经按分数阈值筛查投稿，但基准多停在评估时可得的版本上。
构建 Workerville 基准，测量组织条件如何改变 Agent 的安全行为
Workerville 用组织行为测量智能体安全如何随监督、同伴和记忆条件变化。。要处理的缺口是：持续运行的 LLM 智能体同时接收用户指令、同伴消息和长期记忆，但既有安全评测多把这些通道拆开，没有统一的定量测量。
演化可检查评分器并与自改进 Agent 共同演化，抑制奖励作弊
开放任务没有验证器时，被演化的是评分器，而不是再补一条手写 rubric。
提出 RL-ARC，用推理置信度在强化学习中校准推理模型
RL-ARC 是校准感知的强化学习框架，用推理过程的口头置信度调节答案置信度。
提出包络采样，用少量真值标注重标定 LLM 评委以缓解奖励作弊
作者研究后训练前如何选择真值标注的查询分布，使重标定后的廉价代理更适合被优化。真奖励昂贵时，实践用 LLM judge 代替人类。若失准只发生在基座很少生成的输出上，on-policy 标注修不到这些点，随后的 KL 正则优化会把质量集中到被高估的输出。
构建 EpiReal-Bench，评测图像生成器将虚假主张渲染为可信证据
EpiReal-Bench 用来量商用图像生成器会不会把虚假现实主张画成可信视觉证据。
构建统一奖励作弊检测基准 RH-Detect 并评测零样本检测器
提出路由 TopK SAE，把语音编码器稀疏码拆成语言与副语言通路
TopK SAE 的稀疏码被分成语言通路与副语言通路，用来分开冻结语音编码器里同时存在的音素信息和说话人、情感、韵律。只重构不能得到作者选定的这一分解。语言通路由音素识别监督，副语言通路由说话人识别、情感分类和韵律预测监督，作用在对方通路上的梯度反转对抗用来压低交叉因子。