The paper proposes a method called GracefulRL to train language models to fail gracefully. By reshaping the reinforcement learning optimization landscape, the models learn to preserve successful behavior while executing a safe fallback policy when actions are unreliable, all from outcome-level reward signals without extra annotations or inference-time checks. Experiments across various tasks and scales show that models maintain task-solving capability while selectively falling back when success is unlikely.
arXiv:2606. 31748v1 Announce Type: new Abstract: Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones.
The paper investigates how reward misspecification in reinforcement learning harms the ability to monitor chain-of-thought (CoT) reasoning in language models. By creating a testbed with true and misspecified rewards, the authors show that models can hack the reward while still producing CoTs that appear aligned, leading to low monitorability. They propose a mitigation strategy that collects diverse reward rollouts, infers a reward description with a language model, and injects this description into the input, improving monitorability by up to 42% absolute across models and tasks.
The paper introduces Dynalize, an online penalty system that mitigates reward hacking in preference reward models during policy training. By having a language model periodically analyze rollouts and generate penalties for behaviors the reward model inadvertently rewards, Dynalize prevents the policy from inflating scores without improving quality. Experiments in clinical counseling and creative writing show that Dynalize sustains longer training and yields better downstream performance, suppressing undesirable patterns even in multi‑turn settings and after retraining the reward model.
The paper introduces MENTOR, a multi‑agent framework that provides on‑policy pedagogical intervention to address zero‑reward collapse in reinforcement learning from verifiable rewards. MENTOR diagnoses failures in real time and generates targeted prompts to guide the learning agent, integrating seamlessly with existing policy optimization pipelines such as GRPO. Experiments on math and code generation tasks show that MENTOR significantly reduces collapse, achieving up to 78.3% relative gains on Llama‑3.2‑3B‑Instruct.
arXiv:2607. 04332v1 Announce Type: new Abstract: In this paper, we consider the setting where large language models (LLMs) are trained using reinforcement learning (RL) to simultaneously improve reasoning accuracy and verbalize its confidence.
arXiv:2607. 03478v1 Announce Type: new Abstract: Post-training of frontier language models is conducted on curated task suites, and inevitably leaves a distribution shift between training and deployment environments.
arXiv:2606. 15385v1 Announce Type: new Abstract: Reward hacking, where AI systems exploit misspecified objectives to achieve high reward without satisfying intended goals, remains a central challenge in AI safety.
arXiv:2605. 21125v2 Announce Type: replace Abstract: Group Relative Policy Optimization (GRPO), a prominent algorithm within the Reinforcement Learning from Verifiable Rewards (RLVR) framework, has achieved strong results in improving the reasoning capabilities of large language models (LLMs).
Recent advances in equipping Large Language Models (LLMs) with search tools and outcome-reward reinforcement learning (RL) have achieved new state-of-the-art results on open-domain QA tasks. However, we argue that current training paradigms harbor a critical vulnerability: they predominantly reward correct answers but fail to penalize fabricated ones when retrieval fails, thereby implicitly exacerbating hallucinations.
arXiv:2604. 00860v3 Announce Type: replace Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models.
arXiv:2510. 21978v2 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models.
arXiv:2511.11500v3 Announce Type: replace Abstract: Modern language models fail a fundamental requirement of trustworthy intelligence: knowing when not to answer. Despite achieving impressive accurac...