Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient

Large language models (LLMs) demonstrate strong reasoning abilities in solving complex real-world problems. Yet, the internal mechanisms driving these complex reasoning behaviors remain opaque. Existing interpretability approaches targeting reasoning either identify components (e.g., neurons) correlated with special textual patterns, or rely on human-annotated contrastive pairs to derive control vectors. Consequently, current methods struggle to precisely localize complex reasoning mechanisms or capture sequential influence from model internal workings to the reasoning outputs. In this paper, built on outcome-oriented and sequential-influence-aware principles, we focus on identifying components that have sequential contribution to reasoning behavior where outcomes are cumulated by long-range effects. We propose Integrated Policy Gradient (IPG), a novel framework that attributes reasoning behaviors to model's inner components by propagating compound outcome-based signals such as post reasoning accuracy backward through model inference trajectories. Empirical evaluations demonstrate that our approach achieves more precise localization and enables reliable modulation of reasoning behaviors (e.g., reasoning capability, reasoning strength) across diverse reasoning models.

Reparameterization Flow Policy Optimization

Reparameterization Policy Gradient (RPG) has emerged as a powerful paradigm for model-based reinforcement learning, enabling high sample efficiency by backpropagating gradients through differentiable dynamics. However, prior RPG approaches have been predominantly restricted to Gaussian policies, limiting their performance and failing to leverage recent advances in generative models. In this work, we identify that flow policies, which generate actions via differentiable ODE integration, naturally align with the RPG framework, a connection not established in prior work. However, naively exploiting this synergy proves ineffective, often suffering from training instability and a lack of exploration. We propose Reparameterization Flow Policy Optimization (RFO). RFO computes policy gradients by backpropagating jointly through the flow generation process and system dynamics, unlocking high sample efficiency without requiring intractable log-likelihood calculations. RFO includes two tailored regularization terms for stability and exploration. We also propose a variant of RFO with action chunking. Extensive experiments on diverse locomotion and manipulation tasks, involving both rigid and soft bodies with state or visual inputs, demonstrate the effectiveness of RFO. Notably, on a challenging locomotion task controlling a soft-body quadruped, RFO achieves almost $2\times$ the reward of the state-of-the-art baseline.

ConsistentRFT: Reducing Visual Hallucinations in Flow-based Reinforcement Fine-Tuning

Reinforcement Fine-Tuning (RFT) on flow-based models is crucial for preference alignment. However, they often introduce visual hallucinations like over-optimized details and semantic misalignment. This work preliminarily explores why visual hallucinations arise and how to reduce them. We first investigate RFT methods from a unified perspective, and reveal the core problems stemming from two aspects, exploration and exploitation: (1) limited exploration during stochastic differential equation (SDE) rollouts, leading to an over-emphasis on local details at the expense of global semantics, and (2) trajectory imitation process inherent in policy gradient methods, distorting the model's foundational vector field and its cross-step consistency. Building on this, we propose ConsistentRFT, a general framework to mitigate these hallucinations. Specifically, we design a Dynamic Granularity Rollout (DGR) mechanism to balance exploration between global semantics and local details by dynamically scheduling different noise sources. We then introduce a Consistent Policy Gradient Optimization (CPGO) that preserves the model's consistency by aligning the current policy with a more stable prior. Extensive experiments demonstrate that ConsistentRFT significantly mitigates visual hallucinations, achieving average reductions of 49\% for low-level and 38\% for high-level perceptual hallucinations. Furthermore, ConsistentRFT outperforms other RFT methods on out-of-domain metrics, showing an improvement of 5.1\% (v.s. the baseline's decrease of -0.4\%) over FLUX1.dev. This is \href{https://xiaofeng-tan.github.io/projects/ConsistentRFT}{Project Page}.

An Approximate Ascent Approach To Prove Convergence of PPO

Proximal Policy Optimization (PPO) is among the most widely used deep reinforcement learning algorithms, yet its theoretical foundations remain incomplete. Most importantly, convergence and understanding of fundamental PPO advantages remain widely open. Under standard theory assumptions we show how PPO's policy update scheme (performing multiple epochs of minibatch updates on multi-use rollouts with a surrogate gradient) can be interpreted as approximated policy gradient ascent. We show how to control the bias accumulated by the surrogate gradients and use techniques from random reshuffling to prove a convergence theorem for PPO that sheds light on PPO's success. Additionally, we identify a previously overlooked issue in truncated Generalized Advantage Estimation commonly used in PPO. The geometric weighting scheme induces infinite mass collapse onto the longest $k$-step advantage estimator at episode boundaries. Empirical evaluations show that a simple weight correction can yield substantial improvements in environments with strong terminal signal, such as Lunar Lander.

TodyComm: Task-Oriented Dynamic Communication for Multi-Round LLM-based Multi-Agent System

Multi-round LLM-based multi-agent systems rely on effective communication structures to support collaboration across rounds. However, most existing methods employ a fixed communication topology during inference, which falls short in many realistic applications where the agents' roles may change \textit{across rounds} due to dynamic adversary, task progression, or time-varying constraints such as communication bandwidth. In this paper, we propose addressing this issue through TodyComm, a \textbf{t}ask-\textbf{o}riented \textbf{dy}namic \textbf{comm}unication algorithm. It produces behavior-driven collaboration topologies that adapt to the dynamics at each round, optimizing the utility for the task through policy gradient. Experiments on five benchmarks demonstrate that under both dynamic adversary and communications budgets, TodyComm delivers superior task effectiveness while retaining token efficiency and scalability.

$V_0$: A Generalist Value Model for Any Policy at State Zero

Policy gradient methods rely on a baseline to measure the relative advantage of an action, ensuring the model reinforces behaviors that outperform its current average capability. In the training of Large Language Models (LLMs) using Actor-Critic methods (e.g., PPO), this baseline is typically estimated by a Value Model (Critic) often as large as the policy model itself. However, as the policy continuously evolves, the value model requires expensive, synchronous incremental training to accurately track the shifting capabilities of the policy. To avoid this overhead, Group Relative Policy Optimization (GRPO) eliminates the coupled value model by using the average reward of a group of rollouts as the baseline; yet, this approach necessitates extensive sampling to maintain estimation stability. In this paper, we propose $V_0$, a Generalist Value Model capable of estimating the expected performance of any model on unseen prompts without requiring parameter updates. We reframe value estimation by treating the policy's dynamic capability as an explicit context input; specifically, we leverage a history of instruction-performance pairs to dynamically profile the model, departing from the traditional paradigm that relies on parameter fitting to perceive capability shifts. Focusing on value estimation at State Zero (i.e., the initial prompt, hence $V_0$), our model serves as a critical resource scheduler. During GRPO training, $V_0$ predicts success rates prior to rollout, allowing for efficient sampling budget allocation; during deployment, it functions as a router, dispatching instructions to the most cost-effective and suitable model. Empirical results demonstrate that $V_0$ significantly outperforms heuristic budget allocation and achieves a Pareto-optimal trade-off between performance and cost in LLM routing tasks.

Entropy-Gated Selective Policy Optimization:Token-Level Gradient Allocation for Hybrid Training of Large Language Models

Hybrid training methods for large language models combine supervised fine tuning (SFT) on expert demonstrations with reinforcement learning (RL) on model rollouts, typically at the sample level. We propose Entropy Gated Selective Policy Optimization (EGSPO), a three stage framework that extends sample level mixing with token level gradient modulation. Stage 1, SFT expert learning, establishes a reliable warm up policy using expert demonstrations with a pure SFT loss. Stage 2, RL rollout generation, samples trajectories from the current policy and computes per token predictive entropy. Stage 3, the EGSPO mechanism, applies entropy gated gradient allocation: a predictive entropy module routes high entropy tokens to full PPO updates to encourage exploration, and low entropy tokens to attenuated PPO updates to reduce variance and preserve knowledge. Critically, both branches incorporate the advantage function A_t, ensuring that incorrect trajectories receive consistent negative learning signals and preventing reinforcement of confident errors. EGSPO achieves consistent improvements on mathematical reasoning benchmarks, with gains of 3.8 percent on AIME and 2.9 percent on MATH over the CHORD phi baseline, while incurring only 3.4 percent additional computational overhead.

Flow Policy Gradients for Robot Control

Likelihood-based policy gradient methods are the dominant approach for training robot control policies from rewards. These methods rely on differentiable action likelihoods, which constrain policy outputs to simple distributions like Gaussians. In this work, we show how flow matching policy gradients -- a recent framework that bypasses likelihood computation -- can be made effective for training and fine-tuning more expressive policies in challenging robot control settings. We introduce an improved objective that enables success in legged locomotion, humanoid motion tracking, and manipulation tasks, as well as robust sim-to-real transfer on two humanoid robots. We then present ablations and analysis on training dynamics. Results show how policies can exploit the flow representation for exploration when training from scratch, as well as improved fine-tuning robustness over baselines.

Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards

Sampling efficiency is a key bottleneck in reinforcement learning with verifiable rewards. Existing group-based policy optimization methods, such as GRPO, allocate a fixed number of rollouts for all training prompts. This uniform allocation implicitly treats all prompts as equally informative, and could lead to inefficient computational budget usage and impede training progress. We introduce VIP, a Variance-Informed Predictive allocation strategy that allocates a given rollout budget to the prompts in the incumbent batch to minimize the expected gradient variance of the policy update. At each iteration, VIP uses a lightweight Gaussian process model to predict per-prompt success probabilities based on recent rollouts. These probability predictions are translated into variance estimates, which are then fed into a convex optimization problem to determine the optimal rollout allocations under a hard compute budget constraint. Empirical results show that VIP consistently improves sampling efficiency and achieves higher performance than uniform or heuristic allocation strategies in multiple benchmarks.

Causal Flow Q-Learning for Robust Offline Reinforcement Learning

Expressive policies based on flow-matching have been successfully applied in reinforcement learning (RL) more recently due to their ability to model complex action distributions from offline data. These algorithms build on standard policy gradients, which assume that there is no unmeasured confounding in the data. However, this condition does not necessarily hold for pixel-based demonstrations when a mismatch exists between the demonstrator's and the learner's sensory capabilities, leading to implicit confounding biases in offline data. We address the challenge by investigating the problem of confounded observations in offline RL from a causal perspective. We develop a novel causal offline RL objective that optimizes policies' worst-case performance that may arise due to confounding biases. Based on this new objective, we introduce a practical implementation that learns expressive flow-matching policies from confounded demonstrations, employing a deep discriminator to assess the discrepancy between the target policy and the nominal behavioral policy. Experiments across 25 pixel-based tasks demonstrate that our proposed confounding-robust augmentation procedure achieves a success rate 120\% that of confounding-unaware, state-of-the-art offline RL methods.