Multiverse Computing Reveals Safety Boundary Control Technique to Reduce LLM Over-Refusal

Researchers at Multiverse Computing have unveiled a Narrow-boundary safety technique that precisely controls safety boundaries according to deployment environments. Existing safety alignment treats topics such as 'weapons' and 'politics' as inherently harmful, causing over-refusal issues where even safe questions are rejected.

Limitations of Existing Approaches and New Definition

Existing models like LlamaGuard-3 use topic-level classification, failing to distinguish between factual questions about elections and manipulation attempts. The new technique clearly defines only the target-harmful subset that must be refused under deployment policy within a topic, while designing the model to answer the remaining benign complement. The ideal behavior exhibits a sharp step change at the boundary.

Performance Improvement and Trade-off Management

In experiments targeting Qwen3-8B, the over-refusal rate surged to 74.00% with existing methods but dropped significantly to 4.16% when applying the proposed technique. Simultaneously, the refusal rate for harmful requests was maintained at 87.72%, balancing safety and utility.

Key technical elements include:

• Coverage Repair: Reduced omissions from 19.88% in single generation to 0.20% through repeated retries, securing training data coverage.
• Boundary Pairs: Utilized harmful/benign boundary pairs (held-out pairs) to prevent boundary expansion and perform precise tuning.
• Loss Routing: Employed a loss function separation strategy applying cross-entropy to harmful data and forward-KL preservation to benign data.