Oyster-II is proposed, a reinforcement learning (RL)-based constructive safety alignment framework that adopts a Zero-RL paradigm combined with a multi-stage reinforcement learning strategy that comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions.
Abstract
Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulness, and trustworthiness remains a persistent challenge. Conventional refusal-oriented alignment strategies mitigate harmful content generation but systematically fail to serve legitimate user needs, often withholding information that could safely and constructively address the underlying intent of sensitive queries. Building upon the constructive safety paradigm pioneered by Oyster-I, which moves beyond blanket refusal toward thoughtful, response-oriented safety alignment, we identify two critical limitations of its Supervised Fine-Tuning (SFT)-based scheme: insufficient safety generalization to out-of-distribution scenarios and a phenomenon we term safety chain-of-thought (CoT) over-generalization, wherein safety-oriented reasoning patterns are excessively applied to benign queries, degrading helpfulness and user experience. To address these limitations, we propose Oyster-II, a reinforcement learning (RL)-based constructive safety alignment framework that adopts a Zero-RL paradigm combined with a multi-stage reinforcement learning strategy.Evaluated across extensive benchmarks, Oyster-II comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B.
Experiments across three instruction-tuned models show that HiRoute achieves high safety rates across multiple safety benchmarks while preserving safe-response helpfulness, reducing over-refusal, and maintaining competitive performance on general-purpose tasks.
Fangzhou Chen, Shiji Zhao, Mengyan Wang et al.· 0 citations
While multimodal large language models (MLLMs) extend model capabilities beyond text, they also make safety alignment increasingly challenging. Multimodal safety alignment methods must address cross-modal jailbreaks, safety-awareness failures, and over-sensitive refusals. However, existing methods often rely on retraining or internal-state inspection, limiting their applicability to deployed closed-source MLLMs and motivating test-time safety alignment. We analyze this setting and identify two key obstacles, utility dominance and reasoning inertia, which cause models to overlook latent risks or follow malicious reasoning trajectories. Guided by these insights, we propose ReFrame, a training-free multimodal input reframing framework where two agents share a lightweight locally deployed MLLM: the evidence-generation agent constructs complementary risk and utility evidence, and the rewrite-and-routing agent converts it into a safe proxy prompt and image-routing decision before calling the downstream MLLM, without modifying it or accessing its internal information. Experiments across multiple MLLMs and benchmarks show that ReFrame improves jailbreak defense, safety awareness, and oversensitivity reduction while preserving multimodal utility.
Wenzheng Jiang, Xuankun Rong, Yuanzhao Zhai et al.· 0 citations
Adapting instruction-tuned large language models (LLMs) to downstream domains is increasingly common, yet fine-tuning on imperfect data can erode the safety alignment learned during post-training. Existing safety-preserving fine-tuning methods typically optimize the aligned instruction model directly, which can destabilize refusal behaviors or impose an ''alignment tax'' that limits task adaptation. We propose SAFT (Safety-preserving Adaptation via Fine-tuning Transfer), a safety-preserving adaptation framework that decouples task learning from alignment preservation by learning a safety-guided task update on the paired pretrained base model, rectifying task gradients to avoid conflicting directions with respect to a safety objective, and then transferring the update to the frozen instruction model via parameter-space grafting. Across mathematical reasoning, code generation, and medical question answering on two open-source model families (Llama3.1-8B-Instruct and Gemma3-4B-IT), SAFT improves downstream utility while maintaining low harmfulness under a unified evaluation protocol, and achieves better safety and utility trade-offs than nine baselines. Warning: This paper contains unfiltered content generated by LLMs that may be offensive to readers.
Zhiwen Ruan, Yan Yang, Zhuocheng Liang et al.· Proceedings of the 32nd ACM...· 0 citations
While excelling at solving complex problems, Large Reasoning Models (LRMs) are still constrained by the overthinking issue. Most current studies rely on reward shaping in Rein-forcement Learning (RL) to shorten the Chain-of-Thought (CoT) of LRMs, remaining sample-inefficient and non-robust due to the absence of prioritized exploitation and guided exploration. To address these issues, we propose a novel policy optimization framework with S elf-I mitation and self-G uidance M ech A nisms (SIGMA), which reshapes the exploration and exploitation through two core components: (i) self-imitation exploitation , which enables the prioritized exploitation of high-value prompts and rollouts by introducing a self-imitation loss and a dynamic sampling strategy based on compression rate; (ii) self-guidance exploration , which provides a preference-aware exploration guidance through diverse and pluggable self-rewriting strategies. Experiments across various datasets indicate that our method achieves superior reasoning efficiency without compromising, and even facilitating, the overall accuracy. Furthermore, ablation studies show that the proposed mechanisms can provide flexible control interfaces for the tradeoff between the reasoning accuracy and efficiency of LRMs.
Yufeng Shi, Weilin Luo, Yuxiang Zhang et al.· Annual Meeting of the Associ...· 0 citations
The migration of legacy C programs to Rust has become an important direction for improving software memory safety while alleviating the high cost of manual rewriting. Leveraging large language models (LLMs) for automated C-to-Rust translation has emerged as a promising direction. However, existing LLM-based approaches remain limited. On the one hand, LLMs exhibit limited capability in identifying Rust-specific rules, and inadequate handling of Rust syntax often results in incorrect translations. On the other hand, existing LLMs often struggle to accurately capture the semantics of complex code, resulting in incorrect translations. To address these challenges, we propose a Translation fRAmework Via rule-guided reasoning and rEinforcement Learning, namely TRAVEL, consisting of two modules. The first module employs Monte Carlo Tree Search (MCTS)-based reasoning path construction guided by Rust-specific rules, steering the search toward translation steps that respect the syntactic rules that LLMs frequently violate. The second module introduces reinforcement learning that couples execution feedback with reasoning-quality signals, encouraging the model to construct reasoning paths that accurately capture program semantics, thereby ensuring that the generated Rust code preserves the intended behavior of the original C program. We evaluate TRAVEL on three datasets: xCodeEval (a public benchmark), OS-Bench (functions collected from the Linux kernel), and HW-Bench (an industrial dataset from Huawei). On xCodeEval, TRAVEL outperforms all baselines across three backbone LLMs. In particular, compared to the strongest prompting baseline IRENE, TRAVEL improves computational accuracy (CA) by 26.22% and compilation success rate (CSR) by 18.77%. On HW-Bench and OS-Bench, TRAVEL further improves CSR by 18.28% and 16.51%, respectively, while reducing unsafe rate (UR) by 13.06% and 13.08%, respectively.
Feng Luo, Jiachen Liu, Cuiyun Gao et al.· 0 citations
Large language models (LLMs) are often prompted with multi-level directives, such as system instructions and user queries, that imply a hierarchy of authority. Yet models frequently fail to enforce this structure, especially in multi-step reasoning where errors propagate across intermediate steps. Existing methods rely on oracle completions but lack verifiable reward signals or intermediate traces, limiting their applicability. We introduce a unified supervision framework that embeds programmatically verifiable checkers into synthesized instruction-conflict instances. Each instance pairs a compliance directive with a conflicting one, along with an executable veri-fier that deterministically checks output adherence. This enables alignment without oracle labels or reasoning traces, supporting both instruction-tuned and reasoning models. The framework is instantiated via a synthesis pipeline that includes unit-test–based validation, LLM-assisted repair, and a probabilistic analysis of cleaning reliability. Fine-tuning on the resulting data improves instruction hierarchy adherence and boosts safety robustness, generalizing to adversarial safety benchmarks without task-specific supervision. This highlights verifiable supervision as a scalable foundation for robust alignment. All code, dataset, and verifier pipeline are publicly available at: https://github.com/cycraft-corp/BeyondOracle .
Sian-Yao Huang, Li-Hsien Chang, Che-Yu Lin et al.· Neural Information Processin...· 4 citations