Another 240papers do not have category tags yet and are not included in filtered results.
Proposes SWE-DeceptionBench to evaluate LLM deception in the software development lifecycle
Auditing training labels for label-only backdoors, comparing spaces and statistics, and judging by retraining
The authors ask: can a defender without trusted data decide which training labels to keep, and how much clean data this destroys, and what the final model accuracy is.
Introduces DistressBench, evaluating models' multidimensional support in crisis conversations
DistressBench uses SSH dialogues written and revised by clinical personnel, and evaluates how much necessary care general chat models deliver according to weighted atomic criteria, rather than whether they refuse. When users disclose suicidal ideation or self-harm to a chat model, the authors consider the expected behavior to be recognizing the crisis, calibrating empathy, de-escalating, and referring to human help. Refusal-based evaluations cannot detect failing to act after recognizing the crisis.
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 Task Bank self-replay to mitigate held-out forgetting in continual instruction tuning
The authors study how language models, when learning new tasks, can use their own knowledge for broader self-preservation: given an incomplete list and forgetting that extends beyond the list, can the model generate replay and help choose targets.
Analyzing why optimizing teaching adaptation metrics collapses tutoring choices and diverges from expert judgment
After turning teaching adaptation metrics into a fine-tuning objective, the metrics rise while scores from trained educators decline.
Using fear activation steering to test the hazard judgment criteria of vision-language models
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。
用 BeliefScope 分离固定命题上证据与压力对模型信念的影响
BeliefScope 是一个黑盒行为诊断框架，用来把固定命题上的回答变化拆成证据驱动的修正，和不加新事实的方向性压力。信念修正评测能看到信息基础变化后的更新，谄媚评测能看到用户影响下的顺从或漂移。作者认为，当两种输入都偏向同一结论时，表面偏移本身不能区分来源。
提出 PatchBench-Local 协议，评测激活补丁修复越狱的局部附带损伤
Abstract论文提出 PatchBench 与局部评估协议，揭示了激活补丁在全局能力无损下仍存局部良性退化。
提出热启动 Bandit 的部署前有界离线注入攻击
比较连续后训练阶段如何改变语言模型的说服力
作者用同一人类协议，把 LLM 说服力拆成阴谋数据上的失调 SFT、说服数据上的 SFT，以及其后的 IPO。
提出 OrthoPurify，用正交投影清除视觉语言模型适配器中的劫持方向
OrthoPurify 是部署前的 LVLM 后门净化方法，用一次正交投影去掉被劫持的权重更新方向。
提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门
Abstract论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。
提出 SOUL 稀疏特征策略遗忘，缓解 VLA 状态幻觉
SOUL用稀疏特征引导的策略 unlearning 处理 VLA 的 state hallucination。
测量委托收费激励下模型策略性虚报任务信心
Abstract论文建立人机委托信誉博弈框架，揭示模型策略性虚报置信度且主要造成信息破坏，作者指出校准受制于报告意愿。
提出 CARE，离线认证 VLA 加速配置的回合级失败风险并选出最快合格候选
CARE 在部署前从加速配置中选出最快、且加速诱发失败风险能被有限样本认证为低于用户预算的候选。证据不足就保留全计算参考策略。
提出 World Models' Last Exam 评测视频物理一致性
提出知行差距面板，测量压力下模型是否违背自身判断
本文评估了开源语言模型作为智能体行动时违背自身道德判断的知行差距及其决定因素。
分析记忆与泛化如何决定机器遗忘的 retain 损伤
作者考察学习方式如何左右 unlearning 的 retain damage：训练越依赖跨样本共享的解法，在匹配的遗忘程度上，retain 上的损失越大。