Number 8 of 12 · November 12, 2026

TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation

TERM is a defense that runs while a model answers. It watches how the model's attention shifts from its own starting point and scales down the model's internal updates when something looks off.

What they found

The authors propose TERM, a defense against jailbreaks (prompts built to trick a model into harmful answers). It measures attention entropy, a sign of how spread out the model's focus is. It compares that value to the same prompt's own baseline from the prefill step, when the model first reads the prompt. Larger and longer-lasting deviations lead to stronger, bounded scaling of the model's own residual updates (the internal changes made at each layer). It needs no predefined safety direction. The authors tested it against seven defenses, seven jailbreak attacks and two over-refusal benchmarks, on Llama, Qwen and Gemma models, including Qwen3 from 0.6B to 32B. They report better jailbreak robustness, low over-refusal on benign but sensitive requests, and modest inference overhead. The abstract gives no numbers.

Why it matters

Safety defenses often refuse harmless requests too, which makes models less useful. This work tries to protect models without that cost. It also avoids a fixed target, so the defense adapts to each prompt. The abstract does not give specific gains, so the size of the benefit is unclear.

How it changes the race

If the result holds, builders could add a jailbreak defense at inference time without retraining or a fixed safety direction. This would be a modest, practical gain in the balance between safety and usefulness, not a change in raw capability.

Safety and control

The paper bears directly on jailbreak defense. It is tested on seven attacks, but the abstract does not say how it fares against new or adaptive attacks that target this method.

The abstract, by the authors

Large 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. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.

Quoted from arXiv:2611.13577. The research and its words belong to its authors.

The authors

Yunkai Dang, Wanyu Zhang, Meiyi Zhu, Qi Fan, Tianyu Ding, Wenbin Li

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.