cd /news/ai-safety/is-abliteration-getting-harder-not-u… · home › topics › ai-safety › article
[ARTICLE · art-170375] src=machinebrief.com ↗ pub= topic=ai-safety verified=true sentiment=· neutral

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

A study of five model release pairs edited in Heretic 2.0 found that refusal removal did not become uniformly harder on newer open-weight LLMs, contradicting the developer of Arbitrary-Rank Ablation (ARA), who reported that direction-based abliteration increasingly fails on newer models. After direction-based removal, newer releases retained -1 to +37 percentage points more refusal across 972 held-out harmful prompts judged by WildGuard, and after ARA -26 to +34, with between-pair heterogeneity ranging from high to low (I^2 20-80%). The authors conclude that whether a newer model resisted removal depended on the release pair, method and budget, not on generation alone, and that audits of a single pair, method or prompt source can reach opposite conclusions.

by read1 min views1 publishedNov 16, 2026

arXiv:2611.17846v1 Announce Type: cross Abstract: 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.

── more in #ai-safety 4 stories · sorted by recency
── more on @heretic 2.0 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/is-abliteration-gett…] indexed:0 read:1min 2026-11-16 · —