Semantic Routing Calibration mitigates LLM over-refusal
Large language models that are aligned for safety often reject benign instructions, a problem known as over‑refusal. Prior work has linked the issue to static representation overlap, but the underlying dynamic causes have not been fully explored. The paper investigates over‑refusal by examining internal routing conflicts in transformer attention, focusing on a small set of hypersensitive safety…
Key points
- SRC suppresses hypersensitive safety heads during inference.
- SRC uses dual‑branch logits fusion as a safety regularizer.
- Experiments show SRC reduces over‑refusal while preserving safety performance.
To address this, the authors introduce Semantic Routing Calibration (SRC), a lightweight, training‑free inference framework. SRC identifies the problematic safety heads at inference time and dynamically suppresses them. It also applies a dual‑branch logits fusion that acts as a safety regularizer during decoding, helping the model maintain trustworthy reasoning.
Extensive experiments show that SRC reduces over‑refusal rates while keeping intrinsic safety performance largely intact. The authors claim that the approach preserves safety constraints and improves the model’s ability to handle benign safety‑related queries.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Study finds wiki-indexed LLM outperforms vector RAG for cross-unit questions in ML courses · 1 src
- Study proposes skill habits to fix AI agent inconsistency · 1 src
- Researchers propose latent equivalence learning for enterprise data agents, score 94.67% on benchmark · 1 src
- Researchers extract circuits from language models using Attention routing · 1 src
- ReAdapt improves warm‑introduction and reaction selection accuracy for Gemini‑3‑Flash · 1 src
Comments
via GitHub Discussions