The paper investigates the training‑inference mismatch that arises when large language models are trained with on‑policy reinforcement learning but inference is performed by a separate engine. It identifies two sources of mismatch—numerical discrepancies in the engine and policy lag due to stale weights—and shows that correcting one source can worsen the other. The authors propose a discrepancy‑lag correction (DLC) that centers numerical errors, clips lag ratios, and resynchronizes the sampler, achieving stable training on Qwen models.
arXiv:2610. 11475v1 Announce Type: new Abstract: Population gradient flow is a common tool for reasoning about how neural networks adapt, including after pretraining.
The paper introduces COPC, a Coupled Off-Policy Correction method for asynchronous reinforcement learning of large language models. COPC coordinates policy-side and advantage-side corrections by combining token-level ratio masking with two-sided clipped-ratio weighting of TD residuals, addressing both policy mismatch and advantage staleness. Experiments show COPC outperforms existing asynchronous baselines on tool-integrated mathematical reasoning and search tasks, while maintaining training stability and minimal overhead.
arXiv:2605. 07284v2 Announce Type: replace Abstract: A late-layer change learned during post-training may work on the base model's earlier state, or it may depend on earlier computation learned with it.
Population gradient flow is a common tool for reasoning about how neural networks adapt, including after pretraining. We show that it can mispredict finite-batch stochastic gradient descent (SGD) qualitatively, and we trace the discrepancy to a specific mechanism.
The paper identifies a hidden mismatch in reinforcement‑learning fine‑tuning of large language models: when training uses fp32 master weights but rollouts are served from a bf16 copy produced by round‑to‑nearest conversion, most per‑step updates are too small to affect the served weights, causing a lag of 50–150 steps. The authors propose Fair Cast, a one‑line change that exports the master weights with an unbiased rounding rule (stochastic rounding or error feedback) and computes the loss on the served copy, eliminating the lag without adding hyperparameters and incurring only a modest computational cost. Across multiple models and optimizers, Fair Cast consistently outperforms round‑to‑nearest, improves checkpoint consistency, and reveals the implicit regularization effect of the lag.
The paper investigates selective on‑policy distillation, where a student model is trained only on token positions chosen by a selector. It demonstrates that the commonly used shared learning rate is not neutral: performance varies significantly with the learning rate for different selectors, leading to inconsistent comparisons. The authors attribute this selector‑rate entanglement to the selection process itself and recommend reporting the full arm‑by‑rate matrix for fair evaluation.
arXiv:2606. 08779v1 Announce Type: new Abstract: Reinforcement Learning (RL) has emerged as a pivotal post-training paradigm, yet it frequently suffers from unpredictable sub-optimum performance or even training collapses.
arXiv:2609.36484v1 Announce Type: new Abstract: On-policy distillation (OPD) trains a student to match the teacher's next-token distributions on the student's own trajectories and has yielded substan...
The paper presents an end‑to‑end MXFP8 reinforcement learning implementation on Blackwell GPUs, evaluating its benefits and limitations with Qwen3-8B and Qwen3-30B-A3B models. It finds that matching FP8 precision between sampler and learner improves consistency, that speed gains appear only for large micro‑batches, and that several weight‑synchronization correctness conditions can fail silently. The study also introduces seven explicit checks to detect such failures during training.
The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
The paper presents an analysis of asynchronous reinforcement learning for large language model post‑training, focusing on the impact of stale rollouts. It derives a convergence bound for GRPO‑style algorithms that balances the gradient estimator’s second moment and bias, and introduces a new group‑mass capping GRPO (GMC‑GRPO) method that reduces delay‑dependent bias. Experiments on Qwen3 models and reasoning benchmarks show that GMC‑GRPO outperforms existing baselines in robustness to large rollout delays.
arXiv:2606. 25556v1 Announce Type: cross Abstract: Stepwise group-based RL is an attractive way to train long-horizon LLM agents without a learned critic: it reuses multiple sampled rollouts to estimate local advantages.