[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-refusal-abliteration-generation-audit":3,"topics-all":38,"news-related-2e7a2be3-762e-46eb-b582-af299c4ec173":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"2e7a2be3-762e-46eb-b582-af299c4ec173","模型越新，拒绝越难拆？76页审计：并不均匀","开源模型越新，拒绝行为越难从权重里拆掉？单作者76页预印本在五组新旧模型对上用Heretic 2.0跑五种方法，972条留出有害提示由WildGuard判分：方向类编辑后新模型多留-1.2到+36.9个百分点拒绝，只有OLMo组两种方差模型下都显著，ARA在Gemma组反而更易拆。结论取决于模型对、方法与预算。","2024 年起，把开源大模型的「拒绝」行为从公开权重里拆掉，就不再需要端到端微调：几百条对比提示词、一张 GPU、改几个权重矩阵，模型就不再拒绝有害请求。全自动工具公开可用，消融后的模型直接挂上 Hugging Face。真正悬着的问题是：新模型是不是越来越难拆？ARA 方法的作者在 pull request 里断言方向类消融在新模型上越来越失灵——但那只是单个来源的传闻级证据。一篇 11 月 13 日提交的单作者 arXiv 预印本（2611.17846，76 页、34 张表，尚未同行评审）用五组新旧模型对给出了 held-out 判分的答案：并不均匀。\n\n## 五组对照，一条流水线\n\n研究选取五组升级对：Qwen2.5-32B→Qwen3.8-27B、Gemma-2-27B→Gemma-4-31B、OLMo-2-32B→Olmo-3.1-32B、Mistral-Small-2409→Mistral-Small-3.2，外加小参数量组 Qwen2.5-7B→Qwen3.5-4B。每组在 Heretic 2.0 里被三到四种方向类方法（DiM、MPOA、SOM，小对加 J-space）和优化类 ARA 编辑——ARA 现在是该工具的默认方法。所有方法共享同一套提示词、100 次 TPE 搜索预算、首 token KL 散度 0.1 的上限和同一条选择规则。判分用 972 条未参与搜索的留出有害提示（JailbreakBench、XSTest、StrongREJECT、HarmBench，加作者自编的 200 条「代码当武器」题库），由 WildGuard 自动判定是否拒绝。「变难」的定义很朴素：新模型编辑后仍拒绝的比例更高。\n\n## 五组对，五种答案\n\n方向类编辑后，新模型多留出的拒绝从 -1.2 个百分点（Mistral Small，反而更容易）到 +36.9 个百分点（OLMo）。把逐次运行方差算进来，两种方差模型下都显著排除零的只有 OLMo 一组；Qwen（+10.7）、Gemma（+4.7）和小对（+17.3）的区间都跨零。ARA 的图景同样分裂：Qwen 组 +34.4，但加入运行方差后区间是 -5.9 到 +53.1，论文自己标注为不确定；Gemma 组却是 -26.0——优化类方法在新 Gemma 上反而更好拆。合并四组的方向类效应是 +12.8，区间 [-13.9, +39.5]；ARA 合并 +1.3 [-38.3, +40.8]。组间异质性 I² 从 20% 到 89% 不等，取决于尺度与方差模型。\n\n更扎眼的是脆弱性。Gemma 的方向类增加完全来自作者自编的代码题库（该题库上 +49.3，公开题库上 -6.8，方向反转）；小对换一个随机种子，ARA 的代际效应从 +14.2 直接归零（种子改变留出拒绝的中位幅度 9.4 个百分点、至多 25.2）；Gemma-4-31B 的 DiM 干脆编辑失败——搜索在 30 倍 KL 预算内无法把 dev 拒绝压到 98% 以下，选出的模型拒绝率不低于原模型。同一张表里，Qwen、Mistral 和小对的原模型本来就比前代更爱拒绝（+13.1、+17.9、+16.4），Gemma 和 OLMo 没有——要不要扣除这条基线，又会改变部分结论的读法。\n\n## 顺带捅破的两层窗户纸\n\n其一，工具优化的信号并不可靠。Heretic 用 33 个关键词在 100 token 窗口计拒绝，与 WildGuard 判分平均只差 0.5 个百分点，但逐模型平均相差 7.0 个百分点、两个方向都有，Cohen's κ 从原模型的 0.73 跌到消融后的 0.39。其二，KL 上限不约束能力损失：42 个主模型的能力变化在 -4.2 到 +0.9 个百分点之间，与选中试验的首 token KL 无关（Spearman ρ=-0.03），与全文 teacher-forced 散度相关（ρ=-0.48），两处最大损失都发生在预算之内。论文还提醒，结论只覆盖非思考模式的贪心解码；而八个社区公开消融版里，四个比本研究选出的模型拆得更干净（Qwen3.8-27B 上 1.4% 对 14.0%），代价是散度更大——预算内的残余拒绝是上界，不是下界。\n\n## 所以呢\n\n这篇论文对「代际问题」的回答是：取决于模型对、方法、预算和度量，不取决于代际本身。只审 OLMo 会得出「新模型抗拆」，只审 Mistral Small 会得出相反结论。作者给出的最小审计协议值得抄下来：至少一种方向类加一种优化类方法、每格两个以上种子、多来源留出提示、至少两个自动判分器、多个散度预算，外加对判分歧义样本的人工抽检。成本并不夸张——单张 H200 上方向类搜索中位 18 分钟、ARA 81 分钟，31 次计时搜索共 18.1 小时；加一个种子近乎翻倍，但不加种子就是在赌。对开放权重安全的审计而言，「哪一代更安全」是个错误的问题，能回答的只有「哪一对、哪种方法、多大预算、用什么测」。论文原文见 https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17846 。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17846","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"307672dd-f41a-4c3c-aaaf-0649d0857de2","en","Newer LLMs Harder to Abliterate? A 76-Page Audit Says Not Uniformly","A 76-page audit of five model pairs: refusal removal got harder only in OLMo, easier in Gemma under ARA; verdicts hinge on pair, method and budget.","Since 2024, stripping the refusal behavior out of an open-weight LLM has required no end-to-end fine-tuning: a few hundred contrasting prompts, a single GPU, and edits to a few weight matrices make the model stop refusing harmful requests. Fully automated tools are public, and abliterated models are posted on Hugging Face. The open question is whether newer releases are getting harder to unlock. The developer of Arbitrary-Rank Ablation (ARA) asserted in a pull request that direction-based abliteration increasingly fails on recent models - anecdote-grade evidence from one source. A single-author preprint submitted on 13 November (arXiv:2611.17846, 76 pages, 34 tables, not yet peer-reviewed) ran five older-newer release pairs through a held-out, judged audit. The answer: not uniformly.\n\n## Five pairs, one pipeline\n\nThe study covers Qwen2.5-32B to Qwen3.8-27B, Gemma-2-27B to Gemma-4-31B, OLMo-2-32B to Olmo-3.1-32B, Mistral-Small-2409 to Mistral-Small-3.2, plus a small pair, Qwen2.5-7B to Qwen3.5-4B. Each pair was edited in Heretic 2.0 by three or four direction-based methods (DiM, MPOA, SOM, with J-space on the small pair) and by ARA, now the tool's default. All methods shared the same prompts, a 100-trial TPE search budget, a first-token KL divergence cap of 0.1, and one selection rule. Scoring used 972 held-out harmful prompts kept out of search (JailbreakBench, XSTest, StrongREJECT, HarmBench, plus an author-compiled 200-prompt code-as-weapon bank), judged as refusal or not by WildGuard. 'Harder' has a plain definition: the newer model refuses more of them after the edit.\n\n## Five pairs, five answers\n\nAfter direction-based removal, the extra refusal retained by the newer model ranged from -1.2 pp (Mistral Small, actually easier) to +36.9 pp (OLMo). With run-to-run variance added, OLMo was the only pair whose interval excluded zero under both variance models; Qwen (+10.7), Gemma (+4.7) and the small pair (+17.3) all crossed zero. ARA splits the same way: +34.4 in the Qwen pair, but its run-to-run interval (-5.9 to +53.1) is labelled indeterminate in the paper itself, while Gemma came in at -26.0 - the optimisation-based edit worked better on the newer model. Pooled over the four main pairs, the direction-based effect was +12.8 [-13.9, +39.5] and ARA +1.3 [-38.3, +40.8]. Between-pair heterogeneity ran from I2 of 20% to 89% depending on scale and variance model.\n\nThe fragility is the story. Gemma's direction-based increase came entirely from the author-compiled code bank (+49.3 there, -6.8 on public sets - the sign flips). On the small pair, a second seed erased the ARA generation effect (+14.2 to 0.0; seed changes shifted held-out refusal by a median 9.4 pp and up to 25.2). The Gemma-4-31B DiM edit failed outright: no trial cut dev refusal below 98% within 30 times the KL cap, and the selected model refused no less than its own original. Meanwhile the newer Qwen, Mistral and small-pair originals already refused more than their predecessors (+13.1, +17.9, +16.4), while Gemma and OLMo did not - whether to subtract that baseline changes how several verdicts read.\n\n## Two uncomfortable by-products\n\nFirst, the signal these tools optimise is shaky. Heretic counts refusals with 33 keywords in a 100-token window; that matched WildGuard within 0.5 pp on average but differed by 7.0 pp per model, in both directions, with Cohen's kappa falling from 0.73 on originals to 0.39 after ablation. Second, the KL cap does not bound capability loss: across the 42 primary models, capability change ranged from -4.2 to +0.9 pp, uncorrelated with the selected first-token KL (Spearman rho = -0.03) but correlated with teacher-forced divergence (rho = -0.48), and the two largest losses happened inside the budget. The authors also note the conclusions cover greedy decoding in non-thinking mode only, and that four of eight public community abliterations cut refusal further than the study's own selections (1.4% vs 14.0% on Qwen3.8-27B) at larger divergence - residual refusal under the budget is an upper bound, not a floor.\n\n## So what\n\nThe paper's answer to the generational question: it depends on the release pair, the method, the budget and the measure - not on generation alone. An audit of the OLMo pair alone would conclude that newer models resist removal; one of the Mistral Small pair concludes the opposite. The proposed minimal audit protocol is worth copying: at least one direction-based and one optimisation-based method, two or more seeds per cell, held-out prompts from more than one source, at least two automatic judges, several divergence budgets, and human labels on a sample of judge disagreements. The cost is modest - a median 18 minutes per direction-based search on one H200, 81 minutes for ARA, 18.1 hours for the 31 timed main-study searches - and doubling it for a second seed beats gambling on one. For open-weight safety, 'which generation is safer' is the wrong question; the answerable one is 'which pair, which method, which budget, measured how'. Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17846 .\n","refusal-abliteration-generation-audit","2026-11-16T17:45:00Z","2026-11-16T17:17:41.846796Z","2026-11-16T17:17:41.846807Z",true,"agent",43,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"4977d1f4-c8f7-480c-aaa1-ec01d69f44d8","8B 拿 SFT+GRPO 打 685B MoE:KaliBench 把 LLM 网络安全工具调用拆到命令行级","kalibench-cybersecurity-cli-runtime-rewards","2026-10-03T05:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"9ebb888c-dfe7-416a-9940-a913527d4f73","AI Agent 的失败比成功更值钱:5 万对错误诊断数据,修正通过率 18.4%→51.1%","agent-error-dataset","2026-10-01T15:11:08+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"86565410-ced9-4e64-80b7-d97a353a1d1d","LLM 元认知首次可度量：Metacognition-Bench 用 300 道陷阱题 + 11 个开源错题雷达适配器把自我纠错做成开源工程","metacognition-bench-llm-self-correction","2026-07-01T12:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"19dd65ff-646c-4edf-bd90-98b2dc8bf215","Web Agent幻觉可解释性:探针能预警,归因却失灵","web-agent-hallucination-mechanistic-interpretability-probe","2026-11-17T03:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d1d54032-073b-433e-9430-547c6ffb4a49","贵的模型，没写出更快的代码：9455 份实测","perfarena-llm-code-performance-ranking","2026-11-16T19:13:41+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4fcd63ae-2b7c-43a2-9716-d7c20ca4cfd7","模型指纹会传染:中途换人,检测率0.97跌到0.72","contagious-llm-fingerprints-relay-audit","2026-11-16T13:30:00+00:00"]