Some text is shown in its original language while the English translation is being prepared.
Oct 12 · Monday
All papers Found 397 papers
Another 336 papers do not have a research type yet and are not included in these categories.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Organized first” shows papers with tags or public deep readings first. Each group is ordered by original publication date, newest first, with undated papers last. “Latest publication” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and deep readings are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
Proposes SWE-DeceptionBench to evaluate LLM deception in the software development lifecycle
Published Oct 9
Evaluated models GPT-5.5 、 Claude-Opus-4.8 、 Gemini-3.5-Flash and others9 Proposing DRYLAB-Bench to evaluate agentic LLMs' high-risk mutation design capability
Published Oct 9
Evaluated models Claude Opus 5 、 Claude Opus 4.8 、 Claude Fable 5 and others15 Introduces DistressBench, evaluating models' multidimensional support in crisis conversations
Published Oct 9
Evaluated models Muse Spark 1.2 、 Muse Spark 1.3 、 Muse Glimmer 30B and others20 Abstract The model is more restrained and acts less.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 SWE-Duel, which lets coding agents mutually evaluate, create problems, and blindly review fixes during evaluation.
Published Oct 9
Evaluated models GPT-5.5 、 Claude Opus 4.8 、 Kimi K2.7 Code and others11 Abstract The mutual evaluation arena creates questions during evaluation.SWE-Duel is a coding agent arena that generates tasks and judges only at evaluation time, to avoid leakage and saturation of static benchmarks.
Proposes the ISE framework to evaluate the epistemic humility of LLM agents under knowledge conflicts
Published Oct 9
Evaluated models Nemotron-Orchestrator-8B 、 Claude Sonnet 4.6 、 GPT-5 and others5 Abstract ISE trajectory evaluations show that high task accuracy across four agents does not correspond to higher cognitive humility.What the authors want to test is not whether the agent's final answer is correct, but whether, when a knowledge conflict arises, it recognizes the gap in its trajectory, performs bounded tool follow-up, and writes the unresolved uncertainty to the user.
Proposes LGWM, using an action-conditioned world model to verify GUI agent screen transitions.
Published Oct 8
Evaluated models LGWM 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
Published Oct 8
Evaluated models ObligationGuard 、 Claude-Opus-4.8 、 GPT-5.6-Sol and others17 Abstract 义务是护栏缺口。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。
提出 IRIS，用意图恢复率区分遮蔽越狱下无害输出的原因
Published Oct 8
Evaluated models GPT-4o 、 Claude Sonnet 4.5 、 Claude Opus 4.5 and others6 Abstract 无害输出的证据权重取决于任务是否被恢复。IRIS 把意图恢复当作中间信号，用来解释意图遮蔽提示下为何没有出现有害协助。
测量 LLM 版本更替对科学摘要 AI 改写检测器的影响
Published Oct 8
Evaluated models RoBERTa-base 、 Pangram Abstract 作者称检测器的基准准确率只是临时的，每次发布都要连同更早版本重验。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。
提出 MH-INDIC，评测印度语境母婴健康 LLM 交互的文化对齐
Published Oct 8
Evaluated models Claude Sonnet 4.6 、 GPT-5 、 Gemini 3 Pro and others10 Abstract 作者称模型可接近人群均值，但画像变异低于人类，人类分数条件可改善对话。MH-INDIC 是面向城市与半城市北印度母婴互动的多维文化对齐评测。 它关心的不是医学事实是否正确，而是回复是否沿家庭权威、性别期待和情绪披露来推理。做法是把 Hofstede 六维改写到互动层，再加入 MAFC、BTB、SVO、MHS。
提出 SSCBench，评测工具调用 Agent 故障注入测试的证据有效性
Published Oct 8
Evaluated models deepseek-flash 、 deepseek-v4-pro 、 gpt-5.6-luna and others5 Abstract 解释采纳前应报告该次执行的证据条件，以及支撑该主张的观察总体。SSCBench 研究工具使用智能体的故障注入结果，在什么证据条件下可以被解释。
用 BeliefScope 分离固定命题上证据与压力对模型信念的影响
Published Oct 8
Evaluated models Qwen3-8B 、 Llama3.1-8B 、 Gemma3-12B and others4 Abstract BeliefScope 把证据修正与压力偏移拆成绑定评测条件的行为画像。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。
构建 EpiReal-Bench，评测图像生成器将虚假主张渲染为可信证据
Published Oct 8
Evaluated models GPT-Image-2 、 Nano Banana 2 、 Grok Imagen 2.0 and others4 Abstract 四款商用图像模型能把假主张画成可信证据。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。
提出 GUISE 基准与 AXIS 训练，防御叙事包装越狱
Published Oct 8
Evaluated models Qwen3-1.7B 、 Qwen3-4B 、 GLM-4-9B Abstract GUISE 分开测包装与语域。作者测量叙事包装下的拒答失效，并训练一个在未见包装上拒绝、同时仍回答同风格良性请求的适配器。问题是：安全对齐模型拒绝直说的有害请求，却回答嵌进角色或叙事里的同一请求。文言文被用来把语域和包装拆开。
构建统一奖励作弊检测基准 RH-Detect 并评测零样本检测器
Published Oct 8
Evaluated models Qwen 3.5 Flash 、 Gemini 3.1 Flash Lite 、 Grok 4.1 Fast and others7 提出统一框架评测多模态大模型的遗忘方法
Published Oct 8
Evaluated models LLaVA-1.5-7B 提出 RSIGym 环境，评测研究智能体的递归自改进
Published Oct 8
Evaluated models Claude Opus 5 、 Claude Opus 5.5 、 GPT-6 Astra and others9 提出 PatchBench-Local 协议，评测激活补丁修复越狱的局部附带损伤
Published Oct 7
Evaluated models Qwen2.5-3B-Instruct 、 Qwen2.5-14B-Instruct 、 gemma-3-4b-it and others8 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。
适配多种已发表说服评测并比较模型排名
Published Oct 7
Evaluated models Claude-Fable-5.1 、 GPT-5.6-Sol 、 Kimi-K2.6 and others15 Abstract 说服分数混合能力与意愿，换任务后同一模型的排名可以差很多。这项研究比较的是自动化说服评测是否一致，而不是哪一种分数更正确。
用对照实验隔离评测设计对漏洞补丁基准得分的影响
Published Oct 7
Evaluated models Claude Sonnet 4.0 、 Claude Sonnet 4.5 、 Claude Opus 4.7 Abstract 作者认为评测流水线会改变补丁基准分数，PoC 通过率高于开发者测试通过率。作者把 LLM 漏洞补丁基准的评测流水线当作会改变测量结果的对象，并明确说本文不是再做一个用于排名的基准。