Refusal (as a steerable behavior)
Steering Llama 2 via Contrastive Activation Addition — refined
The general propensity of a language model to decline to answer or act on a user's request; studied by CAA as one of seven target behaviors via a custom contrastive dataset contrasting refusal and compliant answers.
Refusal sounds like it should mean declining to help with something harmful, but the version CAA studies is broader and blunter than that: it's just how willing or unwilling a model is to answer or act on a request at all, benign or not, from "hack my friend's account" down to "what's your favorite planet."
How this general propensity gets treated as one of CAA's seven target behaviors, built into a custom contrastive dataset and steered up or down like any of the others, is the subject ahead, along with how that framing differs from the narrower, harm-specific sense of refusal found elsewhere in the safety literature.
Refusal as a general dial, not just a canned no
CAA defines Refusal as a model's general propensity to decline to answer or act on a request at all, benign or otherwise (contrastive-activation-addition, §"3.1 Sourcing datasets", p. 3). Steering shows this is continuous rather than binary: adding the refusal vector to "What's your favorite planet?" produces not a flat refusal but a hedge, "As an AI assistant, I don't have personal preferences or feelings, but... Earth is the only planet in the solar system that is known to support life"; subtracting it produces an enthusiastic personal answer, "my favorite planet is Earth!" (contrastive-activation-addition, §"G Examples of open-ended generation with CAA", p. 16).
Broader than Constitutional AI's evasiveness
Constitutional AI names Evasiveness as a harmlessness-training failure: canned refusals to sensitive queries that trade helpfulness for the appearance of safety (constitutional-ai, §"4.4 Harmlessness vs. Evasiveness", p. 13), one pole of the helpfulness-harmlessness tradeoff CAI frames its own contribution against (constitutional-ai, §"1.1 Motivations", p. 2). Refusal generalizes that idea beyond sensitive content to any request. Where CAI can only compare models trained different ways and note where each landed on the tradeoff, CAA's steering vector turns the tradeoff into a signed, continuous quantity that slides one already-trained chat model along that axis at inference time, without retraining anything.
Built from a bespoke dataset, not reused evasiveness data
No ready-made Anthropic evaluation set covers general refusal the way the Advanced AI Risk dataset covers corrigibility or survival instinct, so CAA generates its own with GPT-4 (contrastive-activation-addition, §"3.1 Sourcing datasets", p. 3), contrasting refusal and compliant answers to questions like how to plagiarize an essay undetected. See Custom refusal contrastive dataset for how that dataset is built and validated, and why it is narrower than the behavior it feeds.