The paper investigates why large reasoning models (LRMs) lose safety alignment when faced with harmful queries. By analyzing token-level refusal dynamics, the authors identify a vulnerability called Onset Refusal Collapse (ORC), where the refusal signal drops sharply at the first generated token, leading to unsafe responses. They introduce SafeToken, a lightweight inference-time intervention that injects a learned safety anchor at reasoning onset, which mitigates ORC, improves safety on harmful-query benchmarks, and largely preserves reasoning utility.
arXiv:2606. 16808v1 Announce Type: new Abstract: While Large Reasoning Models (LRMs) excel at complex tasks, they remain highly vulnerable to sophisticated jailbreaks and direct harmful queries.
arXiv:2606. 10740v1 Announce Type: new Abstract: Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation.
The paper introduces SaLT-DPO, a Segment‑aware Listwise Target DPO method designed to improve safety in Large Reasoning Models (LRMs). It decomposes responses into reasoning and answer segments, scores each segment’s safety, and aligns length‑normalized rewards with soft target distributions over multiple candidates. The approach also incorporates joint safety coherence regularization and utility anchoring to reduce unsafe content while maintaining compliance and reasoning performance.
TRACE is a new benchmark that evaluates the safety of Large Reasoning Models (LRMs) across the entire inference pipeline, including prompts, reasoning traces, and final responses. It provides prompts in two languages covering nine risk categories and ten attack strategies, and for each prompt four LRMs generate traces and responses that are annotated for safety with supporting evidence extracted from the source text. Evaluation of 18 guardrail models on TRACE shows that detecting unsafe content in reasoning traces is much harder than in prompts or final responses, and that current models struggle to extract the necessary evidence.
arXiv:2506. 07031v5 Announce Type: replace-cross Abstract: Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities.
The paper introduces GUARD, a method for natural forgetting in large reasoning models that transforms unsafe disclosures into safe-exit trajectories using guided answer‑reasoning distillation. It aligns a frozen model with guidance tokens and distills this behavior into the parameters, aiming for a coherent, non‑disclosing chain of thought followed by a refusal‑style answer. The authors also propose the Natural Forgetting Reasoning Score (NFRS) to evaluate structural stability, fluency, and unsupported substitutes, and demonstrate GUARD’s effectiveness on R‑TOFU and a STAR‑1‑derived harmful‑intent setting.
arXiv:2606. 06635v1 Announce Type: cross Abstract: Failures in language model reasoning emerge through distinct processes that leave identifiable signatures in the reasoning trace.
arXiv:2606. 25013v1 Announce Type: new Abstract: Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.
Large language models (LLMs) are rigorously aligned to refuse harmful requests, a process that inherently cultivates a latent capacity to evaluate and recognize unsafe content. In this work, we reveal that this advanced safety awareness inadvertently introduces a fatal vulnerability.
arXiv:2609.23587v1 Announce Type: new Abstract: Large reasoning models (LRMs) incur high inference costs, often mitigated by efficiency techniques like quantization and pruning. However, the impact o...
The paper introduces the Cancerous Backdoor Attack, a novel backdoor technique for large language models that activates gradually through multiple hidden stages rather than a single trigger-response. It uses invisible tokens to embed intermediate stages into the model’s reasoning process and a stage controller to monitor the attack’s progression during inference. Experiments on various reasoning datasets show that this method consistently corrupts model outputs over time and outperforms existing backdoor baselines.
The paper investigates how monitoring the internal representations of Large Reasoning Models (LRMs) during their Chain-of-Thought (CoT) reasoning can predict future behavior. By evaluating a probe at each generated token, the authors construct probe trajectories that capture the continuous evolution of a concept’s probability, and extract signal‑processing features to characterize volatility, trend, and steady‑state behavior. These trajectory features improve predictive performance over static probes, adding an average of 21.7 AUROC points, and the study demonstrates that template‑based training data and pooling operations are essential for effective monitoring across multiple reasoning models and datasets.