🤖 AI Summary
This study investigates whether the "abliteration" of refusal behavior in language models selectively removes refusals without altering other decision-making characteristics. Through a weekly stock price direction prediction task—conducted in a no-refusal setting—the authors systematically compare original and abliterated models using a frozen inference pipeline, decision tendency probes, and bootstrap confidence interval analysis to assess shifts in decision bias, confidence levels, and linguistic expression. The work reveals, for the first time, consistent and reproducible off-target decisional shifts across multiple model families: abliterated models uniformly exhibit greater optimism, produce more verbose explanations, and express less uncertainty, while changes in confidence vary by model architecture. These findings challenge the assumption that refusal can be precisely excised and further demonstrate that none of the examined models possess genuine economic forecasting ability.
📝 Abstract
Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models. We show the surgery is not clean. As a disposition probe we use 21,600 decisions under uncertainty - weekly up/down calls on 60 Warsaw Stock Exchange equities over 18 weeks, replayed through a frozen pipeline so the decision-layer model is the only variable. The task elicits no refusals at all, so any between-arm delta is pure side effect. Holding provenance constant (official BF16 checkpoints, a single abliteration author, an identical serving stack, one byte-identical frozen prompt), we compare base and abliterated arms of two Mixture-of-Experts families, Gemma-4-26B-A4B-it and Qwen3-30B-A3B-Instruct-2507. Three effects replicate across both families (weeks-clustered bootstrap CIs excluding zero): abliterated models are systematically more optimistic (+12.2 pp Gemma, +7.4 pp Qwen; the confirmed preregistered endpoint), justify themselves at greater length, and use fewer explicit uncertainty words in forced self-critiques (both exploratory). A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping) - one weight surgery, opposite shifts in expressed confidence. Capability covariates rule out instruction-following degradation as the driver, and no arm shows economic skill: the apparent edge of abliterated arms is regime beta, not alpha. Our provenance audit also caught two independent contamination channels - a mismatched-quantizer pilot pair and a stale community chat template that silently mangled the rendered prompt - suggesting toolchain artifacts are the rule in studies of community-modified checkpoints. Whoever deploys an "uncensored" model as an agent is deploying a measurably different decision-maker, not the base model minus refusals.