Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair Methods for Off-Policy Correction

Asynchronous reinforcement learning improves rollout throughput for large language model agents by decoupling sample generation from policy optimization, but it also introduces a critical failure mode for PPO-style off-policy correction. In heterogeneous training systems, the total importance ratio should ideally be decomposed into two semantically distinct factors: a \emph{training--inference discrepancy term} that aligns inference-side and training-side distributions at the same behavior-policy version, and a \emph{policy-staleness term} that constrains the update from the historical policy to the current policy. We show that practical asynchronous pipelines with delayed updates and partial rollouts often lose the required historical training-side logits, or old logits. This missing-old-logit problem entangles discrepancy repair with staleness correction, breaks the intended semantics of decoupled correction, and makes clipping and masking thresholds interact undesirably. To address this issue, we study both exact and approximate correction routes. We propose three exact old-logit acquisition strategies: snapshot-based version tracking, a dedicated old-logit model, and synchronization via partial rollout interruption, and compare their system trade-offs. From the perspective of approximate correction, we focus on preserving the benefits of decoupled correction through a more appropriate approximate policy when exact old logits cannot be recovered at low cost, without incurring extra system overhead. Following this analysis, we adopt a revised PPO-EWMA method, which achieves significant gains in both training speed and optimization performance. Code at https://github.com/millioniron/ROLL.

SHARP: Synthesizing High-quality Aligned Reasoning Problems for Large Reasoning Models Reinforcement Learning

Training large reasoning models (LRMs) with reinforcement learning in STEM domains is hindered by the scarcity of high-quality, diverse, and verifiable problem sets. Existing synthesis methods, such as Chain-of-Thought prompting, often generate oversimplified or uncheckable data, limiting model advancement on complex tasks. To address these challenges, we introduce SHARP, a unified approach to Synthesizing High-quality Aligned Reasoning Problems for LRMs reinforcement learning with verifiable rewards (RLVR). SHARP encompasses a strategic set of self-alignment principles -- targeting graduate and Olympiad-level difficulty, rigorous logical consistency, and unambiguous, verifiable answers -- and a structured three-phase framework (Alignment, Instantiation, Inference) that ensures thematic diversity and fine-grained control over problem generation. We implement SHARP by leveraging a state-of-the-art LRM to infer and verify challenging STEM questions, then employ a reinforcement learning loop to refine the model's reasoning through verifiable reward signals. Experiments on benchmarks such as GPQA demonstrate that SHARP-augmented training substantially outperforms existing methods, markedly improving complex reasoning accuracy and pushing LRM performance closer to expert-level proficiency. Our contributions include the SHARP strategy, framework design, end-to-end implementation, and experimental evaluation of its effectiveness in elevating LRM reasoning capabilities.