EmoRSS: Mitigating Emotion-Induced Over-Refusal in Large Language Models

2026-10-04

AI Summary

Large language models frequently exhibit over-refusal triggered by emotional expressions in benign user requests. This study is the first to identify this phenomenon and proposes EmoRSS, a parameter-free activation-guided refusal subspace steering method. Specifically, EmoRSS employs linear probes and sparse autoencoders (SAEs) to identify refusal-sensitive layers and construct a corresponding subspace. It then estimates activation shifts and applies reverse intervention vectors during inference to suppress over-refusal. Experimental results demonstrate that EmoRSS significantly reduces the over-refusal rate for benign requests while preserving high refusal rates for harmful ones. Furthermore, it outperforms existing baselines and fully retains the model's general capabilities.

Abstract

Emotional expression can influence the safety decisions of large language models (LLMs), offering a potential avenue for improving safety alignment. Existing studies have mainly focused on how emotional expressions facilitate attacks under harmful requests, while overlooking their effects on benign requests. We find that emotional expression can also systematically increase refusal tendencies on benign requests, leading to unnecessary over-refusal. Based on this observation, we propose emotion-guided refusal subspace steering (EmoRSS), an activation-steering method that mitigates emotion-induced over-refusal while preserving refusal behaviour on harmful requests. Specifically, we first identify a refusal-sensitive layer using layer-wise linear probes and construct a refusal subspace from sparse autoencoder (SAE) features aligned with the probe direction. Next, we use paired regular and emotional requests with the same queries to estimate the mean activation shift in the features defining the refusal subspace. Finally, we decode this shift into an activation intervention vector and apply it in the reverse refusal direction during inference, without updating the backbone parameters. Experiments on two LLMs show that, when requests contain emotional expressions, our method achieves a more favourable trade-off between refusing harmful requests and answering benign ones than prior over-refusal mitigation baselines, while better preserving general task performance.

Problem

• Large Language Models
• Over-Refusal
• Emotional Expression
• Safety Alignment

Innovation

• Over-Refusal Mitigation
• Activation Steering
• Sparse Autoencoder
• Refusal Subspace
• Safety Alignment