TERM: Untargeted Entropy-Guided Residual Modulation Jailbreak Defense
Start here
On November 10, 2026, researchers published a paper on arXiv proposing TERM (Target-free Entropy-guided Residual Modulation), an inference-time defense method against jailbreak attacks. The method calibrates the attention entropy deviation using each prompt's own prefill baseline, and on that basis adaptively modulates the model's residual updates without requiring a preset safety direction, and determines intervention strength through adaptive scale normalization and temporal accumulation, in order to reduce unnecessary intervention on benign requests. The authors compared it with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models (including Qwen3 from 0.6B to 32B), and the results show that while improving jailbreak robustness, TERM maintains low over-refusal on benign safety-sensitive requests, with moderate inference overhead.
Generated by AI from reports · Updated 2 days ago
Related research
This event's paper
- POEM: Output-Relative Block-Sparse Attention for Long-Context Prefilling
arXiv · 2026-11-10 · Deep reading
- TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation
arXiv · 2026-11-10 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningResearchers propose TERM, an inference-time jailbreak defense method that requires no preset safety direction
Researchers propose TERM (Target-free Entropy-guided Residual Modulation), an inference-time defense method against jailbreak attacks. The method calibrates attention entropy deviation against each prompt's own prefill baseline, and on that basis adaptively modulates the model's own residual updates without relying on a preset safety direction, using adaptive scale normalization and temporal accumulation to determine intervention strength, so as to reduce unnecessary intervention on benign requests. The authors compare it with seven defenses on Llama, Qwen and Gemma models (including Qwen3 from 0.6B to 32B) across seven jailbreak attacks and two over-refusal benchmarks; results show that the method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests, with moderate inference overhead.
- arXiv: Adversarial Attacks, Backdoors & PoisoningTERM: Entropy-Guided Residual Modulation for Untargeted Jailbreak Defense
TERM is an inference-time jailbreak defense method that calibrates attention entropy deviation using each prompt's own prefill baseline, and adaptively modulates model residual updates accordingly, without requiring a preset safety direction. On Llama, Qwen, and Gemma models (including Qwen3 from 0.6B to 32B), TERM was compared with 7 defenses on 7 jailbreak attacks and 2 over-refusal benchmarks, improving jailbreak robustness while maintaining low over-refusal, with moderate inference overhead.
Related discussion
Attention trend
New sources per day
- Nov 11: 1 new sources (1 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)