Jun 2026· arXiv.org· Vol abs/2606.31748· 0 citations· 61 references
Computer Science
TL;DR
The resulting model SEAR deliberately engages in harmful reasoning as exploration while reliably flipping back to a safe answer, demonstrating that this behavior helps mitigate over-refusal and defend against attacks that directly manipulate the reasoning to be harmful.
Abstract
Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones. Though this trade-off can be mitigated by training models with reinforcement learning (RL) to reason before answering, it does not remove the underlying problem that reasoning can often be a"rubber stamp"for a predetermined response. In this paper, we address the safety-refusal trade-off by rethinking how models are trained to reason about safety. Our key insight is that unsafe reasoning can itself serve as a useful exploratory signal. Rather than preemptively blocking harmful thoughts, we encourage the model to sufficiently explore unsafe reasoning but produce a safe response. The harmful exploration improves the model's ability to distinguish harmful from harmless prompts by resolving ambiguity, allowing it to remain safe while complying only when appropriate. We cast this as an adversarial optimization problem in which a reasoning player explores strategies for producing an unsafe response and an answer player ensures that the final output is safe. We train a single model with dense rewards to play both roles within one chain-of-thought, across different segments. To achieve this, we find that process rewards are crucial for stable optimization of competing objectives. Our resulting model SEAR deliberately engages in harmful reasoning as exploration while reliably flipping back to a safe answer. We demonstrate that this behavior helps mitigate over-refusal and defend against attacks that directly manipulate the reasoning to be harmful.
Safety guards are widely used to filter harmful content and are typically trained via supervised fine-tuning on labeled prompt-response pairs. We audit two widely used safety-guard training datasets, WildGuardMix and GR-Train, and find that among responses to harmful prompts, refusal expressions co-occur almost exclusively with unharmful labels. This imbalance motivates what we term the refusal-cue shortcut: inserting a refusal cue into a harmful response could flip the guard's verdict from harmful to unharmful. The shortcut affects not only guards trained on these datasets but also officially released models such as LlamaGuard3 and Qwen3Guard whose training data is undisclosed. It persists across response positions and is generally stronger in smaller variants within a family. To mitigate it, we adapt sparse complementary masking as a lightweight post-hoc intervention that identifies and suppresses a small set of shortcut-associated attention heads and MLP neurons without retraining. On two primary benchmarks, the intervention achieves an approximately 79% relative reduction in response-initial detection failures induced by refusal cues, while preserving standard detection performance. Although optimized using cues at a single response position, the suppression effect transfers to unseen positions and datasets, suggesting that shortcut manifestations across positions are partly mediated by shared internal components. Further analysis provides evidence that shortcut reliance and legitimate refusal recognition are partially functionally separable, as suppressing the shortcut broadly preserves the guard's ability to recognize genuine refusals.
Yu Feng, Chunting Zang, Chen Shen et al.· 0 citations
Safety defenses for large language models (LLMs) have been extensively studied, with existing approaches focusing on attack detection and refusal mechanisms. Such fixed-form direct refusal strategies may introduce the risk of prefix injection attacks. Recent work has explored a new direction that leverages humor as an indirect refusal mechanism to mitigate over-refusal in jailbreak scenarios and reduce prefix injection risks. However, this approach implicitly assumes that humorous responses are safe. Whether humorization itself introduces safety risks remains unexplored. To address this issue, we conduct an exploratory study involving over 30,000 real-world agent interaction records and 45 stand-up comedians, revealing practical safety concerns in LLM-based content humorization. Motivated by these findings, we propose \textsc{HumorSafe}, a novel framework for evaluating latent safety risk propagation during humorization. \textsc{HumorSafe} enables LLMs to learn harmful humorization patterns and use them to transform benign content into humorous content with safety risks. Across five frontier LLMs, we find that LLMs can introduce stereotypes and toxicity during humorization. We further propose \textsc{HumorPIA}, a prompt injection attack that exploits latent risks in humor-based defenses. \textsc{HumorPIA} preserves the appearance of safe humorous refusal while covertly injecting harmful content, allowing latent risks to evade existing detection mechanisms. Experiments show that it increases toxicity by 3.14$\times$ while maintaining an apparent safety rate of 97.8\% even under defense settings. Our findings highlight a gap in existing LLM safety evaluations under humorized settings.
Yu Cui, Ruiqing Yue, Tingyu Li et al.· 0 citations
LongGuard is presented, a framework that evaluates, mechanistically analyzes, and mitigates long-context guardrail failure, and proposes two training-free mitigations - Chunked Detection and Attention-Head Sharpening (AHS) - and a deployment protocol that selects configurations by context length and audit side.
Experiments show that with only 1K synthesized samples, AdvSafe-aligned LRMs achieve significantly stronger jailbreak robustness than existing baselines, with almost no utility degradation, demonstrating that learning unsafety knowledge enables a superior robustness-utility trade-off and generalizes beyond seen attack patterns.
Hongli Shen, Shaopeng Fu, Qinbo Zhang et al.· 0 citations
Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks'CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.
The proposed Recursive Self-Correction approach raises model performance from a Political Neutrality Likert scale baseline of 2.14 to 4.56, averaged across all models, demonstrating effective inference-time mitigation of political bias in LLM-generated summaries.
Tejaswi V. Panchagnula, Bruce Coburn, Bryce J. Dietrich et al.· 0 citations