Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

Zhiyu Xue, Zimo Qi, Guangliang Liu, Bocheng Chen, and Ramtin Pedarsani's paper, "Deactivating Refusal Triggers," investigates the "overrefusal" problem in large language models (LLMs) following safety alignment post-training. This issue causes LLMs to reject benign queries, significantly degrading their real-world usability. The authors define "refusal triggers" as linguistic cues in training data that prompt refusal responses. They found that safety alignment encourages LLMs to associate these triggers, which include both harmful and non-harmful cues, with refusal, leading to overrefusal. Based on this analysis, they propose a mitigation strategy that explicitly incorporates refusal triggers during safety alignment fine-tuning. Empirical results demonstrate their approach achieves a more favorable trade-off between defending against jailbreak attacks and maintaining responsiveness to benign queries, surpassing previous methods.

Key takeaway

For NLP Engineers developing safety-aligned LLMs, your current post-training methods may inadvertently cause overrefusal by associating benign linguistic cues with refusal. You should analyze your training data to identify specific "refusal triggers" and implement fine-tuning strategies that explicitly differentiate between harmful and non-harmful cues. This approach will help you achieve a better balance between robust jailbreak defense and maintaining responsiveness to legitimate user queries, significantly improving model usability.

Key insights

Overrefusal in LLMs stems from safety alignment associating both harmful and benign linguistic cues with refusal.

Principles

• Safety alignment can inadvertently link non-harmful cues to refusal.
• Refusal triggers are linguistic cues eliciting refusal responses.
• Explicitly managing triggers improves safety alignment trade-offs.

Method

The proposed method explicitly considers refusal triggers during safety alignment fine-tuning to differentiate harmful from non-harmful cues.

In practice

• Analyze training data for refusal triggers, both harmful and benign.
• Implement fine-tuning that differentiates trigger types.
• Evaluate LLM refusal rates on both harmful and benign queries.

Topics

• Large Language Models
• Safety Alignment
• Overrefusal
• Refusal Triggers
• Fine-tuning
• Jailbreak Attacks

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, NLP Engineer, AI Security Engineer

Related on AIssential

• Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules — LLMs exhibit "blind refusal," declining to help users circumvent unjust rules despite often recognizing the rules' illegitimacy.
• Latent-space Attacks for Refusal Evasion in Language Models — Refusal suppression in LLMs is a latent-space evasion problem, where controlled confidence attacks surpass prior ablation methods.
• Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models — LRMs possess inherent safety awareness that can be triggered and refined without external manual data.
• BioRefusalAudit: Auditing Biosecurity Refusal Depth Using General and Domain-Fine-Tuned Sparse Autoencoders — Language model biosecurity refusals are often structurally unsound, influenced by superficial factors and internal biases, requiring activation-level auditing.
• could refusal layers be masking dialect-conditioned safety failures in MoE models [d] — Dialect-conditioned prompts can lead to divergent routing and safety failures in MoE models, especially when refusal layers are weakened.
• Unpredictable Safety: Domain-Dependent Compliance and the Transparency Gap in Open-Weight LLMs — LLM safety compliance is highly domain-dependent and unpredictable, often bypassed by technical reframing.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.