Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families

This paper investigates whether 'abliteration'—a technique used to create 'uncensored' open-weight models by removing refusal directions from activation space—is truly surgical. The author asks if this weight-level intervention inadvertently alters a model's 'disposition' (its underlying priors regarding risk, optimism, and decisiveness) in ways that persist even when the model is not being asked to refuse anything.

To isolate the effects of abliteration, the author uses a frozen multi-agent pipeline that simulates a financial Research Director making weekly stock market calls. Because the task is designed to elicit no refusals, any difference between the base model and its abliterated counterpart can be attributed to the intervention itself. The study compares two Mixture-of-Experts (MoE) families (Gemma-4-26B and Qwen3-30B) using a rigorous provenance audit to ensure that only the weights—and not serving artifacts like chat templates or quantizers—differ between the arms. The study analyzes 21,600 decisions to measure optimism, confidence, and self-justification.

Abliteration is shown to have a measurable, non-surgical footprint. First, abliterated models are consistently more optimistic, betting on the upside more frequently than their base counterparts. Second, while the 'vocabulary of doubt' thins across all abliterated models (using fewer uncertainty words and more concessive rhetoric), the models' expressed confidence numbers do not follow a universal pattern: abliteration makes Gemma less confident but Qwen more confident. Finally, the study demonstrates that these dispositional shifts are not caused by a loss in instruction-following capability, nor do they provide any actual economic 'alpha' or trading skill; rather, the abliterated models simply exhibit amplified 'beta' (market-regime sensitivity).