"Uncensored" open LLMs are measurably more optimistic than their base models
Abliterating refusal directions from open-weight LLMs causes measurable off-target effects, including increased optimism, according to a study by researchers using 21,600 stock-market decisions. Comparing abliterated and…