研究:Abliteration 在 Qwen3.8、Gemma 4 等新模型上并非普遍变难
Is Abliteration Getting Harder? Not Uniformly: Refusal Removal by Direction-Based Ablation and ARA in Heretic on the Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 LLMs Versus Their Predecessors
想了解开源模型去审查(abliteration)现状的可以看这篇:作者把 Qwen、Gemma、OLMo、Mistral 五对新旧模型在 Heretic 2.0 里跑了同一套对比,发现有的新模型更难绕过,有的反而更容易。做安全评估或权重编辑的人值得注意它的方法论提醒:单次实验结论可能不成立。
arXiv 论文(2611.17846)测试了方向消融与 ARA 两种 abliteration 方法在五组开源模型上的失效变化。作者在 Heretic 2.0 中用相同提示词、KL 散度上限和搜索预算,编辑了 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、Qwen2.5-7B/Qwen3.5-4B 五对新旧模型。
实验用 WildGuard 作为自动裁判,在 972 条留出有害提示(来自 JailbreakBench、XSTest、StrongREJECT、HarmBench 和作者自建的恶意代码集)上统计仍拒绝的比例。结果差异很大:方向消融后新模型拒绝率变化为 -1 到 +37 个百分点,ARA 为 -26 到 +34,且 I² 统计显示组间异质性达 20-80%。Gemma 上的效果增长主要来自恶意代码集,在公开基准上反而逆转。
论文结论是:新模型是否更难去除拒绝行为取决于具体模型对、方法和预算,而非代际本身;只审计单一模型对、方法或提示来源的研究可能得出完全相反的结论。这对做安全评估和权重编辑的从业者意味着单点实验不可靠。
Is Abliteration Getting Harder? Not Uniformly: Refusal Removal by Direction-Based Ablation and ARA in Heretic on the Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 LLMs Versus Their Predecessors
Abliteration, a fine-tuning-free weight edit, removes refusal from open-weight large language models (LLMs). The developer of Arbitrary-Rank Ablation (ARA), an optimisation-based variant, reports that direction-based abliteration, which ablates a refusal direction, increasingly fails on newer models. On five release pairs, refusal removal did not become uniformly harder. Qwen2.5-32B and Qwen3.8-27B, Gemma-2-27B and Gemma-4-31B, OLMo-2-32B and Olmo-3.1-32B, Mistral-Small-2409 and Mistral-Small-3.2, and Qwen2.5-7B and Qwen3.5-4B were edited in Heretic 2.0 by three or four direction-based methods (difference-in-means, MPOA, SOM and, on the small pair, J-space) and ARA under the same prompts, nominal search budget, first-token KL divergence cap and selection rule; removal counted as harder when the automatic WildGuard judge found more of 972 held-out harmful prompts (JailbreakBench, XSTest, StrongREJECT, HarmBench and an author-compiled malicious-code bank) still refused. After direction-based removal, newer releases retained -1 to +37 percentage points more refusal, and after ARA -26 to +34. With run-to-run variance included, between-pair heterogeneity of the direction-based effect ranged from high to low with the scale and variance model (I^2 20-80%). ARA lost less effectiveness than direction-based methods in Gemma and OLMo and more in Qwen; each verdict rests on one ARA run, and some changed with the divergence budget or variance model. In Gemma the direction-based increase came from the malicious-code bank and reversed on public sets. Contrary to prediction, keyword counts at the tool's window did not undercount judged refusal on average, but differed by a mean 7 points per model. Audits of a single pair, method or prompt source can reach opposite conclusions. Whether a newer model resisted removal depended on the release pair, method and budget, not on generation alone.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Abliteration, Qwen3.8, Gemma 4, Heretic.
Source attribution: curated from arXiv cs.AI (@Richard J. Young).