Ling Pan

Bingxu Liu, Jiashun Liu, Johan Obando-Ceron et al.

While Proximal Policy Optimization (PPO) demonstrates strong performance in stationary settings, we show that its standard optimization paradigm struggles in continual and non-stationary environments. The failure does not stem from insufficient model capacity or overly restrictive clipping. Instead, PPO performs persistent, directionally inefficient local updates, which indicates a lack of geometry-aware guidance for accumulating meaningful behavioral change and ultimately hindering transitions toward new behavior patterns. Although divergence-based regularization introduces partial geometric awareness, its monotonically increasing penalties implicitly discourage large policy deviations, even when such shifts are necessary for effective adaptation. To address this limitation, we propose Gaussian Trust Region Policy Optimization (GTR), which reshapes the trust region using a Gaussian kernel. The resulting constraint is bounded and non-monotonic, providing strong local stability while progressively relaxing under sustained high-advantage updates. To further improve robustness, we introduce a Mixture Gaussian Anchor that adapts to recent policy trajectories, reducing variance induced by stale references. GTR is architecture-agnostic and achieves strong performance across games, simulated robotic control, open-world exploration, and language model post-training. These results demonstrate that geometry-aware trust-region design can be a promising direction for robust reinforcement learning in complex non-stationary environments. Our code is available at https://anonymous.4open.science/r/GTR_demo/README.md.

Minsu Kim, Joohwan Ko, Taeyoung Yun et al. · mila

GFlowNets are probabilistic models that sequentially generate compositional structures through a stochastic policy. Among GFlowNets, temperature-conditional GFlowNets can introduce temperature-based controllability for exploration and exploitation. We propose \textit{Logit-scaling GFlowNets} (Logit-GFN), a novel architectural design that greatly accelerates the training of temperature-conditional GFlowNets. It is based on the idea that previously proposed approaches introduced numerical challenges in the deep network training, since different temperatures may give rise to very different gradient profiles as well as magnitudes of the policy's logits. We find that the challenge is greatly reduced if a learned function of the temperature is used to scale the policy's logits directly. Also, using Logit-GFN, GFlowNets can be improved by having better generalization capabilities in offline learning and mode discovery capabilities in online learning, which is empirically verified in various biological and chemical tasks. Our code is available at \url{https://github.com/dbsxodud-11/logit-gfn}

Ling Pan, Dinghuai Zhang, Aaron Courville et al. · mila

The Generative Flow Network is a probabilistic framework where an agent learns a stochastic policy for object generation, such that the probability of generating an object is proportional to a given reward function. Its effectiveness has been shown in discovering high-quality and diverse solutions, compared to reward-maximizing reinforcement learning-based methods. Nonetheless, GFlowNets only learn from rewards of the terminal states, which can limit its applicability. Indeed, intermediate rewards play a critical role in learning, for example from intrinsic motivation to provide intermediate feedback even in particularly challenging sparse reward tasks. Inspired by this, we propose Generative Augmented Flow Networks (GAFlowNets), a novel learning framework to incorporate intermediate rewards into GFlowNets. We specify intermediate rewards by intrinsic motivation to tackle the exploration problem in sparse reward environments. GAFlowNets can leverage edge-based and state-based intrinsic rewards in a joint way to improve exploration. Based on extensive experiments on the GridWorld task, we demonstrate the effectiveness and efficiency of GAFlowNet in terms of convergence, performance, and diversity of solutions. We further show that GAFlowNet is scalable to a more complex and large-scale molecule generation domain, where it achieves consistent and significant performance improvement.

Can Chang, Ni Mu, Jiajun Wu et al. · stanford

A critical challenge in multi-agent reinforcement learning(MARL) is for multiple agents to efficiently accomplish complex, long-horizon tasks. The agents often have difficulties in cooperating on common goals, dividing complex tasks, and planning through several stages to make progress. We propose to address these challenges by guiding agents with programs designed for parallelization, since programs as a representation contain rich structural and semantic information, and are widely used as abstractions for long-horizon tasks. Specifically, we introduce Efficient Multi-Agent Reinforcement Learning with Parallel Program Guidance(E-MAPP), a novel framework that leverages parallel programs to guide multiple agents to efficiently accomplish goals that require planning over $10+$ stages. E-MAPP integrates the structural information from a parallel program, promotes the cooperative behaviors grounded in program semantics, and improves the time efficiency via a task allocator. We conduct extensive experiments on a series of challenging, long-horizon cooperative tasks in the Overcooked environment. Results show that E-MAPP outperforms strong baselines in terms of the completion rate, time efficiency, and zero-shot generalization ability by a large margin.

Dinghuai Zhang, Ling Pan, Ricky T. Q. Chen et al. · mila

Generative Flow Networks (GFlowNets) are a new family of probabilistic samplers where an agent learns a stochastic policy for generating complex combinatorial structure through a series of decision-making steps. Despite being inspired from reinforcement learning, the current GFlowNet framework is relatively limited in its applicability and cannot handle stochasticity in the reward function. In this work, we adopt a distributional paradigm for GFlowNets, turning each flow function into a distribution, thus providing more informative learning signals during training. By parameterizing each edge flow through their quantile functions, our proposed \textit{quantile matching} GFlowNet learning algorithm is able to learn a risk-sensitive policy, an essential component for handling scenarios with risk uncertainty. Moreover, we find that the distributional approach can achieve substantial improvement on existing benchmarks compared to prior methods due to our enhanced training algorithm, even in settings with deterministic rewards.

Ling Pan, Nikolay Malkin, Dinghuai Zhang et al. · mila

Generative Flow Networks or GFlowNets are related to Monte-Carlo Markov chain methods (as they sample from a distribution specified by an energy function), reinforcement learning (as they learn a policy to sample composed objects through a sequence of steps), generative models (as they learn to represent and sample from a distribution) and amortized variational methods (as they can be used to learn to approximate and sample from an otherwise intractable posterior, given a prior and a likelihood). They are trained to generate an object $x$ through a sequence of steps with probability proportional to some reward function $R(x)$ (or $\exp(-\mathcal{E}(x))$ with $\mathcal{E}(x)$ denoting the energy function), given at the end of the generative trajectory. Like for other RL settings where the reward is only given at the end, the efficiency of training and credit assignment may suffer when those trajectories are longer. With previous GFlowNet work, no learning was possible from incomplete trajectories (lacking a terminal state and the computation of the associated reward). In this paper, we consider the case where the energy function can be applied not just to terminal states but also to intermediate states. This is for example achieved when the energy function is additive, with terms available along the trajectory. We show how to reparameterize the GFlowNet state flow function to take advantage of the partial reward already accrued at each state. This enables a training objective that can be applied to update parameters even with incomplete trajectories. Even when complete trajectories are available, being able to obtain more localized credit and gradients is found to speed up training convergence, as demonstrated across many simulations.

Ling Pan, Dinghuai Zhang, Moksh Jain et al. · mila

Generative Flow Networks (or GFlowNets for short) are a family of probabilistic agents that learn to sample complex combinatorial structures through the lens of "inference as control". They have shown great potential in generating high-quality and diverse candidates from a given energy landscape. However, existing GFlowNets can be applied only to deterministic environments, and fail in more general tasks with stochastic dynamics, which can limit their applicability. To overcome this challenge, this paper introduces Stochastic GFlowNets, a new algorithm that extends GFlowNets to stochastic environments. By decomposing state transitions into two steps, Stochastic GFlowNets isolate environmental stochasticity and learn a dynamics model to capture it. Extensive experimental results demonstrate that Stochastic GFlowNets offer significant advantages over standard GFlowNets as well as MCMC- and RL-based approaches, on a variety of standard benchmarks with stochastic dynamics.

Ling Pan, Moksh Jain, Kanika Madan et al. · mila

Generative Flow Networks (GFlowNets) are amortized samplers that learn stochastic policies to sequentially generate compositional objects from a given unnormalized reward distribution. They can generate diverse sets of high-reward objects, which is an important consideration in scientific discovery tasks. However, as they are typically trained from a given extrinsic reward function, it remains an important open challenge about how to leverage the power of pre-training and train GFlowNets in an unsupervised fashion for efficient adaptation to downstream tasks. Inspired by recent successes of unsupervised pre-training in various domains, we introduce a novel approach for reward-free pre-training of GFlowNets. By framing the training as a self-supervised problem, we propose an outcome-conditioned GFlowNet (OC-GFN) that learns to explore the candidate space. Specifically, OC-GFN learns to reach any targeted outcomes, akin to goal-conditioned policies in reinforcement learning. We show that the pre-trained OC-GFN model can allow for a direct extraction of a policy capable of sampling from any new reward functions in downstream tasks. Nonetheless, adapting OC-GFN on a downstream task-specific reward involves an intractable marginalization over possible outcomes. We propose a novel way to approximate this marginalization by learning an amortized predictor enabling efficient fine-tuning. Extensive experimental results validate the efficacy of our approach, demonstrating the effectiveness of pre-training the OC-GFN, and its ability to swiftly adapt to downstream tasks and discover modes more efficiently. This work may serve as a foundation for further exploration of pre-training strategies in the context of GFlowNets.

Yuxiao Ye, Haoran He, Fangyuan Kong et al.

Text-guided image editing has advanced rapidly with diffusion models and unified multimodal foundation models. However, most existing methods remain confined to single-turn settings, overlooking the more realistic scenario of multi-turn in-context editing, where users iteratively refine an image through a sequence of instructions. In this setting, a model must follow each new instruction while preserving accumulated session-level constraints, challenged by two coupled failure modes: long-context dilution, where sparse textual constraints become difficult to recover from growing interleaved image-text histories, and state contamination, where earlier editing mistakes degrade subsequent generations. We introduce Edit-R2, a novel reinforcement learning post-training framework for unified multimodal models. Edit-R2 reconstructs the operative session intent, which effectively consolidates scattered historical constraints into an explicit reasoning trace before each editing turn. It further enables multi-turn RL over both reasoning and generation through a unified objective that jointly optimizes intent reconstruction generation in discrete text space and flow-matching image generation in continuous latent space, while a trajectory filtering mechanism suppresses corrupted rollouts to stabilize training under state contamination. To support systematic evaluation, we introduce MICE-Bench, a large-scale benchmark for multi-turn in-context editing with automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA) over accumulated session constraints. Experiments show that Edit-R2 substantially improves multi-turn in-context editing and achieves competitive performance compared against strong baselines.

Weixun Wang, XiaoXiao Xu, Wanhe An et al.

Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. We introduce the Agentic Learning Ecosystem (ALE), a foundational infrastructure that optimizes the production pipeline for agentic model. ALE consists of three components: ROLL, a post-training framework for weight optimization; ROCK, a sandbox environment manager for trajectory generation; and iFlow CLI, an agent framework for efficient context engineering. We release ROME, an open-source agent grounded by ALE and trained on over one million trajectories. Our approach includes data composition protocols for synthesizing complex behaviors and a novel policy optimization algorithm, Interaction-Perceptive Agentic Policy Optimization (IPA), which assigns credit over semantic interaction chunks rather than individual tokens to improve long-horizon training stability. Empirically, we evaluate ROME within a structured setting and introduce Terminal Bench Pro, a benchmark with improved scale and contamination control. ROME demonstrates strong performance across benchmarks like SWE-bench Verified and Terminal Bench, proving the effectiveness of ALE.

Zhuoran Li, Xing Wang, Ling Pan et al.

Topology impacts important network performance metrics, including link utilization, throughput and latency, and is of central importance to network operators. However, due to the combinatorial nature of network topology, it is extremely difficult to obtain an optimal solution, especially since topology planning in networks also often comes with management-specific constraints. As a result, local optimization with hand-tuned heuristic methods from human experts is often adopted in practice. Yet, heuristic methods cannot cover the global topology design space while taking into account constraints, and cannot guarantee to find good solutions. In this paper, we propose a novel deep reinforcement learning (DRL) algorithm for graph searching, called DRL-GS, for network topology optimization. DRL-GS consists of three novel components, including a verifier to validate the correctness of a generated network topology, a graph neural network (GNN) to efficiently approximate topology rating, and a DRL agent to conduct a topology search. DRL-GS can efficiently search over relatively large topology space and output topology with satisfactory performance. We conduct a case study based on a real-world network scenario, and our experimental results demonstrate the superior performance of DRL-GS in terms of both efficiency and performance.

Yiqin Tan, Pihe Hu, Ling Pan et al.

Training deep reinforcement learning (DRL) models usually requires high computation costs. Therefore, compressing DRL models possesses immense potential for training acceleration and model deployment. However, existing methods that generate small models mainly adopt the knowledge distillation-based approach by iteratively training a dense network. As a result, the training process still demands massive computing resources. Indeed, sparse training from scratch in DRL has not been well explored and is particularly challenging due to non-stationarity in bootstrap training. In this work, we propose a novel sparse DRL training framework, "the Rigged Reinforcement Learning Lottery" (RLx2), which builds upon gradient-based topology evolution and is capable of training a sparse DRL model based entirely on a sparse network. Specifically, RLx2 introduces a novel multi-step TD target mechanism with a dynamic-capacity replay buffer to achieve robust value learning and efficient topology exploration in sparse models. It also reaches state-of-the-art sparse training performance in several tasks, showing 7.5\times-20\times model compression with less than 3% performance degradation and up to 20\times and 50\times FLOPs reduction for training and inference, respectively.

Pihe Hu, Ling Pan, Yu Chen et al.

Multi-user delay constrained scheduling is important in many real-world applications including wireless communication, live streaming, and cloud computing. Yet, it poses a critical challenge since the scheduler needs to make real-time decisions to guarantee the delay and resource constraints simultaneously without prior information of system dynamics, which can be time-varying and hard to estimate. Moreover, many practical scenarios suffer from partial observability issues, e.g., due to sensing noise or hidden correlation. To tackle these challenges, we propose a deep reinforcement learning (DRL) algorithm, named Recurrent Softmax Delayed Deep Double Deterministic Policy Gradient ($\mathtt{RSD4}$), which is a data-driven method based on a Partially Observed Markov Decision Process (POMDP) formulation. $\mathtt{RSD4}$ guarantees resource and delay constraints by Lagrangian dual and delay-sensitive queues, respectively. It also efficiently tackles partial observability with a memory mechanism enabled by the recurrent neural network (RNN) and introduces user-level decomposition and node-level merging to ensure scalability. Extensive experiments on simulated/real-world datasets demonstrate that $\mathtt{RSD4}$ is robust to system dynamics and partially observable environments, and achieves superior performances over existing DRL and non-DRL-based methods.

Dilxat Muhtar, Jiashun Liu, Wei Gao et al.

Reinforcement Learning (RL) has emerged as a powerful paradigm for training LLM-based agents, yet remains limited by low sample efficiency, stemming not only from sparse outcome feedback but also from the agent's inability to leverage prior experience across episodes. While augmenting agents with historical experience offers a promising remedy, existing approaches suffer from a critical weakness: the experience distilled from history is either stored statically or fail to coevolve with the improving actor, causing a progressive misalignment between the experience and the actor's evolving capability that diminishes its utility over the course of training. Inspired by complementary learning systems in neuroscience, we present Complementary RL to achieve seamless co-evolution of an experience extractor and a policy actor within the RL optimization loop. Specifically, the actor is optimized via sparse outcome-based rewards, while the experience extractor is optimized according to whether its distilled experiences demonstrably contribute to the actor's success, thereby evolving its experience management strategy in lockstep with the actor's growing capabilities. Empirically, Complementary RL outperforms outcome-based agentic RL baselines that do not learn from experience, achieving 10% performance improvement in single-task scenarios and exhibits robust scalability in multi-task settings. These results establish Complementary RL as a paradigm for efficient experience-driven agent learning.

Ziqi Wang, Jiashun Liu, Ling Pan

Traditional continuous deep reinforcement learning (RL) algorithms employ deterministic or unimodal Gaussian actors, which cannot express complex multimodal decision distributions. This limitation can hinder their performance in diversity-critical scenarios. There have been some attempts to design online multimodal RL algorithms based on diffusion or amortized actors. However, these actors are intractable, making existing methods struggle with balancing performance, decision diversity, and efficiency simultaneously. To overcome this challenge, we first reformulate existing intractable multimodal actors within a unified framework, and prove that they can be directly optimized by policy gradient via reparameterization. Then, we propose a distance-based diversity regularization that does not explicitly require decision probabilities. We identify two diversity-critical domains, namely multi-goal achieving and generative RL, to demonstrate the advantages of multimodal policies and our method, particularly in terms of few-shot robustness. In conventional MuJoCo benchmarks, our algorithm also shows competitive performance. Moreover, our experiments highlight that the amortized actor is a promising policy model class with strong multimodal expressivity and high performance. Our code is available at https://github.com/PneuC/DrAC

Xiaoye Wang, Nicole Xi Zhang, Hongyu He et al.

Recent advancements in artificial intelligence (AI), particularly in large language models (LLMs), have unlocked significant potential to enhance the quality and efficiency of medical care. By introducing a novel way to interact with AI and data through natural language, LLMs offer new opportunities for medical practitioners, patients, and researchers. However, as AI and LLMs become more powerful and especially achieve superhuman performance in some medical tasks, public concerns over their safety have intensified. These concerns about AI safety have emerged as the most significant obstacles to the adoption of AI in medicine. In response, this review examines emerging risks in AI utilization during the LLM era. First, we explore LLM-specific safety challenges from functional and communication perspectives, addressing issues across data collection, model training, and real-world application. We then consider inherent safety problems shared by all AI systems, along with additional complications introduced by LLMs. Last, we discussed how safety issues of using AI in clinical practice and healthcare system operation would undermine trust among patient, clinicians and the public, and how to build confidence in these systems. By emphasizing the development of safe AI, we believe these technologies can be more rapidly and reliably integrated into everyday medical practice to benefit both patients and clinicians.

Zhuoran Li, Ling Pan, Longbo Huang

We present a novel Diffusion Offline Multi-agent Model (DOM2) for offline Multi-Agent Reinforcement Learning (MARL). Different from existing algorithms that rely mainly on conservatism in policy design, DOM2 enhances policy expressiveness and diversity based on diffusion. Specifically, we incorporate a diffusion model into the policy network and propose a trajectory-based data-augmentation scheme in training. These key ingredients make our algorithm more robust to environment changes and achieve significant improvements in performance, generalization and data-efficiency. Our extensive experimental results demonstrate that DOM2 outperforms existing state-of-the-art methods in multi-agent particle and multi-agent MuJoCo environments, and generalizes significantly better in shifted environments thanks to its high expressiveness and diversity. Furthermore, DOM2 shows superior data efficiency and can achieve state-of-the-art performance with $20+$ times less data compared to existing algorithms.

Haoran He, Yuxiao Ye, Jie Liu et al.

Fine-tuning diffusion models via online reinforcement learning (RL) has shown great potential for enhancing text-to-image alignment. However, since precisely specifying a ground-truth objective for visual tasks remains challenging, the models are often optimized using a proxy reward that only partially captures the true goal. This mismatch often leads to reward hacking, where proxy scores increase while real image quality deteriorates and generation diversity collapses. While common solutions add regularization against the reference policy to prevent reward hacking, they compromise sample efficiency and impede the exploration of novel, high-reward regions, as the reference policy is usually sub-optimal. To address the competing demands of sample efficiency, effective exploration, and mitigation of reward hacking, we propose Gated and Adaptive Regularization with Diversity-aware Optimization (GARDO), a versatile framework compatible with various RL algorithms. Our key insight is that regularization need not be applied universally; instead, it is highly effective to selectively penalize a subset of samples that exhibit high uncertainty. To address the exploration challenge, GARDO introduces an adaptive regularization mechanism wherein the reference model is periodically updated to match the capabilities of the online policy, ensuring a relevant regularization target. To address the mode collapse issue in RL, GARDO amplifies the rewards for high-quality samples that also exhibit high diversity, encouraging mode coverage without destabilizing the optimization process. Extensive experiments across diverse proxy rewards and hold-out unseen metrics consistently show that GARDO mitigates reward hacking and enhances generation diversity without sacrificing sample efficiency or exploration, highlighting its effectiveness and robustness.

Zihe Liu, Jiashun Liu, Yancheng He et al.

Reinforcement learning for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for employing RL techniques and a fragmented understanding of their underlying mechanisms. Additionally, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups, and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we reveal that a minimalist combination of two techniques can unlock the learning capability of critic-free policies using vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies like GRPO and DAPO.

Baoshun Tong, Haoran He, Ling Pan et al.

Vision-Language-Action (VLA) models have achieved remarkable success in robotic manipulation. However, their robustness to linguistic nuances remains a critical, under-explored safety concern, posing a significant safety risk to real-world deployment. Red teaming, or identifying environmental scenarios that elicit catastrophic behaviors, is an important step in ensuring the safe deployment of embodied AI agents. Reinforcement learning (RL) has emerged as a promising approach in automated red teaming that aims to uncover these vulnerabilities. However, standard RL-based adversaries often suffer from severe mode collapse due to their reward-maximizing nature, which tends to converge to a narrow set of trivial or repetitive failure patterns, failing to reveal the comprehensive landscape of meaningful risks. To bridge this gap, we propose a novel \textbf{D}iversity-\textbf{A}ware \textbf{E}mbodied \textbf{R}ed \textbf{T}eaming (\textbf{DAERT}) framework, to expose the vulnerabilities of VLAs against linguistic variations. Our design is based on evaluating a uniform policy, which is able to generate a diverse set of challenging instructions while ensuring its attack effectiveness, measured by execution failures in a physical simulator. We conduct extensive experiments across different robotic benchmarks against two state-of-the-art VLAs, including $π_0$ and OpenVLA. Our method consistently discovers a wider range of more effective adversarial instructions that reduce the average task success rate from 93.33\% to 5.85\%, demonstrating a scalable approach to stress-testing VLA agents and exposing critical safety blind spots before real-world deployment.

Xinran Li, Ling Pan, Jun Zhang

In multi-agent reinforcement learning (MARL), parameter sharing is commonly employed to enhance sample efficiency. However, the popular approach of full parameter sharing often leads to homogeneous policies among agents, potentially limiting the performance benefits that could be derived from policy diversity. To address this critical limitation, we introduce \emph{Kaleidoscope}, a novel adaptive partial parameter sharing scheme that fosters policy heterogeneity while still maintaining high sample efficiency. Specifically, Kaleidoscope maintains one set of common parameters alongside multiple sets of distinct, learnable masks for different agents, dictating the sharing of parameters. It promotes diversity among policy networks by encouraging discrepancy among these masks, without sacrificing the efficiencies of parameter sharing. This design allows Kaleidoscope to dynamically balance high sample efficiency with a broad policy representational capacity, effectively bridging the gap between full parameter sharing and non-parameter sharing across various environments. We further extend Kaleidoscope to critic ensembles in the context of actor-critic algorithms, which could help improve value estimations.Our empirical evaluations across extensive environments, including multi-agent particle environment, multi-agent MuJoCo and StarCraft multi-agent challenge v2, demonstrate the superior performance of Kaleidoscope compared with existing parameter sharing approaches, showcasing its potential for performance enhancement in MARL. The code is publicly available at \url{https://github.com/LXXXXR/Kaleidoscope}.

Pihe Hu, Shaolong Li, Zhuoran Li et al.

Deep Multi-agent Reinforcement Learning (MARL) relies on neural networks with numerous parameters in multi-agent scenarios, often incurring substantial computational overhead. Consequently, there is an urgent need to expedite training and enable model compression in MARL. This paper proposes the utilization of dynamic sparse training (DST), a technique proven effective in deep supervised learning tasks, to alleviate the computational burdens in MARL training. However, a direct adoption of DST fails to yield satisfactory MARL agents, leading to breakdowns in value learning within deep sparse value-based MARL models. Motivated by this challenge, we introduce an innovative Multi-Agent Sparse Training (MAST) framework aimed at simultaneously enhancing the reliability of learning targets and the rationality of sample distribution to improve value learning in sparse models. Specifically, MAST incorporates the Soft Mellowmax Operator with a hybrid TD-($λ$) schema to establish dependable learning targets. Additionally, it employs a dual replay buffer mechanism to enhance the distribution of training samples. Building upon these aspects, MAST utilizes gradient-based topology evolution to exclusively train multiple MARL agents using sparse networks. Our comprehensive experimental investigation across various value-based MARL algorithms on multiple benchmarks demonstrates, for the first time, significant reductions in redundancy of up to $20\times$ in Floating Point Operations (FLOPs) for both training and inference, with less than $3\%$ performance degradation.

Siyuan Yang, Yang Zhang, Haoran He et al.

Vision-Language-Action (VLA) models, trained via flow-matching or diffusion objectives, excel at learning complex behaviors from large-scale, multi-modal datasets (e.g., human teleoperation, scripted policies). However, since VLAs incorporate diverse data modes in the pre-training stage, and the finetuning dataset often contains demonstration data collected in a kinematically suboptimal or undesirable way, it exists redundant action modes that are irrelevant to the success action modes of the downstream task. Specifically, we observe a critical inference-time fragility among various sampled noises after supervised finetuning of pre-trained VLAs. In this paper, we attribute this instability to the distribution shift between the VLA policy and the policy induced by stable success modes of the downstream task dataset. Thus, we propose \textbf{TACO}, a test-time-scaling (TTS) framework that applies a lightweight pseudo-count estimator as a high-fidelity verifier of action chunks. The VLA models integrated with TACO can execute the actions with maximum pseudo-count from all sampled action chunks, thereby preventing distribution shifts while preserving the generalization ability of VLAs since the constraint is applied only during inference. Our method resembles the classical anti-exploration principle in offline reinforcement learning (RL), and being gradient-free, it incurs significant computational benefits compared to RL update, especially for flow or diffusion-based VLAs which are difficult to perform RL update due to denoising process. Extensive experiments across four simulation benchmarks (RoboTwin2.0, Robotwin, LIBERO, SimplerEnv) and a dual-arm platform demonstrate that our method significantly improves the inference stability and success rates in downstream-task adaptations.

Zarif Ikram, Ling Pan, Dianbo Liu

Generative Flow Networks (GFlowNets) are a family of probabilistic generative models that learn to sample compositional objects proportional to their rewards. One big challenge of GFlowNets is training them effectively when dealing with long time horizons and sparse rewards. To address this, we propose Evolution guided generative flow networks (EGFN), a simple but powerful augmentation to the GFlowNets training using Evolutionary algorithms (EA). Our method can work on top of any GFlowNets training objective, by training a set of agent parameters using EA, storing the resulting trajectories in the prioritized replay buffer, and training the GFlowNets agent using the stored trajectories. We present a thorough investigation over a wide range of toy and real-world benchmark tasks showing the effectiveness of our method in handling long trajectories and sparse rewards. We release the code at http://github.com/zarifikram/egfn.

Jiashun Liu, Johan Obando-Ceron, Han Lu et al.

Most recent RL for LLMs (RL4LLM) methods avoid explicit critics, replacing them with average advantage baselines. This shift is largely pragmatic: conventional value functions are computationally expensive to train at LLM scale and often fail under sparse rewards and long reasoning horizons. We revisit this bottleneck from an architectural perspective and introduce Asymmetric Proximal Policy Optimization (AsyPPO), a simple and scalable framework that restores the critics role while remaining efficient in large-model settings. AsyPPO employs a set of lightweight mini-critics, each trained on disjoint prompt shards. This design encourages diversity while preserving calibration, reducing value-estimation bias. Beyond robust estimation, AsyPPO leverages inter-critic uncertainty to refine the policy update: (i) masking advantages in states where critics agree and gradients add little learning signal, and (ii) filtering high-divergence states from entropy regularization, suppressing spurious exploration. After training on open-source data with only 5,000 samples, AsyPPO consistently improves learning stability and performance across multiple benchmarks over strong baselines, such as GRPO, achieving performance gains of more than six percent on Qwen3-4b-Base and about three percent on Qwen3-8b-Base and Qwen3-14b-Base over classic PPO, without additional tricks. These results highlight the importance of architectural innovations for scalable, efficient algorithms.

Dinghuai Zhang, Hanjun Dai, Nikolay Malkin et al.

Combinatorial optimization (CO) problems are often NP-hard and thus out of reach for exact algorithms, making them a tempting domain to apply machine learning methods. The highly structured constraints in these problems can hinder either optimization or sampling directly in the solution space. On the other hand, GFlowNets have recently emerged as a powerful machinery to efficiently sample from composite unnormalized densities sequentially and have the potential to amortize such solution-searching processes in CO, as well as generate diverse solution candidates. In this paper, we design Markov decision processes (MDPs) for different combinatorial problems and propose to train conditional GFlowNets to sample from the solution space. Efficient training techniques are also developed to benefit long-range credit assignment. Through extensive experiments on a variety of different CO tasks with synthetic and realistic data, we demonstrate that GFlowNet policies can efficiently find high-quality solutions. Our implementation is open-sourced at https://github.com/zdhNarsil/GFlowNet-CombOpt.

Yiqin Tan, Ling Pan, Longbo Huang

Deep reinforcement learning has achieved remarkable performance in various domains by leveraging deep neural networks for approximating value functions and policies. However, using neural networks to approximate value functions or policy functions still faces challenges, including low sample efficiency and overfitting. In this paper, we introduce OMNet, a novel learning paradigm utilizing multiple subnetworks within a single network, offering diverse outputs efficiently. We provide a systematic pipeline, including initialization, training, and sampling with OMNet. OMNet can be easily applied to various deep reinforcement learning algorithms with minimal additional overhead. Through comprehensive evaluations conducted on MuJoCo benchmark, our findings highlight OMNet's ability to strike an effective balance between performance and computational cost.

Zarif Ikram, Ling Pan, Dianbo Liu

Due to limited resources and fast economic growth, designing optimal transportation road networks with traffic simulation and validation in a cost-effective manner is vital for developing countries, where extensive manual testing is expensive and often infeasible. Current rule-based road design generators lack diversity, a key feature for design robustness. Generative Flow Networks (GFlowNets) learn stochastic policies to sample from an unnormalized reward distribution, thus generating high-quality solutions while preserving their diversity. In this work, we formulate the problem of linking incident roads to the circular junction of a roundabout by a Markov decision process, and we leverage GFlowNets as the Junction-Art road generator. We compare our method with related methods and our empirical results show that our method achieves better diversity while preserving a high validity score.

Haoran He, Chenjia Bai, Ling Pan et al.

Learning a generalist embodied agent capable of completing multiple tasks poses challenges, primarily stemming from the scarcity of action-labeled robotic datasets. In contrast, a vast amount of human videos exist, capturing intricate tasks and interactions with the physical world. Promising prospects arise for utilizing actionless human videos for pre-training and transferring the knowledge to facilitate robot policy learning through limited robot demonstrations. However, it remains a challenge due to the domain gap between humans and robots. Moreover, it is difficult to extract useful information representing the dynamic world from human videos, because of its noisy and multimodal data structure. In this paper, we introduce a novel framework to tackle these challenges, which leverages a unified discrete diffusion to combine generative pre-training on human videos and policy fine-tuning on a small number of action-labeled robot videos. We start by compressing both human and robot videos into unified video tokens. In the pre-training stage, we employ a discrete diffusion model with a mask-and-replace diffusion strategy to predict future video tokens in the latent space. In the fine-tuning stage, we harness the imagined future videos to guide low-level action learning with a limited set of robot data. Experiments demonstrate that our method generates high-fidelity future videos for planning and enhances the fine-tuned policies compared to previous state-of-the-art approaches with superior performance. Our project website is available at https://video-diff.github.io/.

Haoran He, Yang Zhang, Liang Lin et al.

Video generative models pre-trained on large-scale internet datasets have achieved remarkable success, excelling at producing realistic synthetic videos. However, they often generate clips based on static prompts (e.g., text or images), limiting their ability to model interactive and dynamic scenarios. In this paper, we propose Dynamic World Simulation (DWS), a novel approach to transform pre-trained video generative models into controllable world simulators capable of executing specified action trajectories. To achieve precise alignment between conditioned actions and generated visual changes, we introduce a lightweight, universal action-conditioned module that seamlessly integrates into any existing model. Instead of focusing on complex visual details, we demonstrate that consistent dynamic transition modeling is the key to building powerful world simulators. Building upon this insight, we further introduce a motion-reinforced loss that enhances action controllability by compelling the model to capture dynamic changes more effectively. Experiments demonstrate that DWS can be versatilely applied to both diffusion and autoregressive transformer models, achieving significant improvements in generating action-controllable, dynamically consistent videos across games and robotics domains. Moreover, to facilitate the applications of the learned world simulator in downstream tasks such as model-based reinforcement learning, we propose prioritized imagination to improve sample efficiency, demonstrating competitive performance compared with state-of-the-art methods.

Elaine Lau, Stephen Zhewen Lu, Ling Pan et al.

Generative Flow Networks (GFlowNets; GFNs) are a family of energy-based generative methods for combinatorial objects, capable of generating diverse and high-utility samples. However, consistently biasing GFNs towards producing high-utility samples is non-trivial. In this work, we leverage connections between GFNs and reinforcement learning (RL) and propose to combine the GFN policy with an action-value estimate, $Q$, to create greedier sampling policies which can be controlled by a mixing parameter. We show that several variants of the proposed method, QGFN, are able to improve on the number of high-reward samples generated in a variety of tasks without sacrificing diversity.

Jingtong Gao, Ling Pan, Yejing Wang et al.

Reinforcement Learning (RL) has emerged as a pivotal method for improving the reasoning capabilities of Large Language Models (LLMs). However, prevalent RL approaches such as Proximal Policy Optimization (PPO) and Group-Regularized Policy Optimization (GRPO) face critical limitations due to their reliance on sparse outcome-based rewards and inadequate mechanisms for incentivizing exploration. These limitations result in inefficient guidance for reasoning. Specifically, sparse rewards fail to deliver sufficient feedback, particularly for challenging problems. Furthermore, such rewards induce systematic biases that prioritize exploitation of familiar trajectories over novel solution discovery. These shortcomings critically hinder performance in complex reasoning tasks, which inherently demand iterative refinement across intermediate steps. To address these challenges, we propose an Intrinsic Motivation guidEd exploratioN meThOd foR LLM Reasoning (i-MENTOR), a method designed to deliver dense rewards and amplify exploration in the RL-based paradigm. i-MENTOR introduces three innovations: trajectory-aware exploration rewards that mitigate bias in token-level strategies while maintaining computational efficiency; error-conditioned reward allocation to ensure efficient exploration on challenging samples while intrinsically stabilizing training; and advantage-preserving integration that maintains advantage distribution integrity while incorporating exploratory guidance. Experiments across 4 public datasets demonstrate i-MENTOR's effectiveness, achieving a 22.23\% improvement on AIME 2024.

Taeyoung Yun, Dinghuai Zhang, Jinkyoo Park et al.

Recent advances in text-to-image diffusion models have achieved impressive image generation capabilities. However, it remains challenging to control the generation process with desired properties (e.g., aesthetic quality, user intention), which can be expressed as black-box reward functions. In this paper, we focus on prompt adaptation, which refines the original prompt into model-preferred prompts to generate desired images. While prior work uses reinforcement learning (RL) to optimize prompts, we observe that applying RL often results in generating similar postfixes and deterministic behaviors. To this end, we introduce \textbf{P}rompt \textbf{A}daptation with \textbf{G}FlowNets (\textbf{PAG}), a novel approach that frames prompt adaptation as a probabilistic inference problem. Our key insight is that leveraging Generative Flow Networks (GFlowNets) allows us to shift from reward maximization to sampling from an unnormalized density function, enabling both high-quality and diverse prompt generation. However, we identify that a naive application of GFlowNets suffers from mode collapse and uncovers a previously overlooked phenomenon: the progressive loss of neural plasticity in the model, which is compounded by inefficient credit assignment in sequential prompt generation. To address this critical challenge, we develop a systematic approach in PAG with flow reactivation, reward-prioritized sampling, and reward decomposition for prompt adaptation. Extensive experiments validate that PAG successfully learns to sample effective and diverse prompts for text-to-image generation. We also show that PAG exhibits strong robustness across various reward functions and transferability to different text-to-image models.

Kaixuan Jiang, Yang Liu, Weixing Chen et al.

Embodied Question Answering (EQA) is a challenging task in embodied intelligence that requires agents to dynamically explore 3D environments, actively gather visual information, and perform multi-step reasoning to answer questions. However, current EQA approaches suffer from critical limitations in exploration efficiency, dataset design, and evaluation metrics. Moreover, existing datasets often introduce biases or prior knowledge, leading to disembodied reasoning, while frontier-based exploration strategies struggle in cluttered environments and fail to ensure fine-grained exploration of task-relevant areas. To address these challenges, we construct the EXPloration-awaRe Embodied queStion anSwering Benchmark (EXPRESS-Bench), the largest dataset designed specifically to evaluate both exploration and reasoning capabilities. EXPRESS-Bench consists of 777 exploration trajectories and 2,044 question-trajectory pairs. To improve exploration efficiency, we propose Fine-EQA, a hybrid exploration model that integrates frontier-based and goal-oriented navigation to guide agents toward task-relevant regions more effectively. Additionally, we introduce a novel evaluation metric, Exploration-Answer Consistency (EAC), which ensures faithful assessment by measuring the alignment between answer grounding and exploration reliability. Extensive experimental comparisons with state-of-the-art EQA models demonstrate the effectiveness of our EXPRESS-Bench in advancing embodied exploration and question reasoning.

Haoran He, Jiajun Liang, Xintao Wang et al.

As the marginal cost of scaling computation (data and parameters) during model pre-training continues to increase substantially, test-time scaling (TTS) has emerged as a promising direction for improving generative model performance by allocating additional computation at inference time. While TTS has demonstrated significant success across multiple language tasks, there remains a notable gap in understanding the test-time scaling behaviors of image and video generative models (diffusion-based or flow-based models). Although recent works have initiated exploration into inference-time strategies for vision tasks, these approaches face critical limitations: being constrained to task-specific domains, exhibiting poor scalability, or falling into reward over-optimization that sacrifices sample diversity. In this paper, we propose \textbf{Evo}lutionary \textbf{Search} (EvoSearch), a novel, generalist, and efficient TTS method that effectively enhances the scalability of both image and video generation across diffusion and flow models, without requiring additional training or model expansion. EvoSearch reformulates test-time scaling for diffusion and flow models as an evolutionary search problem, leveraging principles from biological evolution to efficiently explore and refine the denoising trajectory. By incorporating carefully designed selection and mutation mechanisms tailored to the stochastic differential equation denoising process, EvoSearch iteratively generates higher-quality offspring while preserving population diversity. Through extensive evaluation across both diffusion and flow architectures for image and video generation tasks, we demonstrate that our method consistently outperforms existing approaches, achieves higher diversity, and shows strong generalizability to unseen evaluation metrics. Our project is available at the website https://tinnerhrhe.github.io/evosearch.

Jiashun Liu, Zihao Wu, Johan Obando-Ceron et al. · mila

Deep reinforcement learning (RL) agents frequently suffer from neuronal activity loss, which impairs their ability to adapt to new data and learn continually. A common method to quantify and address this issue is the tau-dormant neuron ratio, which uses activation statistics to measure the expressive ability of neurons. While effective for simple MLP-based agents, this approach loses statistical power in more complex architectures. To address this, we argue that in advanced RL agents, maintaining a neuron's learning capacity, its ability to adapt via gradient updates, is more critical than preserving its expressive ability. Based on this insight, we shift the statistical objective from activations to gradients, and introduce GraMa (Gradient Magnitude Neural Activity Metric), a lightweight, architecture-agnostic metric for quantifying neuron-level learning capacity. We show that GraMa effectively reveals persistent neuron inactivity across diverse architectures, including residual networks, diffusion models, and agents with varied activation functions. Moreover, resetting neurons guided by GraMa (ReGraMa) consistently improves learning performance across multiple deep RL algorithms and benchmarks, such as MuJoCo and the DeepMind Control Suite.

Jiashun Liu, Johan Obando-Ceron, Pablo Samuel Castro et al. · mila

Off-policy deep reinforcement learning (RL) typically leverages replay buffers for reusing past experiences during learning. This can help improve sample efficiency when the collected data is informative and aligned with the learning objectives; when that is not the case, it can have the effect of "polluting" the replay buffer with data which can exacerbate optimization challenges in addition to wasting environment interactions due to wasteful sampling. We argue that sampling these uninformative and wasteful transitions can be avoided by addressing the sunk cost fallacy, which, in the context of deep RL, is the tendency towards continuing an episode until termination. To address this, we propose learn to stop (LEAST), a lightweight mechanism that enables strategic early episode termination based on Q-value and gradient statistics, which helps agents recognize when to terminate unproductive episodes early. We demonstrate that our method improves learning efficiency on a variety of RL algorithms, evaluated on both the MuJoCo and DeepMind Control Suite benchmarks.

Han Lu, Zichen Liu, Shaopan Xiong et al.

Synchronous Reinforcement Learning (RL) post-training has emerged as a crucial step for enhancing Large Language Models (LLMs) with diverse capabilities. However, many systems designed to accelerate RL post-training still suffer from low resource utilization and limited scalability. We present ROLL Flash, a system that extends ROLL with native support for asynchronous RL post-training. ROLL Flash is built upon two core design principles: fine-grained parallelism and rollout-train decoupling. Guided by these principles, ROLL Flash provides flexible programming interfaces that enable a fully asynchronous training architecture and support efficient rollout mechanisms, including queue scheduling and environment-level asynchronous execution. Through comprehensive theoretical analysis and extensive experiments, we demonstrate that ROLL Flash significantly improves resource utilization and scalability over synchronous RL post-training. ROLL Flash achieves up to 2.24x speedup on RLVR tasks and 2.72x on agentic tasks, using the same GPU budget as synchronous baselines. Furthermore, we implement several popular off-policy algorithms and verify that asynchronous training can achieve performance on par with synchronous training.

Haoran He, Yuxiao Ye, Qingpeng Cai et al.

RL with Verifiable Rewards (RLVR) has emerged as a promising paradigm for improving the reasoning abilities of large language models (LLMs). Current methods rely primarily on policy optimization frameworks like PPO and GRPO, which follow generalized policy iteration that alternates between evaluating the current policy's value and improving the policy based on evaluation. While effective, they often suffer from training instability and diversity collapse, requiring complex heuristic tricks and careful tuning. We observe that standard RLVR in math reasoning can be formalized as a specialized finite-horizon Markov Decision Process with deterministic state transitions, tree-structured dynamics, and binary terminal rewards. Though large in scale, the underlying structure is simpler than general-purpose control settings for which popular RL algorithms (e.g., PPO) were developed, suggesting that several sophisticated techniques in existing methods may be reduced or even omitted. Based on this insight, we prove a surprising result: the optimal action can be recovered from the Q-function of a fixed uniformly random policy, thereby bypassing the generalized policy iteration loop and its associated heuristics. We introduce Random Policy Valuation for Diverse Reasoning (ROVER) to translate this principle into a practical and scalable algorithm for LLM math reasoning, a minimalist yet highly effective RL method that samples actions from a softmax over these uniform-policy Q-values. ROVER preserves diversity throughout training, allowing sustained exploration of multiple valid pathways. Across multiple base models and standard math reasoning benchmarks, ROVER demonstrates superior performance in both \textbf{quality} (\textbf{+8.2} on pass@1, \textbf{+16.8} on pass@256) and \textbf{diversity} (\textbf{+17.6\%}), despite its radical simplification compared to strong, complicated existing methods.

Haoran He, Emmanuel Bengio, Qingpeng Cai et al.

The Generative Flow Network (GFlowNet) is a probabilistic framework in which an agent learns a stochastic policy and flow functions to sample objects proportionally to an unnormalized reward function. A number of recent works explored connections between GFlowNets and maximum entropy (MaxEnt) RL, which modifies the standard objective of RL agents by learning an entropy-regularized objective. However, the relationship between GFlowNets and standard RL remains largely unexplored, despite the inherent similarities in their sequential decision-making nature. While GFlowNets can discover diverse solutions through specialized flow-matching objectives, connecting them can simplify their implementation through established RL principles and improve RL's diverse solution discovery capabilities. In this paper, we bridge this gap by revealing a fundamental connection between GFlowNets and one RL's most basic components -- policy evaluation. Surprisingly, we find that the value function obtained from evaluating a uniform policy is closely associated with the flow functions in GFlowNets through the lens of flow iteration under certain structural conditions. Building upon these insights, we introduce a rectified random policy evaluation (RPE) algorithm, which achieves the same reward-matching effect as GFlowNets based on simply evaluating a fixed random policy in these cases, offering a new perspective. Empirical results across extensive benchmarks demonstrate that RPE achieves competitive results compared to previous approaches, shedding light on the previously overlooked connection between (non-MaxEnt) RL and GFlowNets.

Chunhui Li, Cheng-Hao Liu, Dianbo Liu et al.

Generative Flow Networks (GFlowNets), a new family of probabilistic samplers, have recently emerged as a promising framework for learning stochastic policies that generate high-quality and diverse objects proportionally to their rewards. However, existing GFlowNets often suffer from low data efficiency due to the direct parameterization of edge flows or reliance on backward policies that may struggle to scale up to large action spaces. In this paper, we introduce Bifurcated GFlowNets (BN), a novel approach that employs a bifurcated architecture to factorize the flows into separate representations for state flows and edge-based flow allocation. This factorization enables BN to learn more efficiently from data and better handle large-scale problems while maintaining the convergence guarantee. Through extensive experiments on standard evaluation benchmarks, we demonstrate that BN significantly improves learning efficiency and effectiveness compared to strong baselines.

Haoran He, Can Chang, Huazhe Xu et al.

Generative Flow Networks (GFlowNets), a new family of probabilistic samplers, have demonstrated remarkable capabilities to generate diverse sets of high-reward candidates, in contrast to standard return maximization approaches (e.g., reinforcement learning) which often converge to a single optimal solution. Recent works have focused on developing goal-conditioned GFlowNets, which aim to train a single GFlowNet capable of achieving different outcomes as the task specifies. However, training such models is challenging due to extremely sparse rewards, particularly in high-dimensional problems. Moreover, previous methods suffer from the limited coverage of explored trajectories during training, which presents more pronounced challenges when only offline data is available. In this work, we propose a novel method called \textbf{R}etrospective \textbf{B}ackward \textbf{S}ynthesis (\textbf{RBS}) to address these critical problems. Specifically, RBS synthesizes new backward trajectories in goal-conditioned GFlowNets to enrich training trajectories with enhanced quality and diversity, thereby introducing copious learnable signals for effectively tackling the sparse reward problem. Extensive empirical results show that our method improves sample efficiency by a large margin and outperforms strong baselines on various standard evaluation benchmarks.

Haoran He, Peilin Wu, Chenjia Bai et al.

Reinforcement Learning (RL) has recently achieved remarkable success in robotic control. However, most works in RL operate in simulated environments where privileged knowledge (e.g., dynamics, surroundings, terrains) is readily available. Conversely, in real-world scenarios, robot agents usually rely solely on local states (e.g., proprioceptive feedback of robot joints) to select actions, leading to a significant sim-to-real gap. Existing methods address this gap by either gradually reducing the reliance on privileged knowledge or performing a two-stage policy imitation. However, we argue that these methods are limited in their ability to fully leverage the available privileged knowledge, resulting in suboptimal performance. In this paper, we formulate the sim-to-real gap as an information bottleneck problem and therefore propose a novel privileged knowledge distillation method called the Historical Information Bottleneck (HIB). In particular, HIB learns a privileged knowledge representation from historical trajectories by capturing the underlying changeable dynamic information. Theoretical analysis shows that the learned privileged knowledge representation helps reduce the value discrepancy between the oracle and learned policies. Empirical experiments on both simulated and real-world tasks demonstrate that HIB yields improved generalizability compared to previous methods. Videos of real-world experiments are available at https://sites.google.com/view/history-ib .

Ling Pan, Longbo Huang, Tengyu Ma et al.

Conservatism has led to significant progress in offline reinforcement learning (RL) where an agent learns from pre-collected datasets. However, as many real-world scenarios involve interaction among multiple agents, it is important to resolve offline RL in the multi-agent setting. Given the recent success of transferring online RL algorithms to the multi-agent setting, one may expect that offline RL algorithms will also transfer to multi-agent settings directly. Surprisingly, we empirically observe that conservative offline RL algorithms do not work well in the multi-agent setting -- the performance degrades significantly with an increasing number of agents. Towards mitigating the degradation, we identify a key issue that non-concavity of the value function makes the policy gradient improvements prone to local optima. Multiple agents exacerbate the problem severely, since the suboptimal policy by any agent can lead to uncoordinated global failure. Following this intuition, we propose a simple yet effective method, Offline Multi-Agent RL with Actor Rectification (OMAR), which combines the first-order policy gradients and zeroth-order optimization methods to better optimize the conservative value functions over the actor parameters. Despite the simplicity, OMAR achieves state-of-the-art results in a variety of multi-agent control tasks.

Ling Pan, Tabish Rashid, Bei Peng et al.

Tackling overestimation in $Q$-learning is an important problem that has been extensively studied in single-agent reinforcement learning, but has received comparatively little attention in the multi-agent setting. In this work, we empirically demonstrate that QMIX, a popular $Q$-learning algorithm for cooperative multi-agent reinforcement learning (MARL), suffers from a more severe overestimation in practice than previously acknowledged, and is not mitigated by existing approaches. We rectify this with a novel regularization-based update scheme that penalizes large joint action-values that deviate from a baseline and demonstrate its effectiveness in stabilizing learning. Furthermore, we propose to employ a softmax operator, which we efficiently approximate in a novel way in the multi-agent setting, to further reduce the potential overestimation bias. Our approach, Regularized Softmax (RES) Deep Multi-Agent $Q$-Learning, is general and can be applied to any $Q$-learning based MARL algorithm. We demonstrate that, when applied to QMIX, RES avoids severe overestimation and significantly improves performance, yielding state-of-the-art results in a variety of cooperative multi-agent tasks, including the challenging StarCraft II micromanagement benchmarks.

Ling Pan, Qingpeng Cai, Longbo Huang

A widely-used actor-critic reinforcement learning algorithm for continuous control, Deep Deterministic Policy Gradients (DDPG), suffers from the overestimation problem, which can negatively affect the performance. Although the state-of-the-art Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm mitigates the overestimation issue, it can lead to a large underestimation bias. In this paper, we propose to use the Boltzmann softmax operator for value function estimation in continuous control. We first theoretically analyze the softmax operator in continuous action space. Then, we uncover an important property of the softmax operator in actor-critic algorithms, i.e., it helps to smooth the optimization landscape, which sheds new light on the benefits of the operator. We also design two new algorithms, Softmax Deep Deterministic Policy Gradients (SD2) and Softmax Deep Double Deterministic Policy Gradients (SD3), by building the softmax operator upon single and double estimators, which can effectively improve the overestimation and underestimation bias. We conduct extensive experiments on challenging continuous control tasks, and results show that SD3 outperforms state-of-the-art methods.

Ling Pan, Qingpeng Cai, Longbo Huang

Recent years have witnessed a tremendous improvement of deep reinforcement learning. However, a challenging problem is that an agent may suffer from inefficient exploration, particularly for on-policy methods. Previous exploration methods either rely on complex structure to estimate the novelty of states, or incur sensitive hyper-parameters causing instability. We propose an efficient exploration method, Multi-Path Policy Optimization (MPPO), which does not incur high computation cost and ensures stability. MPPO maintains an efficient mechanism that effectively utilizes a population of diverse policies to enable better exploration, especially in sparse environments. We also give a theoretical guarantee of the stable performance. We build our scheme upon two widely-adopted on-policy methods, the Trust-Region Policy Optimization algorithm and Proximal Policy Optimization algorithm. We conduct extensive experiments on several MuJoCo tasks and their sparsified variants to fairly evaluate the proposed method. Results show that MPPO significantly outperforms state-of-the-art exploration methods in terms of both sample efficiency and final performance.

Qingpeng Cai, Ling Pan, Pingzhong Tang

Reinforcement learning algorithms such as the deep deterministic policy gradient algorithm (DDPG) has been widely used in continuous control tasks. However, the model-free DDPG algorithm suffers from high sample complexity. In this paper we consider the deterministic value gradients to improve the sample efficiency of deep reinforcement learning algorithms. Previous works consider deterministic value gradients with the finite horizon, but it is too myopic compared with infinite horizon. We firstly give a theoretical guarantee of the existence of the value gradients in this infinite setting. Based on this theoretical guarantee, we propose a class of the deterministic value gradient algorithm (DVG) with infinite horizon, and different rollout steps of the analytical gradients by the learned model trade off between the variance of the value gradients and the model bias. Furthermore, to better combine the model-based deterministic value gradient estimators with the model-free deterministic policy gradient estimator, we propose the deterministic value-policy gradient (DVPG) algorithm. We finally conduct extensive experiments comparing DVPG with state-of-the-art methods on several standard continuous control benchmarks. Results demonstrate that DVPG substantially outperforms other baselines.

Ling Pan, Qingpeng Cai, Qi Meng et al.

Value function estimation is an important task in reinforcement learning, i.e., prediction. The Boltzmann softmax operator is a natural value estimator and can provide several benefits. However, it does not satisfy the non-expansion property, and its direct use may fail to converge even in value iteration. In this paper, we propose to update the value function with dynamic Boltzmann softmax (DBS) operator, which has good convergence property in the setting of planning and learning. Experimental results on GridWorld show that the DBS operator enables better estimation of the value function, which rectifies the convergence issue of the softmax operator. Finally, we propose the DBS-DQN algorithm by applying dynamic Boltzmann softmax updates in deep Q-network, which outperforms DQN substantially in 40 out of 49 Atari games.

Qingpeng Cai, Ling Pan, Pingzhong Tang

We study a reinforcement learning setting, where the state transition function is a convex combination of a stochastic continuous function and a deterministic function. Such a setting generalizes the widely-studied stochastic state transition setting, namely the setting of deterministic policy gradient (DPG). We firstly give a simple example to illustrate that the deterministic policy gradient may be infinite under deterministic state transitions, and introduce a theoretical technique to prove the existence of the policy gradient in this generalized setting. Using this technique, we prove that the deterministic policy gradient indeed exists for a certain set of discount factors, and further prove two conditions that guarantee the existence for all discount factors. We then derive a closed form of the policy gradient whenever exists. Furthermore, to overcome the challenge of high sample complexity of DPG in this setting, we propose the Generalized Deterministic Policy Gradient (GDPG) algorithm. The main innovation of the algorithm is a new method of applying model-based techniques to the model-free algorithm, the deep deterministic policy gradient algorithm (DDPG). GDPG optimize the long-term rewards of the model-based augmented MDP subject to a constraint that the long-rewards of the MDP is less than the original one. We finally conduct extensive experiments comparing GDPG with state-of-the-art methods and the direct model-based extension method of DDPG on several standard continuous control benchmarks. Results demonstrate that GDPG substantially outperforms DDPG, the model-based extension of DDPG and other baselines in terms of both convergence and long-term rewards in most environments.