Mingfei Sun

Micah Carroll, Orr Paradise, Jessy Lin et al. · berkeley

Randomly masking and predicting word tokens has been a successful approach in pre-training language models for a variety of downstream tasks. In this work, we observe that the same idea also applies naturally to sequential decision-making, where many well-studied tasks like behavior cloning, offline reinforcement learning, inverse dynamics, and waypoint conditioning correspond to different sequence maskings over a sequence of states, actions, and returns. We introduce the UniMASK framework, which provides a unified way to specify models which can be trained on many different sequential decision-making tasks. We show that a single UniMASK model is often capable of carrying out many tasks with performance similar to or better than single-task models. Additionally, after fine-tuning, our UniMASK models consistently outperform comparable single-task models. Our code is publicly available at https://github.com/micahcarroll/uniMASK.

Benjamin Ellis, Jonathan Cook, Skander Moalla et al. · deepmind

The availability of challenging benchmarks has played a key role in the recent progress of machine learning. In cooperative multi-agent reinforcement learning, the StarCraft Multi-Agent Challenge (SMAC) has become a popular testbed for centralised training with decentralised execution. However, after years of sustained improvement on SMAC, algorithms now achieve near-perfect performance. In this work, we conduct new analysis demonstrating that SMAC lacks the stochasticity and partial observability to require complex *closed-loop* policies. In particular, we show that an *open-loop* policy conditioned only on the timestep can achieve non-trivial win rates for many SMAC scenarios. To address this limitation, we introduce SMACv2, a new version of the benchmark where scenarios are procedurally generated and require agents to generalise to previously unseen settings (from the same distribution) during evaluation. We also introduce the extended partial observability challenge (EPO), which augments SMACv2 to ensure meaningful partial observability. We show that these changes ensure the benchmark requires the use of *closed-loop* policies. We evaluate state-of-the-art algorithms on SMACv2 and show that it presents significant challenges not present in the original benchmark. Our analysis illustrates that SMACv2 addresses the discovered deficiencies of SMAC and can help benchmark the next generation of MARL methods. Videos of training are available at https://sites.google.com/view/smacv2.

Micah Carroll, Jessy Lin, Orr Paradise et al. · berkeley

Randomly masking and predicting word tokens has been a successful approach in pre-training language models for a variety of downstream tasks. In this work, we observe that the same idea also applies naturally to sequential decision making, where many well-studied tasks like behavior cloning, offline RL, inverse dynamics, and waypoint conditioning correspond to different sequence maskings over a sequence of states, actions, and returns. We introduce the FlexiBiT framework, which provides a unified way to specify models which can be trained on many different sequential decision making tasks. We show that a single FlexiBiT model is simultaneously capable of carrying out many tasks with performance similar to or better than specialized models. Additionally, we show that performance can be further improved by fine-tuning our general model on specific tasks of interest.

Massimiliano Patacchiola, Mingfei Sun, Katja Hofmann et al.

In this paper we explore few-shot imitation learning for control problems, which involves learning to imitate a target policy by accessing a limited set of offline rollouts. This setting has been relatively under-explored despite its relevance to robotics and control applications. State-of-the-art methods developed to tackle few-shot imitation rely on meta-learning, which is expensive to train as it requires access to a distribution over tasks (rollouts from many target policies and variations of the base environment). Given this limitation we investigate an alternative approach, fine-tuning, a family of methods that pretrain on a single dataset and then fine-tune on unseen domain-specific data. Recent work has shown that fine-tuners outperform meta-learners in few-shot image classification tasks, especially when the data is out-of-domain. Here we evaluate to what extent this is true for control problems, proposing a simple yet effective baseline which relies on two stages: (i) training a base policy online via reinforcement learning (e.g. Soft Actor-Critic) on a single base environment, (ii) fine-tuning the base policy via behavioral cloning on a few offline rollouts of the target policy. Despite its simplicity this baseline is competitive with meta-learning methods on a variety of conditions and is able to imitate target policies trained on unseen variations of the original environment. Importantly, the proposed approach is practical and easy to implement, as it does not need any complex meta-training protocol. As a further contribution, we release an open source dataset called iMuJoCo (iMitation MuJoCo) consisting of 154 variants of popular OpenAI-Gym MuJoCo environments with associated pretrained target policies and rollouts, which can be used by the community to study few-shot imitation learning and offline reinforcement learning.

Tim Pearce, Tabish Rashid, Anssi Kanervisto et al.

Diffusion models have emerged as powerful generative models in the text-to-image domain. This paper studies their application as observation-to-action models for imitating human behaviour in sequential environments. Human behaviour is stochastic and multimodal, with structured correlations between action dimensions. Meanwhile, standard modelling choices in behaviour cloning are limited in their expressiveness and may introduce bias into the cloned policy. We begin by pointing out the limitations of these choices. We then propose that diffusion models are an excellent fit for imitating human behaviour, since they learn an expressive distribution over the joint action space. We introduce several innovations to make diffusion models suitable for sequential environments; designing suitable architectures, investigating the role of guidance, and developing reliable sampling strategies. Experimentally, diffusion models closely match human demonstrations in a simulated robotic control task and a modern 3D gaming environment.

Zichuan Lin, Xiapeng Wu, Mingfei Sun et al.

Recent success in Deep Reinforcement Learning (DRL) methods has shown that policy optimization with respect to an off-policy distribution via importance sampling is effective for sample reuse. In this paper, we show that the use of importance sampling could introduce high variance in the objective estimate. Specifically, we show in a principled way that the variance of importance sampling estimate grows quadratically with importance ratios and the large ratios could consequently jeopardize the effectiveness of surrogate objective optimization. We then propose a technique called sample dropout to bound the estimation variance by dropping out samples when their ratio deviation is too high. We instantiate this sample dropout technique on representative policy optimization algorithms, including TRPO, PPO, and ESPO, and demonstrate that it consistently boosts the performance of those DRL algorithms on both continuous and discrete action controls, including MuJoCo, DMControl and Atari video games. Our code is open-sourced at \url{https://github.com/LinZichuan/sdpo.git}.

Mingfei Sun, Benjamin Ellis, Anuj Mahajan et al.

Trust Region Policy Optimization (TRPO) is an iterative method that simultaneously maximizes a surrogate objective and enforces a trust region constraint over consecutive policies in each iteration. The combination of the surrogate objective maximization and the trust region enforcement has been shown to be crucial to guarantee a monotonic policy improvement. However, solving a trust-region-constrained optimization problem can be computationally intensive as it requires many steps of conjugate gradient and a large number of on-policy samples. In this paper, we show that the trust region constraint over policies can be safely substituted by a trust-region-free constraint without compromising the underlying monotonic improvement guarantee. The key idea is to generalize the surrogate objective used in TRPO in a way that a monotonic improvement guarantee still emerges as a result of constraining the maximum advantage-weighted ratio between policies. This new constraint outlines a conservative mechanism for iterative policy optimization and sheds light on practical ways to optimize the generalized surrogate objective. We show that the new constraint can be effectively enforced by being conservative when optimizing the generalized objective function in practice. We call the resulting algorithm Trust-REgion-Free Policy Optimization (TREFree) as it is free of any explicit trust region constraints. Empirical results show that TREFree outperforms TRPO and Proximal Policy Optimization (PPO) in terms of policy performance and sample efficiency.

Haoxuan Pan, Deheng Ye, Xiaoming Duan et al.

We revisit the estimation bias in policy gradients for the discounted episodic Markov decision process (MDP) from Deep Reinforcement Learning (DRL) perspective. The objective is formulated theoretically as the expected returns discounted over the time horizon. One of the major policy gradient biases is the state distribution shift: the state distribution used to estimate the gradients differs from the theoretical formulation in that it does not take into account the discount factor. Existing discussion of the influence of this bias was limited to the tabular and softmax cases in the literature. Therefore, in this paper, we extend it to the DRL setting where the policy is parameterized and demonstrate how this bias can lead to suboptimal policies theoretically. We then discuss why the empirically inaccurate implementations with shifted state distribution can still be effective. We show that, despite such state distribution shift, the policy gradient estimation bias can be reduced in the following three ways: 1) a small learning rate; 2) an adaptive-learning-rate-based optimizer; and 3) KL regularization. Specifically, we show that a smaller learning rate, or, an adaptive learning rate, such as that used by Adam and RSMProp optimizers, makes the policy optimization robust to the bias. We further draw connections between optimizers and the optimization regularization to show that both the KL and the reverse KL regularization can significantly rectify this bias. Moreover, we provide extensive experiments on continuous control tasks to support our analysis. Our paper sheds light on how successful PG algorithms optimize policies in the DRL setting, and contributes insights into the practical issues in DRL.

Yingxiao Huo, Satya Prakash Dash, Radu Stoican et al.

Natural gradients have long been studied in deep reinforcement learning due to their fast convergence properties and covariant weight updates. However, computing natural gradients requires inversion of the Fisher Information Matrix (FIM) at each iteration, which is computationally prohibitive in nature. In this paper, we present an efficient and scalable natural policy optimization technique that leverages a rank-1 approximation to full inverse-FIM. We theoretically show that under certain conditions, a rank-1 approximation to inverse-FIM converges faster than policy gradients and, under some conditions, enjoys the same sample complexity as stochastic policy gradient methods. We benchmark our method on a diverse set of environments and show that it achieves superior performance to standard actor-critic and trust-region baselines.

Satya Prakash Dash, Hossein Abdi, Wei Pan et al.

Gradient regularization (GR) has been shown to improve the generalizability of trained models. While Natural Gradient Descent has been shown to accelerate optimization in the initial phase of training, little attention has been paid to how the training dynamics of second-order optimizers can benefit from GR. In this work, we propose Gradient-Regularized Natural Gradients (GRNG), a family of scalable second-order optimizers that integrate explicit gradient regularization with natural gradient updates. Our framework provides two complementary algorithms: a frequentist variant that avoids explicit inversion of the Fisher Information Matrix (FIM) via structured approximations, and a Bayesian variant based on a Regularized-Kalman formulation that eliminates the need for FIM inversion entirely. We establish convergence guarantees for GRNG, showing that gradient regularization improves stability and enables convergence to global minima. Empirically, we demonstrate that GRNG consistently enhances both optimization speed and generalization compared to first-order methods (SGD, AdamW) and second-order baselines (K-FAC, Sophia), with strong results on vision and language benchmarks. Our findings highlight gradient regularization as a principled and practical tool to unlock the robustness of natural gradient methods for large-scale deep learning.

Mingfei Sun

Natural policy gradients improve optimization by accounting for the geometry of distribution space, but their practical use is limited by the cost of estimating and inverting the Fisher matrix. We present Randomized Advantage Transformation (RAT), a method for estimating Tikhonov-regularized natural policy gradients via direct backpropagation. By applying the Woodbury formula, we reformulate the regularized natural policy gradients as vanilla policy gradients with a transformed advantage. RAT computes this transformation efficiently via randomized block Kaczmarz iterations on on-policy mini-batches, avoiding explicit Fisher construction, conjugate-gradient solvers, and architecture-specific approximations. We provide convergence guarantees for RAT and demonstrate empirically that it matches or exceeds established natural-gradient methods across continuous and visual control benchmarks, while remaining simple to implement and compatible with various architectures.

Fei Jiang, Jiyang Xia, Junjie Yu et al.

Quantifying the impacts of air pollution on health and climate relies on key atmospheric particle properties such as toxicity and hygroscopicity. However, these properties typically require complex observational techniques or expensive particle-resolved numerical simulations, limiting the availability of labeled data. We therefore estimate these hard-to-measure particle properties from routinely available observations (e.g., air pollutant concentrations and meteorological conditions). Because routine observations only indirectly reflect particle composition and structure, the mapping from routine observations to particle properties is noisy and input-dependent, yielding a heteroscedastic regression setting. With a limited and costly labeling budget, the central challenge is to select which samples to measure or simulate. While active learning is a natural approach, most acquisition strategies rely on predictive uncertainty. Under heteroscedastic noise, this signal conflates reducible epistemic uncertainty with irreducible aleatoric uncertainty, causing limited budgets to be wasted in noise-dominated regions. To address this challenge, we propose a confidence-aware active learning framework (CAAL) for efficient and robust sample selection in heteroscedastic settings. CAAL consists of two components: a decoupled uncertainty-aware training objective that separately optimises the predictive mean and noise level to stabilise uncertainty estimation, and a confidence-aware acquisition function that dynamically weights epistemic uncertainty using predicted aleatoric uncertainty as a reliability signal. Experiments on particle-resolved numerical simulations and real atmospheric observations show that CAAL consistently outperforms standard AL baselines. The proposed framework provides a practical and general solution for the efficient expansion of high-cost atmospheric particle property databases.

Aditya Kapoor, Yash Bhisikar, Benjamin Freed et al.

Effective communication in multi-agent reinforcement learning (MARL) is critical for success but constrained by bandwidth, yet past approaches have been limited to complex gating mechanisms that only decide \textit{whether} to communicate, not \textit{how precisely}. Learning to optimize message precision at the bit-level is fundamentally harder, as the required discretization step breaks gradient flow. We address this by generalizing Differentiable Discrete Communication Learning (DDCL), a framework for end-to-end optimization of discrete messages. Our primary contribution is an extension of DDCL to support unbounded signals, transforming it into a universal, plug-and-play layer for any MARL architecture. We verify our approach with three key results. First, through a qualitative analysis in a controlled environment, we demonstrate \textit{how} agents learn to dynamically modulate message precision according to the informational needs of the task. Second, we integrate our variant of DDCL into four state-of-the-art MARL algorithms, showing it reduces bandwidth by over an order of magnitude while matching or exceeding task performance. Finally, we provide direct evidence for the \enquote{Bitter Lesson} in MARL communication: a simple Transformer-based policy leveraging DDCL matches the performance of complex, specialized architectures, questioning the necessity of bespoke communication designs.

Zhengkai Sun, Dibyakanti Kumar, Alejandro F Frangi et al.

Transformers have revolutionized machine learning and deploying attention layers in the model is increasingly standard across a myriad of applications. Further, for large models, it is common to implement Low Rank Adaptation (LoRA), whereby a factorized parameterization of them is trained, to achieve a surprisingly beneficial accuracy-size trade-off. In this work, via a unified framework we rigorously establish trainability of such models under stochastic methods. We prove that for any mild regularization, the empirical regression loss on a attention layer and LoRA on a shallow neural net, both induce Poincaré inequality for the corresponding Gibbs' measure. Then it follows via invoking recent results that a certain SDE, which mimics the SGD, minimizes the corresponding losses. In both the cases, our first-of-its-kind results of trainability on attention and nets, do not rely on any assumptions on the data or the size of the architecture.

Jinshu Huang, Mingfei Sun, Chunlin Wu

Deep neural networks (DNNs) have significantly advanced machine learning, with model depth playing a central role in their successes. The dynamical system modeling approach has recently emerged as a powerful framework, offering new mathematical insights into the structure and learning behavior of DNNs. In this work, we establish generalization error bounds for both discrete- and continuous-time residual networks (ResNets) by combining Rademacher complexity, flow maps of dynamical systems, and the convergence behavior of ResNets in the deep-layer limit. The resulting bounds are of order $O(1/\sqrt{S})$ with respect to the number of training samples $S$, and include a structure-dependent negative term, yielding depth-uniform and asymptotic generalization bounds under milder assumptions. These findings provide a unified understanding of generalization across both discrete- and continuous-time ResNets, helping to close the gap in both the order of sample complexity and assumptions between the discrete- and continuous-time settings.

Hossein Abdi, Mingfei Sun, Wei Pan

Vision-language pre-trained models, such as CLIP, have established new benchmarks in multimodal data mining. In such models, few-shot fine-tuning is a major challenge to achieve optimal performance on both in-distribution (ID) and out-of-distribution (OOD) datasets, especially when labeled data is scarce. Most existing fine-tuning approaches rely on first-order gradient-based optimizers, which typically suffer from slow convergence, sensitivity to step-size hyperparameters, and poor generalization in OOD settings. In contrast, second-order methods utilize local curvature information of the loss landscape to adjust the update step size. This is particularly beneficial for CLIP models, whose non-convex loss functions often contain sharp critical points. In such cases, natural gradient direction can offer more substantial and efficient per-iteration updates when fine-tuning with limited data. Natural Gradient Descent (NGD) is obtained by preconditioning the standard gradient with the inverse Fisher Information Matrix (FIM), which is computationally expensive for large models. To address this, we propose a Bayesian approximation of NGD using a Kalman filter for CLIP models. Our method combines the benefits of second-order optimization with Bayesian inference, which enhances generalization while providing uncertainty quantification. Extensive experiments conducted on diverse image classification datasets demonstrate that our algorithm consistently achieves superior--or comparable--ID performance and improved OOD robustness compared to state-of-the-art baselines. To the best of our knowledge, this work represents the first successful application of Kalman filtering to fine-tuning CLIP-based models, which enables more robust and efficient learning in vision-language tasks.

Maytus Piriyajitakonkij, Mingfei Sun, Mengmi Zhang et al.

Robot navigation under visual corruption presents a formidable challenge. To address this, we propose a Test-time Adaptation (TTA) method, named as TTA-Nav, for point-goal navigation under visual corruptions. Our "plug-and-play" method incorporates a top-down decoder to a pre-trained navigation model. Firstly, the pre-trained navigation model gets a corrupted image and extracts features. Secondly, the top-down decoder produces the reconstruction given the high-level features extracted by the pre-trained model. Then, it feeds the reconstruction of a corrupted image back to the pre-trained model. Finally, the pre-trained model does forward pass again to output action. Despite being trained solely on clean images, the top-down decoder can reconstruct cleaner images from corrupted ones without the need for gradient-based adaptation. The pre-trained navigation model with our top-down decoder significantly enhances navigation performance across almost all visual corruptions in our benchmarks. Our method improves the success rate of point-goal navigation from the state-of-the-art result of 46% to 94% on the most severe corruption. This suggests its potential for broader application in robotic visual navigation. Project page: https://sites.google.com/view/tta-nav

Junjie Yu, John S. Schreck, David John Gagne et al.

Reinforcement learning (RL)-based heating, ventilation, and air conditioning (HVAC) control has emerged as a promising technology for reducing building energy consumption while maintaining indoor thermal comfort. However, the efficacy of such strategies is influenced by the background climate and their implementation may potentially alter both the indoor climate and local urban climate. This study proposes an integrated framework combining RL with an urban climate model that incorporates a building energy model, aiming to evaluate the efficacy of RL-based HVAC control across different background climates, impacts of RL strategies on indoor climate and local urban climate, and the transferability of RL strategies across cities. Our findings reveal that the reward (defined as a weighted combination of energy consumption and thermal comfort) and the impacts of RL strategies on indoor climate and local urban climate exhibit marked variability across cities with different background climates. The sensitivity of reward weights and the transferability of RL strategies are also strongly influenced by the background climate. Cities in hot climates tend to achieve higher rewards across most reward weight configurations that balance energy consumption and thermal comfort, and those cities with more varying atmospheric temperatures demonstrate greater RL strategy transferability. These findings underscore the importance of thoroughly evaluating RL-based HVAC control strategies in diverse climatic contexts. This study also provides a new insight that city-to-city learning will potentially aid the deployment of RL-based HVAC control.

Beining Zhang, Aditya Kapoor, Mingfei Sun

Multi-agent reinforcement learning (MARL) often relies on \emph{parameter sharing (PS)} to scale efficiently. However, purely shared policies can stifle each agent's unique specialization, reducing overall performance in heterogeneous environments. We propose \textbf{Low-Rank Agent-Specific Adaptation (LoRASA)}, a novel approach that treats each agent's policy as a specialized ``task'' fine-tuned from a shared backbone. Drawing inspiration from parameter-efficient transfer methods, LoRASA appends small, low-rank adaptation matrices to each layer of the shared policy, naturally inducing \emph{parameter-space sparsity} that promotes both specialization and scalability. We evaluate LoRASA on challenging benchmarks including the StarCraft Multi-Agent Challenge (SMAC) and Multi-Agent MuJoCo (MAMuJoCo), implementing it atop widely used algorithms such as MAPPO and A2PO. Across diverse tasks, LoRASA matches or outperforms existing baselines \emph{while reducing memory and computational overhead}. Ablation studies on adapter rank, placement, and timing validate the method's flexibility and efficiency. Our results suggest LoRASA's potential to establish a new norm for MARL policy parameterization: combining a shared foundation for coordination with low-rank agent-specific refinements for individual specialization.

Hossein Abdi, Mingfei Sun, Andi Zhang et al.

Training large models with millions or even billions of parameters from scratch incurs substantial computational costs. Parameter Efficient Fine-Tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), address this challenge by adapting only a reduced number of parameters to specific tasks with gradient-based optimizers. In this paper, we cast PEFT as an optimal filtering/state estimation problem and present Low-Rank Kalman Optimizer (LoKO) to estimate the optimal trainable parameters in an online manner. We leverage the low-rank decomposition in LoRA to significantly reduce matrix sizes in Kalman iterations and further capitalize on a diagonal approximation of the covariance matrix to effectively decrease computational complexity from quadratic to linear in the number of trainable parameters. Moreover, we discovered that the initialization of the covariance matrix within the Kalman algorithm and the accurate estimation of the observation noise covariance are the keys in this formulation, and we propose robust approaches that work well across a vast range of well-established computer vision and language models. Our results show that LoKO converges with fewer iterations and yields better performance models compared to commonly used optimizers with LoRA in both image classifications and language tasks. Our study opens up the possibility of leveraging the Kalman filter as an effective optimizer for the online fine-tuning of large models.

Matteo Tucat, Anirbit Mukherjee, Procheta Sen et al.

We present and analyze a novel regularized form of the gradient clipping algorithm, proving that it converges to global minima of the loss surface of deep neural networks under the squared loss, provided that the layers are of sufficient width. The algorithm presented here, dubbed $δ-$GClip, introduces a modification to gradient clipping that leads to a first-of-its-kind example of a step size scheduling for gradient descent that provably minimizes training losses of deep neural nets. We also present empirical evidence that our theoretically founded $δ-$GClip algorithm is competitive with the state-of-the-art deep learning heuristics on various neural architectures including modern transformer based architectures. The modification we do to standard gradient clipping is designed to leverage the PL* condition, a variant of the Polyak-Lojasiewicz inequality which was recently proven to be true for sufficiently wide neural networks at any depth within a neighbourhood of the initialization.

Victor M. Yeom-Song, Severi Rissanen, Arno Solin et al.

Diffusion models struggle to produce samples that respect constraints, a common requirement in scientific applications. Recent approaches have introduced regularization terms in the loss or guidance methods during sampling to enforce such constraints, but they bias the generative model away from the true data distribution. This is a problem when the constraint is misspecified, which is a common issue in scientific applications where constraint formulation is challenging. We propose to integrate guidance-inspired adjustments to the denoiser, instead of the loss or sampling loop. This achieves a soft inductive bias towards constraint-compliant samples. We show that these softly constrained denoisers exploit constraint knowledge to improve compliance over standard denoisers, while maintaining enough flexibility to deviate from it in case of misspecification with observed data.

Guorui Quan, Mingfei Sun, Manuel López-Ibáñez

The art of heuristic design has traditionally been a human pursuit. While Large Language Models (LLMs) can generate code for search heuristics, their application has largely been confined to adjusting simple functions within human-crafted frameworks, leaving their capacity for broader innovation an open question. To investigate this, we tasked an LLM with building a complete solver for the constrained 3D Packing Problem. Direct code generation quickly proved fragile, prompting us to introduce two supports: constraint scaffolding--prewritten constraint-checking code--and iterative self-correction--additional refinement cycles to repair bugs and produce a viable initial population. Notably, even within a vast search space in a greedy process, the LLM concentrated its efforts almost exclusively on refining the scoring function. This suggests that the emphasis on scoring functions in prior work may reflect not a principled strategy, but rather a natural limitation of LLM capabilities. The resulting heuristic was comparable to a human-designed greedy algorithm, and when its scoring function was integrated into a human-crafted metaheuristic, its performance rivaled established solvers, though its effectiveness waned as constraints tightened. Our findings highlight two major barriers to automated heuristic design with current LLMs: the engineering required to mitigate their fragility in complex reasoning tasks, and the influence of pretrained biases, which can prematurely narrow the search for novel solutions.

Maytus Piriyajitakonkij, Rujikorn Charakorn, Weicheng Tao et al.

Language is a powerful communicative and cognitive tool. It enables humans to express thoughts, share intentions, and reason about complex phenomena. Despite our fluency in using and understanding language, the question of how it arises and evolves over time remains unsolved. A leading hypothesis in linguistics and anthropology posits that language evolved to meet the ecological and social demands of early human cooperation. Language did not arise in isolation, but through shared survival goals. Inspired by this view, we investigate the emergence of language in multi-agent Foraging Games. These environments are designed to reflect the cognitive and ecological constraints believed to have influenced the evolution of communication. Agents operate in a shared grid world with only partial knowledge about other agents and the environment, and must coordinate to complete games like picking up high-value targets or executing temporally ordered actions. Using end-to-end deep reinforcement learning, agents learn both actions and communication strategies from scratch. We find that agents develop communication protocols with hallmark features of natural language: arbitrariness, interchangeability, displacement, cultural transmission, and compositionality. We quantify each property and analyze how different factors, such as population size, social dynamics, and temporal dependencies, shape specific aspects of the emergent language. Our framework serves as a platform for studying how language can evolve from partial observability, temporal reasoning, and cooperative goals in embodied multi-agent settings. We will release all data, code, and models publicly.

Aditya Kapoor, Kale-ab Tessera, Mayank Baranwal et al.

Credit assignmen, disentangling each agent's contribution to a shared reward, is a critical challenge in cooperative multi-agent reinforcement learning (MARL). To be effective, credit assignment methods must preserve the environment's optimal policy. Some recent approaches attempt this by enforcing return equivalence, where the sum of distributed rewards must equal the team reward. However, their guarantees are conditional on a learned model's regression accuracy, making them unreliable in practice. We introduce Temporal-Agent Reward Redistribution (TAR$^2$), an approach that decouples credit modeling from this constraint. A neural network learns unnormalized contribution scores, while a separate, deterministic normalization step enforces return equivalence by construction. We demonstrate that this method is equivalent to a valid Potential-Based Reward Shaping (PBRS), which guarantees the optimal policy is preserved regardless of model accuracy. Empirically, on challenging SMACLite and Google Research Football (GRF) benchmarks, TAR$^2$ accelerates learning and achieves higher final performance than strong baselines. These results establish our method as an effective solution for the agent-temporal credit assignment problem.

Aditya Kapoor, Sushant Swamy, Kale-ab Tessera et al.

In multi-agent environments, agents often struggle to learn optimal policies due to sparse or delayed global rewards, particularly in long-horizon tasks where it is challenging to evaluate actions at intermediate time steps. We introduce Temporal-Agent Reward Redistribution (TAR$^2$), a novel approach designed to address the agent-temporal credit assignment problem by redistributing sparse rewards both temporally and across agents. TAR$^2$ decomposes sparse global rewards into time-step-specific rewards and calculates agent-specific contributions to these rewards. We theoretically prove that TAR$^2$ is equivalent to potential-based reward shaping, ensuring that the optimal policy remains unchanged. Empirical results demonstrate that TAR$^2$ stabilizes and accelerates the learning process. Additionally, we show that when TAR$^2$ is integrated with single-agent reinforcement learning algorithms, it performs as well as or better than traditional multi-agent reinforcement learning methods.

Hao Zeng, Jiaqi Wang, Avirup Das et al.

Feasible solutions are crucial for Integer Programming (IP) since they can substantially speed up the solving process. In many applications, similar IP instances often exhibit similar structures and shared solution distributions, which can be potentially modeled by deep learning methods. Unfortunately, existing deep-learning-based algorithms, such as Neural Diving and Predict-and-search framework, are limited to generating only partial feasible solutions, and they must rely on solvers like SCIP and Gurobi to complete the solutions for a given IP problem. In this paper, we propose a novel framework that generates complete feasible solutions end-to-end. Our framework leverages contrastive learning to characterize the relationship between IP instances and solutions, and learns latent embeddings for both IP instances and their solutions. Further, the framework employs diffusion models to learn the distribution of solution embeddings conditioned on IP representations, with a dedicated guided sampling strategy that accounts for both constraints and objectives. We empirically evaluate our framework on four typical datasets of IP problems, and show that it effectively generates complete feasible solutions with a high probability (> 89.7 \%) without the reliance of Solvers and the quality of solutions is comparable to the best heuristic solutions from Gurobi. Furthermore, by integrating our method's sampled partial solutions with the CompleteSol heuristic from SCIP, the resulting feasible solutions outperform those from state-of-the-art methods across all datasets, exhibiting a 3.7 to 33.7\% improvement in the gap to optimal values, and maintaining a feasible ratio of over 99.7\% for all datasets.

Hanfang Lyu, Yuanchen Bai, Xin Liang et al.

Preference-based learning aims to align robot task objectives with human values. One of the most common methods to infer human preferences is by pairwise comparisons of robot task trajectories. Traditional comparison-based preference labeling systems seldom support labelers to digest and identify critical differences between complex trajectories recorded in videos. Our formative study (N = 12) suggests that individuals may overlook non-salient task features and establish biased preference criteria during their preference elicitation process because of partial observations. In addition, they may experience mental fatigue when given many pairs to compare, causing their label quality to deteriorate. To mitigate these issues, we propose FARPLS, a Feature-Augmented Robot trajectory Preference Labeling System. FARPLS highlights potential outliers in a wide variety of task features that matter to humans and extracts the corresponding video keyframes for easy review and comparison. It also dynamically adjusts the labeling order according to users' familiarities, difficulties of the trajectory pair, and level of disagreements. At the same time, the system monitors labelers' consistency and provides feedback on labeling progress to keep labelers engaged. A between-subjects study (N = 42, 105 pairs of robot pick-and-place trajectories per person) shows that FARPLS can help users establish preference criteria more easily and notice more relevant details in the presented trajectories than the conventional interface. FARPLS also improves labeling consistency and engagement, mitigating challenges in preference elicitation without raising cognitive loads significantly

Anuj Mahajan, Mikayel Samvelyan, Tarun Gupta et al.

Collective intelligence is a fundamental trait shared by several species of living organisms. It has allowed them to thrive in the diverse environmental conditions that exist on our planet. From simple organisations in an ant colony to complex systems in human groups, collective intelligence is vital for solving complex survival tasks. As is commonly observed, such natural systems are flexible to changes in their structure. Specifically, they exhibit a high degree of generalization when the abilities or the total number of agents changes within a system. We term this phenomenon as Combinatorial Generalization (CG). CG is a highly desirable trait for autonomous systems as it can increase their utility and deployability across a wide range of applications. While recent works addressing specific aspects of CG have shown impressive results on complex domains, they provide no performance guarantees when generalizing towards novel situations. In this work, we shed light on the theoretical underpinnings of CG for cooperative multi-agent systems (MAS). Specifically, we study generalization bounds under a linear dependence of the underlying dynamics on the agent capabilities, which can be seen as a generalization of Successor Features to MAS. We then extend the results first for Lipschitz and then arbitrary dependence of rewards on team capabilities. Finally, empirical analysis on various domains using the framework of multi-agent reinforcement learning highlights important desiderata for multi-agent algorithms towards ensuring CG.

Mingfei Sun, Sam Devlin, Jacob Beck et al.

We present trust region bounds for optimizing decentralized policies in cooperative Multi-Agent Reinforcement Learning (MARL), which holds even when the transition dynamics are non-stationary. This new analysis provides a theoretical understanding of the strong performance of two recent actor-critic methods for MARL, which both rely on independent ratios, i.e., computing probability ratios separately for each agent's policy. We show that, despite the non-stationarity that independent ratios cause, a monotonic improvement guarantee still arises as a result of enforcing the trust region constraint over all decentralized policies. We also show this trust region constraint can be effectively enforced in a principled way by bounding independent ratios based on the number of agents in training, providing a theoretical foundation for proximal ratio clipping. Finally, our empirical results support the hypothesis that the strong performance of IPPO and MAPPO is a direct result of enforcing such a trust region constraint via clipping in centralized training, and tuning the hyperparameters with regards to the number of agents, as predicted by our theoretical analysis.

Mingfei Sun, Vitaly Kurin, Guoqing Liu et al.

Proximal Policy Optimization (PPO) methods learn a policy by iteratively performing multiple mini-batch optimization epochs of a surrogate objective with one set of sampled data. Ratio clipping PPO is a popular variant that clips the probability ratios between the target policy and the policy used to collect samples. Ratio clipping yields a pessimistic estimate of the original surrogate objective, and has been shown to be crucial for strong performance. We show in this paper that such ratio clipping may not be a good option as it can fail to effectively bound the ratios. Instead, one can directly optimize the original surrogate objective for multiple epochs; the key is to find a proper condition to early stop the optimization epoch in each iteration. Our theoretical analysis sheds light on how to determine when to stop the optimization epoch, and call the resulting algorithm Early Stopping Policy Optimization (ESPO). We compare ESPO with PPO across many continuous control tasks and show that ESPO significantly outperforms PPO. Furthermore, we show that ESPO can be easily scaled up to distributed training with many workers, delivering strong performance as well.

Zhengye Yang, Mingfei Sun, Hongzhe Ye et al.

Social distancing can reduce the infection rates in respiratory pandemics such as COVID-19. Traffic intersections are particularly suitable for monitoring and evaluation of social distancing behavior in metropolises. We propose and evaluate a privacy-preserving social distancing analysis system (B-SDA), which uses bird's-eye view video recordings of pedestrians who cross traffic intersections. We devise algorithms for video pre-processing, object detection and tracking which are rooted in the known computer-vision and deep learning techniques, but modified to address the problem of detecting very small objects/pedestrians captured by a highly elevated camera. We propose a method for incorporating pedestrian grouping for detection of social distancing violations. B-SDA is used to compare pedestrian behavior based on pre-pandemic and pandemic videos in a major metropolitan area. The accomplished pedestrian detection performance is $63.0\%$ $AP_{50}$ and the tracking performance is $47.6\%$ MOTA. The social distancing violation rate of $15.6\%$ during the pandemic is notably lower than $31.4\%$ pre-pandemic baseline, indicating that pedestrians followed CDC-prescribed social distancing recommendations. The proposed system is suitable for deployment in real-world applications.

Mingfei Sun, Sam Devlin, Katja Hofmann et al.

Sample efficiency is crucial for imitation learning methods to be applicable in real-world applications. Many studies improve sample efficiency by extending adversarial imitation to be off-policy regardless of the fact that these off-policy extensions could either change the original objective or involve complicated optimization. We revisit the foundation of adversarial imitation and propose an off-policy sample efficient approach that requires no adversarial training or min-max optimization. Our formulation capitalizes on two key insights: (1) the similarity between the Bellman equation and the stationary state-action distribution equation allows us to derive a novel temporal difference (TD) learning approach; and (2) the use of a deterministic policy simplifies the TD learning. Combined, these insights yield a practical algorithm, Deterministic and Discriminative Imitation (D2-Imitation), which operates by first partitioning samples into two replay buffers and then learning a deterministic policy via off-policy reinforcement learning. Our empirical results show that D2-Imitation is effective in achieving good sample efficiency, outperforming several off-policy extension approaches of adversarial imitation on many control tasks.

Mingfei Sun, Anuj Mahajan, Katja Hofmann et al.

We present SoftDICE, which achieves state-of-the-art performance for imitation learning. SoftDICE fixes several key problems in ValueDICE, an off-policy distribution matching approach for sample-efficient imitation learning. Specifically, the objective of ValueDICE contains logarithms and exponentials of expectations, for which the mini-batch gradient estimate is always biased. Second, ValueDICE regularizes the objective with replay buffer samples when expert demonstrations are limited in number, which however changes the original distribution matching problem. Third, the re-parametrization trick used to derive the off-policy objective relies on an implicit assumption that rarely holds in training. We leverage a novel formulation of distribution matching and consider an entropy-regularized off-policy objective, which yields a completely offline algorithm called SoftDICE. Our empirical results show that SoftDICE recovers the expert policy with only one demonstration trajectory and no further on-policy/off-policy samples. SoftDICE also stably outperforms ValueDICE and other baselines in terms of sample efficiency on Mujoco benchmark tasks.

Deheng Ye, Guibin Chen, Peilin Zhao et al.

We present JueWu-SL, the first supervised-learning-based artificial intelligence (AI) program that achieves human-level performance in playing multiplayer online battle arena (MOBA) games. Unlike prior attempts, we integrate the macro-strategy and the micromanagement of MOBA-game-playing into neural networks in a supervised and end-to-end manner. Tested on Honor of Kings, the most popular MOBA at present, our AI performs competitively at the level of High King players in standard 5v5 games.

Christian Schroeder de Witt, Tarun Gupta, Denys Makoviichuk et al.

Most recently developed approaches to cooperative multi-agent reinforcement learning in the \emph{centralized training with decentralized execution} setting involve estimating a centralized, joint value function. In this paper, we demonstrate that, despite its various theoretical shortcomings, Independent PPO (IPPO), a form of independent learning in which each agent simply estimates its local value function, can perform just as well as or better than state-of-the-art joint learning approaches on popular multi-agent benchmark suite SMAC with little hyperparameter tuning. We also compare IPPO to several variants; the results suggest that IPPO's strong performance may be due to its robustness to some forms of environment non-stationarity.

Mingfei Sun, Zhenhui Peng, Meng Xia et al.

Robot Learning from Demonstration (RLfD) is a technique for robots to derive policies from instructors' examples. Although the reciprocal effects of student engagement on teacher behavior are widely recognized in the educational community, it is unclear whether the same phenomenon holds true for RLfD. To fill this gap, we first design three types of robot engagement behavior (attention, imitation, and a hybrid of the two) based on the learning literature. We then conduct, in a simulation environment, a within-subject user study to investigate the impact of different robot engagement cues on humans compared to a "without-engagement" condition. Results suggest that engagement communication significantly changes the human's estimation of the robots' capability and significantly raises their expectation towards the learning outcomes, even though we do not run actual learning algorithms in the experiments. Moreover, imitation behavior affects humans more than attention does in all metrics, while their combination has the most profound influences on humans. We also find that communicating engagement via imitation or the combined behavior significantly improve humans' perception towards the quality of demonstrations, even if all demonstrations are of the same quality.

Deheng Ye, Zhao Liu, Mingfei Sun et al.

We study the reinforcement learning problem of complex action control in the Multi-player Online Battle Arena (MOBA) 1v1 games. This problem involves far more complicated state and action spaces than those of traditional 1v1 games, such as Go and Atari series, which makes it very difficult to search any policies with human-level performance. In this paper, we present a deep reinforcement learning framework to tackle this problem from the perspectives of both system and algorithm. Our system is of low coupling and high scalability, which enables efficient explorations at large scale. Our algorithm includes several novel strategies, including control dependency decoupling, action mask, target attention, and dual-clip PPO, with which our proposed actor-critic network can be effectively trained in our system. Tested on the MOBA game Honor of Kings, our AI agent, called Tencent Solo, can defeat top professional human players in full 1v1 games.

Mingfei Sun, Xiaojuan Ma

Imitation learning targets deriving a mapping from states to actions, a.k.a. policy, from expert demonstrations. Existing methods for imitation learning typically require any actions in the demonstrations to be fully available, which is hard to ensure in real applications. Though algorithms for learning with unobservable actions have been proposed, they focus solely on state information and overlook the fact that the action sequence could still be partially available and provide useful information for policy deriving. In this paper, we propose a novel algorithm called Action-Guided Adversarial Imitation Learning (AGAIL) that learns a policy from demonstrations with incomplete action sequences, i.e., incomplete demonstrations. The core idea of AGAIL is to separate demonstrations into state and action trajectories, and train a policy with state trajectories while using actions as auxiliary information to guide the training whenever applicable. Built upon the Generative Adversarial Imitation Learning, AGAIL has three components: a generator, a discriminator, and a guide. The generator learns a policy with rewards provided by the discriminator, which tries to distinguish state distributions between demonstrations and samples generated by the policy. The guide provides additional rewards to the generator when demonstrated actions for specific states are available. We compare AGAIL to other methods on benchmark tasks and show that AGAIL consistently delivers comparable performance to the state-of-the-art methods even when the action sequence in demonstrations is only partially available.

Mingfei Sun, Yiqing Mou, Hongwen Xie et al.

Equipping social and service robots with the ability to perceive human emotional intensities during an interaction is in increasing demand. Most of existing work focuses on determining which emotion(s) participants are expressing from facial expressions but largely overlooks the emotional intensities spontaneously revealed by other social cues, especially body languages. In this paper, we present a real-time method for robots to capture fluctuations of participants' emotional intensities from their body poses. Unlike conventional joint-position-based approaches, our method adopts local joint transformations as pose descriptors which are invariant to subject body differences as well as the pose sensor positions. In addition, we use a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) architecture to take the specific emotion context into account when estimating emotional intensities from body poses. The dataset evaluation suggests that the proposed method is effective and performs better than baseline method on the test dataset. Also, a series of succeeding field tests on a physical robot demonstrates that the proposed method effectively estimates subjects emotional intensities in real-time. Furthermore, the robot equipped with our method is perceived to be more emotion-sensitive and more emotionally intelligent.