Jiacai Liu

Yingru Li, Jiacai Liu, Jiawei Xu et al.

Policy gradient methods for Large Language Models optimize a policy $π_θ$ via a surrogate objective computed from samples of a rollout policy $π_{\text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences -- backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness -- causing off-policy mismatch ($π_{\text{roll}} \neq π_θ$) and approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. To address this, we derive a family of bounds -- both KL-based and TV-based -- including a Pinsker-Marginal bound ($O(T^{3/2})$), a Mixed bound ($O(T)$), and an Adaptive bound that strictly generalizes the Pinsker-Marginal bound via per-position importance-ratio decomposition. Taking the minimum over all bounds yields the tightest known guarantee across all divergence regimes. Crucially, all bounds depend on the maximum token-level divergence $D_{\mathrm{KL}}^{\mathrm{tok,max}}$ (or $D_{\mathrm{TV}}^{\mathrm{tok,max}}$), a sequence-level quantity that cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences violating the trust region, enabling the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.

Yingru Li, Jiawei Xu, Jiacai Liu et al.

Reinforcement Learning (RL) for Large Language Models (LLMs) faces a fundamental tension: the numerical divergence between high-throughput inference engines and numerically precise training engines. Although these systems share the same parameters, they produce slightly different probability distributions, creating a training-inference mismatch. We prove that the bound on the log-probability divergence arising from this mismatch scales as $(1-p)$, where $p$ is the token probability. This scaling induces a highly asymmetric effect: the bound vanishes for high-probability tokens but remains significant for low-probability tokens in the distribution tail. When sampled, these tail tokens introduce systematically biased errors that accumulate over sequences, thereby destabilizing gradient estimation. Instead of applying post-hoc corrections, we propose Dynamic Vocabulary Pruning (DVP), which constrains the RL objective to a dynamically determined ''safe'' vocabulary that excludes the extreme tail. This strategy trades large, destabilizing numerical errors for a small, bounded optimization bias. We validate DVP empirically by demonstrating stable training, and theoretically by deriving strict bounds on the induced bias.

Yaxiang Zhang, Yingru Li, Jiacai Liu et al.

Reinforcement Learning (RL) for training Large Language Models is notoriously unstable. While recent studies attribute this to "training inference mismatch stemming" from inconsistent hybrid engines, standard remedies, such as Importance Sampling, might fail during extended training runs. In this work, we analyze this instability through the lens of optimization, demonstrating that gradient noise and training-inference mismatch escalate in tandem as training progresses. Meanwhile, we find that the mismatch can be effectively suppressed by shrinking the update size. Taken together, we deduce that the mismatch is not merely a static numerical discrepancy, but a dynamic failure coupled with the model's optimization. Based on this insight, we propose a simple yet effective solution: a specialized Learning Rate (LR) scheduler. Instead of pre-defined decay schedule in traditional LR scheduler, our method dynamically triggers LR decay based on response length, which we identify as a reliable early-warning signal for impending instability. Empirical evidence suggests that by reducing the learning rate as gradient noise rises, we can consistently stabilize RL training and keep the training-inference mismatch at a safe level.

Jiacai Liu, Chaojie Wang, Chris Yuhao Liu et al.

The role of reinforcement learning (RL) in enhancing the reasoning of large language models (LLMs) is becoming increasingly significant. Despite the success of RL in many scenarios, there are still many challenges in improving the reasoning of LLMs. One challenge is the sparse reward, which makes optimization difficult for RL and necessitates a large amount of data samples. Another challenge stems from the inherent instability of RL, particularly when using Actor-Critic (AC) methods to derive optimal policies, which often leads to unstable training processes. To address these issues, we introduce Direct Advantage Policy Optimization (DAPO), an novel step-level offline RL algorithm. Unlike standard alignment that rely solely outcome rewards to optimize policies (such as DPO), DAPO employs a critic function to predict the reasoning accuracy at each step, thereby generating dense signals to refine the generation strategy. Additionally, the Actor and Critic components in DAPO are trained independently, avoiding the co-training instability observed in standard AC algorithms like PPO. We train DAPO on mathematical and code query datasets and then evaluate its performance on multiple benchmarks. Our results show that DAPO can effectively enhance the mathematical and code capabilities on both SFT models and RL models, demonstrating the effectiveness of DAPO.