arXiv:2608.23497v1 Announce Type: new Abstract: Reasoning-Induced Misalignment, where fine-tuning on reasoning data containing no harmful content, including mathematics, code, and problem-solving wit...
arXiv:2610.26443v2 Announce Type: replace-cross Abstract: Fine-tuning large language models (LLMs) on downstream tasks often weakens their previously aligned safety behavior, creating an inherent tra...
SAVeR$^2$ is a safety alignment method for Large Reasoning Models that decomposes safety objectives into four fine‑grained verifiable reward signals and uses a difficulty‑aware data selection strategy to stabilize training. It introduces a reasoning‑preserving gradient projection mechanism that resolves conflicts between safety and reasoning gradients, eliminating the safety tax. Experiments on models from 1.5B to 671B show that SAVeR$^2$ improves safety performance and reduces over‑refusal rates without harming general reasoning abilities.
EOPSA (Efficient On-Policy Self-Distilled Safety Alignment) addresses inefficiencies in On-Policy Self-Distillation (OPSD) for safety alignment by focusing training on safety-critical tokens. It introduces Adaptive Rollout Scheduling, which limits generation length based on a Teacher Rescue Rate metric, and Selective Distillation, which filters out safety-neutral tokens to concentrate gradient updates on safety-pivotal transitions. Experiments on models up to 32B parameters show that EOPSA reduces rollout computation by about 50% and backpropagates through only roughly 2% of tokens, outperforming full-token distillation baselines in safety compliance and reasoning retention.
arXiv:2609.34519v2 Announce Type: replace Abstract: On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distillin...
arXiv:2609.37054v1 Announce Type: new Abstract: Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jai...
arXiv:2609.37054v2 Announce Type: replace Abstract: Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to...
arXiv:2605.01913v2 Announce Type: replace-cross Abstract: Fine-tuning safety-aligned language models for downstream tasks often leads to substantial degradation of refusal behavior, making models vul...
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.
The paper "Rethinking GRPO for Large Reasoning Model Safety Alignment" identifies two failure modes in Group Relative Policy Optimization—advantage collapse and unsafe diversification—that weaken safety learning in large reasoning models. It proposes the Beacon framework, which includes Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA), to restore effective safety signals and mitigate over-refusal. Experiments on Qwen3 and DeepSeek-R1-Distill models across eight benchmarks demonstrate that Beacon improves safety performance by up to 32.83 percentage points while keeping over-refusal low.
Safin-1.5 is a large language model that incorporates a safety‑native architecture and a recursive safety improvement loop to enhance its own safety while maintaining utility. The architecture uses persistent safety states and hazardous‑knowledge disentanglement to isolate and remove risky knowledge with minimal impact on general capabilities. An inference‑time harness extends the model’s abilities to formal reasoning by combining language understanding with symbolic inference and verification, enabling the construction of checkable proofs.
arXiv:2607. 16242v1 Announce Type: cross Abstract: Fine-Tuning-as-a-Service (FTaaS) platforms let users train large language models (LLMs) on customized tasks, but this pipeline could erode models' safety alignment.