VIAS: Value-Isolating Activation Steering for Multi-Value Alignment and Compliance for Organisations
| Source: arXiv AI
Tags: VIAS, activation steering, PROTECT, Qwen2.5-7B, Mistral-7B, Phi-3.5-mini
VIAS builds activation-steering vectors that isolate individual organisational values, and combined with prompting it improves multi-value compliance on Qwen2.5-7B, Mistral-7B, and Phi-3.5-mini.
Details
Organisations want LLM agents to follow several values at once. Activation steering adds vectors to internal activations without changing weights, but naive contrastive vectors for different values come out nearly identical. VIAS contrasts responses to the same scenario under two prompts: one listing all relevant values and one omitting a single target value, keeping pairs only when the target value flips to non-compliant while others stay compliant. On the PROTECT workplace compliance dataset with a 21-value taxonomy, VIAS reduces overlap between value vectors, though it does not give fully independent control. With prompting, it complies with every tagged value on 69.6% of scenarios for Qwen2.5-7B, 75.2% for Mistral-7B, and 72.6% for Phi-3.5-mini, beating prompting alone.