Item titled POEM describes the TERM jailbreak defense
Original title: POEM: Output-Relative Block-Sparse Attention for Long-Context Prefilling
The supplied abstract is TERM, tested against seven jailbreaks on Llama, Qwen, and Gemma, including Qwen3 from 0.6B to 32B.
The item title names POEM, an output-relative block-sparse attention method, but the body describes Target-free Entropy-guided Residual Modulation (TERM). TERM is an inference-time defense that scales a model's own residual updates using prompt-calibrated attention-entropy deviations, without a predefined safety direction. The authors compare it with seven defenses on seven jailbreak attacks and two over-refusal benchmarks. They report better jailbreak robustness, low over-refusal, and modest inference overhead.
Unconfirmed: The title describes POEM and block-sparse attention, while the abstract describes TERM; which paper this item is cannot be confirmed.
- Intervention strength uses adaptive scale normalization and temporal accumulation.
- Entropy is calibrated to each prompt's own prefill baseline.
- Evaluation covers Llama, Qwen, and Gemma, including Qwen3 from 0.6B to 32B.
- The abstract claims low over-refusal on benign safety-sensitive requests.
From the original article
1 min read at arxiv.orgLarge language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses.
By Haoyi Wu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.