The BaseResearcharXiv
TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation
According to the abstract, Yunkai Dang, Wanyu Zhang, Meiyi Zhu and three co-authors propose TERM, an inference-time defense against jailbreak prompts. It uses each prompt's attention entropy, calibrated against that prompt's own baseline, to adjust the model's residual updates while limiting refusals of benign requests.
Why it matters
The same thing, for different readers.
For researchers
The method adjusts a model's internal updates per prompt without a predefined steering direction, aiming to cut over-refusal; its performance figures are not in the excerpt.
The record
- Questions
- Are AI companies putting stronger safeguards into practice?
- People
- Yunkai Dang (Wrote it) · Wanyu Zhang (Wrote it) · Meiyi Zhu (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of November 11, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation”, November 10, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-10-term-jailbreak-defense-entropy
Our notesLight lane: triage excerpt only; no results or models tested were visible.