Abliterlitics Benchmarks 8 Uncensored Qwen 27B Variants Across 167 GPU Hours
Abliterlitics conducted an 11-day evaluation using 167 GPU hours to compare eight uncensored (abliterated) variants of the Qwen 27B model against its base version across KL divergence, weight checks, and HarmBench refusal metrics. The study found that minimal, surgical weight modifications vastly outperformed aggressive edits, which degraded reasoning capabilities and caused models to get trapped in infinite thinking loops. This research provides empirical proof for the open-source AI community that targeted refusal-direction removal preserves baseline capabilities much better than broad tensor modifications. It also highlights new failure modes in reasoning models, where aggressive alignment removal causes models to exhaust their token budget during internal chain-of-thought processing. The top-performing variant, 'orcarouter' (82.2% Attack Success Rate), modified only a single direction at layer 38 across 131 matrices and achieved the highest copyright unlock at 39%. In contrast, the most heavily edited variant, 'obliteratus' (modifying 841 of 850 tensors), caused 44.8% of responses to stall in infinite thinking loops and suffered severe intelligence degradation.
## BACKGROUND
Abliteration is a post-training technique that removes an LLM's refusal behavior by identifying the linear 'refusal direction' in activation space and modifying model weights to ablate it. Research by Andy Arditi et al. demonstrated that refusal in language models is often mediated by a single linear direction, allowing for precise weight intervention. Benchmarks like HarmBench evaluate how compliance and safety boundaries change when models are subjected to adversarial red-teaming queries.