Skip to content
NewA calmer, faster ZeroHour
StoryAI safety & securityinfo

TERM paper proposes inference-time jailbreak defense without steering

1 source · 2 articlesFirst reported Updated ()
What's new

First merged summary.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

A new paper describes TERM, an inference-time jailbreak defense tested on Llama, Qwen, and Gemma with low reported over-refusal.

Researchers propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense against jailbreak prompts. It calibrates attention entropy to each prompt’s prefill baseline and uses that signal to bound scaling of the model’s own residual updates, without a predefined safety direction or learned safety region. The authors compare TERM 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. They report stronger robustness, low over-refusal on benign safety-sensitive requests, and modest inference overhead. Intervention strength is set by adaptive scale normalization and temporal accumulation.

Key facts
  • TERM scales residual updates using prompt-calibrated attention-entropy deviations.
  • No predefined steering direction or learned safety region is used.
  • Tests cover seven defenses, seven attacks, and two over-refusal benchmarks.
  • Models include Llama, Qwen, and Gemma, plus Qwen3 from 0.6B to 32B.
  • Authors report modest inference overhead and low over-refusal.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

What each outlet adds

beyond the shared facts
  • arXiv cs.CRNames scale normalization and temporal accumulation as strength controls.
  • arXiv cs.CR (second item)Item title says POEM, but the abstract body describes TERM.
AI models
LlamaQwenGemmaQwen3

Coverage

oldest first · one row per article · UTC
  1. · 1d ago arXiv cs.CRFirst report

    TERM jailbreak defense modulates residuals from attention entropy

    Compared with seven defenses on seven attacks, it reports stronger robustness and low over-refusal at modest inference cost.

  2. · 1d ago arXiv cs.CR

    Item titled POEM describes the TERM jailbreak defense

    The supplied abstract is TERM, tested against seven jailbreaks on Llama, Qwen, and Gemma, including Qwen3 from 0.6B to 32B.