强化学习

Recent embodied agents are primarily built based on reinforcement learning (RL) or large language models (LLMs). Among them, RL agents are efficient for deployment but only perform very few tasks. By contrast, giant LLM agents (often more than 1000B parameters) present strong generalization while demanding enormous computing resources. In this work, we combine their advantages while avoiding the drawbacks by conducting the proposed referee RL on our developed large auto-regressive model (LARM). Specifically, LARM is built upon a lightweight LLM (fewer than 5B parameters) and directly outputs the next action to execute rather than text. We mathematically reveal that classic RL feedbacks vanish in long-horizon embodied exploration and introduce a giant LLM based referee to handle this reward vanishment during training LARM. In this way, LARM learns to complete diverse open-world tasks without human intervention. Especially, LARM successfully harvests enchanted diamond equipment in Minecraft, which demands significantly longer decision-making chains than the highest achievements of prior best methods.

A fundamental challenge in formal theorem proving by LLMs is the lack of high-quality training data. Although reinforcement learning or expert iteration partially mitigates this issue by alternating between LLM generating proofs and finetuning them on correctly generated ones, performance quickly plateaus due to the scarcity of correct proofs (sparse rewards). To keep improving the models with limited data, we draw inspiration from mathematicians, who continuously develop new results, partly by proposing novel conjectures or exercises (which are often variants of known results) and attempting to solve them. We design the Self-play Theorem Prover (STP) that simultaneously takes on two roles, conjecturer and prover, each providing training signals to the other. The conjecturer is trained iteratively on previously generated conjectures that are barely provable by the current prover, which incentivizes it to generate increasingly challenging conjectures over time. The prover attempts to prove the conjectures with standard expert iteration. We evaluate STP with both Lean and Isabelle formal versifiers. With 51.3 billion tokens generated during the training in Lean, STP proves 28.5% of the statements in the LeanWorkbook dataset, doubling the previous best result of 13.1% achieved through expert iteration. The final model achieves state-of-the-art performance among whole-proof generation methods on miniF2F-test (65.0\%), ProofNet-test (23.9\%) and PutnamBench (8/644) with pass@3200.

The Dreamer algorithm has recently obtained remarkable performance across diverse environment domains by training powerful agents with simulated trajectories. However, the compressed nature of its world model's latent space can result in the loss of crucial information, negatively affecting the agent's performance. Recent approaches, such as $\Delta$-IRIS and DIAMOND, address this limitation by training more accurate world models. However, these methods require training agents directly from pixels, which reduces training efficiency and prevents the agent from benefiting from the inner representations learned by the world model. In this work, we propose an alternative approach to world modeling that is both accurate and efficient. We introduce EMERALD (Efficient MaskEd latent tRAnsformer worLD model), a world model using a spatial latent state with MaskGIT predictions to generate accurate trajectories in latent space and improve the agent performance. On the Crafter benchmark, EMERALD achieves new state-of-the-art performance, becoming the first method to surpass human experts performance within 10M environment steps. Our method also succeeds to unlock all 22 Crafter achievements at least once during evaluation.

Zero-shot coordination (ZSC), the ability to adapt to a new partner in a cooperative task, is a critical component of human-compatible AI. While prior work has focused on training agents to cooperate on a single task, these specialized models do not generalize to new tasks, even if they are highly similar. Here, we study how reinforcement learning on a **distribution of environments with a single partner** enables learning general cooperative skills that support ZSC with **many new partners on many new problems**. We introduce *two* Jax-based, procedural generators that create billions of solvable coordination challenges. We develop a new paradigm called **Cross-Environment Cooperation (CEC)**, and show that it outperforms competitive baselines quantitatively and qualitatively when collaborating with real people. Our findings suggest that learning to collaborate across many unique scenarios encourages agents to develop general norms, which prove effective for collaboration with different partners. Together, our results suggest a new route toward designing generalist cooperative agents capable of interacting with humans without requiring human data.

Policy gradient methods are notorious for having a large variance and high sample complexity. To mitigate this, we introduce SoftTreeMax---a generalization of softmax that employs planning. In SoftTreeMax, we extend the traditional logits with the multi-step discounted cumulative reward, topped with the logits of future states. We analyze SoftTreeMax and explain how tree expansion helps to reduce its gradient variance. We prove that the variance depends on the chosen tree-expansion policy. Specifically, we show that the closer the induced transitions are to being state-independent, the stronger the variance decay. With approximate forward models, we prove that the resulting gradient bias diminishes with the approximation error while retaining the same variance reduction. Ours is the first result to bound the gradient bias for an approximate model. In a practical implementation of SoftTreeMax we utilize a parallel GPU-based simulator for fast and efficient tree expansion. Using this implementation in Atari, we show that SoftTreeMax reduces the gradient variance by three orders of magnitude. This leads to better sample complexity and improved performance compared to distributed PPO.

Traditional approaches to studying decision-making in neuroscience focus on simplified behavioral tasks where animals perform repetitive, stereotyped actions to receive explicit rewards. While informative, these methods constrain our understanding of decision-making to short timescale behaviors driven by explicit goals. In natural environments, animals exhibit more complex, long-term behaviors driven by intrinsic motivations that are often unobservable. Recent works in time-varying inverse reinforcement learning (IRL) aim to capture shifting motivations in long-term, freely moving behaviors. However, a crucial challenge remains: animals make decisions based on their history, not just their current state. To address this, we introduce SWIRL (SWitching IRL), a novel framework that extends traditional IRL by incorporating time-varying, history-dependent reward functions. SWIRL models long behavioral sequences as transitions between short-term decision-making processes, each governed by a unique reward function. SWIRL incorporates biologically plausible history dependency to capture how past decisions and environmental contexts shape behavior, offering a more accurate description of animal decision-making. We apply SWIRL to simulated and real-world animal behavior datasets and show that it outperforms models lacking history dependency, both quantitatively and qualitatively. This work presents the first IRL model to incorporate history-dependent policies and rewards to advance our understanding of complex, naturalistic decision-making in animals.

Recent works have demonstrated the vulnerability of Deep Reinforcement Learning (DRL) algorithms against training-time, backdoor poisoning attacks. The objectives of these attacks are twofold: induce pre-determined, adversarial behavior in the agent upon observing a fixed trigger during deployment while allowing the agent to solve its intended task during training. Prior attacks assume arbitrary control over the agent's rewards, inducing values far outside the environment's natural constraints. This results in brittle attacks that fail once the proper reward constraints are enforced. Thus, in this work we propose a new class of backdoor attacks against DRL which are the first to achieve state of the art performance under strict reward constraints. These ``inception'' attacks manipulate the agent's training data -- inserting the trigger into prior observations and replacing high return actions with those of the targeted adversarial behavior. We formally define these attacks and prove they achieve both adversarial objectives against arbitrary Markov Decision Processes (MDP). Using this framework we devise an online inception attack which achieves an 100% attack success rate on multiple environments under constrained rewards while minimally impacting the agent's task performance.

Recent advancements in Large Language Models (LLMs) and Reinforcement Learning (RL) have shown significant promise in decision-making tasks. Nevertheless, for large-scale industrial decision problems, both approaches face distinct challenges: LLMs lack real-time long-sequence decision-making capabilities, while RL struggles with sample efficiency in vast action spaces. To bridge this gap, we propose **A**gents **C**o-**E**volution (ACE), a synergistic framework between LLMs and RL agent for large-scale decision-making scenarios. ACE introduces a dual-role trajectory refinement mechanism where LLMs act as both Policy Actor and Value Critic during RL's training: the Actor refines suboptimal actions via multi-step reasoning and environment validation, while the Critic performs temporal credit assignment through trajectory-level reward shaping. Concurrently, RL agent enhance LLMs' task-specific decision-making via prioritized experience replay. Through extensive experiments across multiple power grid operation challenges with action spaces exceeding 60K discrete actions, ACE demonstrates superior performance over existing RL methods and LLM-based methods.

Self-play has powered breakthroughs in two-player and multi-player games. Here we show that self-play is a surprisingly effective strategy in another domain. We show that robust and naturalistic driving emerges entirely from self-play in simulation at unprecedented scale -- 1.6 billion km of driving. This is enabled by Gigaflow, a batched simulator that can synthesize and train on 42 years of subjective driving experience per hour on a single 8-GPU node. The resulting policy achieves state-of-the-art performance on three independent autonomous driving benchmarks. The policy outperforms the prior state of the art when tested on recorded real-world scenarios, amidst human drivers, without ever seeing human data during training. The policy is realistic when assessed against human references and achieves unprecedented robustness, averaging 17.5 years of continuous driving between incidents in simulation.

The quest for Continual Learning (CL) seeks to empower neural networks with the ability to learn and adapt incrementally. Central to this pursuit is addressing the stability-plasticity dilemma, which involves striking a balance between two conflicting objectives: preserving previously learned knowledge and acquiring new knowledge. While numerous CL methods aim to achieve this trade-off, they often overlook the impact of network architecture on stability and plasticity, restricting the trade-off to the parameter level. In this paper, we delve into the conflict between stability and plasticity at the architectural level. We reveal that under an equal parameter constraint, deeper networks exhibit better plasticity, while wider networks are characterized by superior stability. To address this architectural-level dilemma, we introduce a novel framework denoted Dual-Arch, which serves as a plug-in component for CL. This framework leverages the complementary strengths of two distinct and independent networks: one dedicated to plasticity and the other to stability. Each network is designed with a specialized and lightweight architecture, tailored to its respective objective. Extensive experiments demonstrate that Dual-Arch enhances the performance of existing CL methods while being up to 87% more compact in terms of parameters.

Behavioral diversity, expert imitation, fairness, safety goals and others give rise to preferences in sequential decision making domains that do not decompose additively across time. We introduce the class of convex Markov games that allow general convex preferences over occupancy measures. Despite infinite time horizon and strictly higher generality than Markov games, pure strategy Nash equilibria exist. Furthermore, equilibria can be approximated empirically by performing gradient descent on an upper bound of exploitability. Our experiments reveal novel solutions to classic repeated normal-form games, find fair solutions in a repeated asymmetric coordination game, and prioritize safe long-term behavior in a robot warehouse environment. In the prisoner's dilemma, our algorithm leverages transient imitation to find a policy profile that deviates from observed human play only slightly, yet achieves higher per-player utility while also being three orders of magnitude less exploitable.

This study presents an innovative reinforcement learning (RL) control approach to facilitate soft exosuit-assisted human walking. Our goal is to address the ongoing challenges in developing reliable RL-based methods for controlling physical devices. To overcome key obstacles—such as limited data, the absence of a simulator for human-robot interaction during walking, the need for low computational overhead in real-time deployment, and the demand for rapid adaptation to achieve personalized control while ensuring human safety—we propose an online Adaptation from an offline Imitating Expert Policy (AIP) approach. Our offline learning mimics human expert actions through real human walking demonstrations without robot assistance. The resulted policy is then used to initialize online actor-critic learning, the goal of which is to optimally personalize robot assistance. In addition to being fast and robust, our online RL method also posses important properties such as learning convergence, dynamic stability, and solution optimality. We have successfully demonstrated our simple and robust framework for safe robot control on all five tested human participants, without selectively presenting results. The qualitative performance guarantees provided by our online RL, along with the consistent experimental validation of AIP control, represent the first demonstration of online adaptation for softsuit control personalization and serve as important evidence for the use of online RL in controlling a physical device to solve a real-life problem.

Reinforcement learning (RL) using world models has found significant recent successes. However, when a sudden change to world mechanics or properties occurs then agent performance and reliability can dramatically decline. We refer to the sudden change in visual properties or state transitions as novelties. Implementing novelty detection within generated world model frameworks is a crucial task for protecting the agent when deployed. In this paper, we propose straightforward bounding approaches to incorporate novelty detection into world model RL agents by utilizing the misalignment of the world model's hallucinated states and the true observed states as a novelty score. We provide effective approaches to detecting novelties in a distribution of transitions learned by an agent in a world model. Finally, we show the advantage of our work in a novel environment compared to traditional machine learning novelty detection methods as well as currently accepted RL-focused novelty detection algorithms.

Real-world human decision-making often relies on strategic planning, where *high-level* goals guide the formulation of sub-goals and subsequent actions, as evidenced by domains such as healthcare, business, and urban policy. Despite notable successes in controlled settings, conventional reinforcement learning (RL) follows a *bottom-up* paradigm, which can struggle to adapt to real-world complexities such as sparse rewards and limited exploration budgets. While methods like hierarchical RL and environment shaping provide partial solutions, they frequently rely on either ad-hoc designs (e.g. choose the set of high-level actions) or purely data-driven discovery of high-level actions that still requires significant exploration. In this paper, we introduce a *top-down* framework for RL that explicitly leverages *human-like strategy* to reduce sample complexity, guide exploration, and enable high-level decision-making. We first formalize the *Strategy Problem*, which frames policy generation as finding distributions over policies that balance *specificity* and *value*. Building on this definition, we propose the *Strategist* agent—an iterative framework that leverages large language models (LLMs) to synthesize domain knowledge into a structured representation of actionable strategies and sub-goals. We further develop a *reward shaping methodology* that translates these strategies expressed in natural language into quantitative feedback for RL methods. Empirically, we demonstrate a significantly faster convergence than conventional PPO. Taken together, our findings highlight that *top-down strategic exploration* opens new avenues for enhancing RL on real-world decision problems.

This paper addresses the challenge of discontinuity in goal-achievement capabilities observed in Goal-conditioned Reinforcement Learning (GCRL) algorithms. Through a theoretical analysis, we identify that the reuse of successful trajectories or policies during training can aid in achieving adjacent goals of achievable goals. However, the policy discrepancy between achievable and adjacent goals must be carefully managed to avoid both overly trivial and excessively large differences, which can respectively hinder policy performance. To tackle this issue, we propose a margin-based policy self-regularization approach that optimizes the policy discrepancies between adjacent desired goals to a minimal acceptable threshold. This method can be integrated into popular GCRL algorithms, such as GC-SAC, HER, and GC-PPO. Systematic evaluations across two robotic arm control tasks and a complex fixed-wing aircraft control task demonstrate that our approach significantly improves the continuity of goal-achievement abilities of GCRL algorithms, thereby enhancing their overall performance.

Reinforcement Learning with Human Feedback (RLHF) is a widely used fine-tuning approach that aligns machine learning models, particularly Language Models (LMs) with human preferences. There are typically multiple objectives driving the preference, hence humans find it easier to express per-objective comparisons rather than a global preference between two choices, e.g. compare two papers on their novelty, clarity, correctness, etc. Multi-Objective RLHF aims to use per-objective preference feedback and achieve a Pareto optimal tradeoff among these objectives by aggregating them into a single unified objective for optimization. However, nearly all prior works rely on linear aggregation, which rules out policies that favor specific objectives such as the worst one. The only existing approach using non-linear aggregation is computationally expensive due to its reward-based nature and the need for retraining whenever the aggregation parameters change. In this work, we address this limitation by transforming the non-linear aggregation maximization problem into a series of sub-problems. Each sub-problem involves only linear aggregation, making it computationally efficient to solve. We further extend our framework to handle multi-group scenarios, where each group has distinct weights for the objectives. Our method enables achieving consensus or maximizing the aggregated objective across all groups. Theoretically, we demonstrate that our algorithmic framework achieves sublinear regret and can be easily adapted to a reward-free algorithm. Empirically, leveraging our theoretical insights, we propose a nearly training-free algorithm once the optimal policies for individual objectives are obtained.

Multi-kernel clustering (MKC) has emerged as a powerful method for capturing diverse data patterns, offering robust and generalized representations of data structures. However, the increasing deployment of MKC in real-world applications raises concerns about its vulnerability to adversarial perturbations. While adversarial robustness has been extensively studied in other domains, its impact on MKC remains largely unexplored. In this paper, we address the challenge of assessing the adversarial robustness of MKC methods in a black-box setting. Specifically, we propose *AdvMKC*, a novel reinforcement-learning-based adversarial attack framework designed to inject imperceptible perturbations into data and mislead MKC methods. AdvMKC leverages proximal policy optimization with an advantage function to overcome the instability of clustering results during optimization. Additionally, it introduces a generator-clusterer framework, where a generator produces adversarial perturbations, and a clusterer approximates MKC behavior, significantly reducing computational overhead. We provide theoretical insights into the impact of adversarial perturbations on MKC and validate these findings through experiments. Evaluations across seven datasets and eleven MKC methods (seven traditional and four robust) demonstrate AdvMKC's effectiveness, robustness, and transferability.

Trajectory prediction for surrounding agents is a challenging task in autonomous driving due to its inherent uncertainty and underlying multimodality. Unlike prevailing data-driven methods that primarily rely on supervised learning, in this paper, we introduce a novel **G**raph-**o**riented **I**nverse **R**einforcement **L**earning (GoIRL) framework, which is an IRL-based predictor equipped with vectorized context representations. We develop a feature adaptor to effectively aggregate lane-graph features into grid space, enabling seamless integration with the maximum entropy IRL paradigm to infer the reward distribution and obtain the policy that can be sampled to induce multiple plausible plans. Furthermore, conditioned on the sampled plans, we implement a hierarchical parameterized trajectory generator with a refinement module to enhance prediction accuracy and a probability fusion strategy to boost prediction confidence. Extensive experimental results showcase our approach not only achieves state-of-the-art performance on the large-scale Argoverse & nuScenes motion forecasting benchmarks but also exhibits superior generalization abilities compared to existing supervised models.

Motivated by the first priority of safety in many real-world applications, we propose \textsc{MaxSafe}, a chance-constrained bi-level optimization framework for safe reinforcement learning. \textsc{MaxSafe} first minimizes the unsafe probability and then maximizes the return among the safest policies. We provide a tailored Q-learning algorithm for the \textsc{MaxSafe} objective, featuring a novel learning process for \emph{optimal action masks} with theoretical convergence guarantees. To enable the application of our algorithm to large-scale experiments, we introduce two key techniques: \emph{safety polarization} and \emph{safety prioritized experience replay}. Safety polarization generalizes the optimal action masking by polarizing the Q-function, which assigns low values to unsafe state-action pairs, effectively discouraging their selection. In parallel, safety prioritized experience replay enhances the learning of optimal action masks by prioritizing samples based on temporal-difference (TD) errors derived from our proposed state-action reachability estimation functions. This approach efficiently addresses the challenges posed by sparse cost signals. Experiments on diverse autonomous driving and safe control tasks show that our methods achieve near-maximal safety and an optimal reward-safety trade-off.

Self-attention has emerged as a core component of modern neural architectures, yet its theoretical underpinnings remain elusive. In this paper, we study self-attention through the lens of *interacting entities*, ranging from agents in multi-agent reinforcement learning to alleles in genetic sequences, and show that a single layer linear self-attention can *efficiently* represent, learn, and generalize functions capturing pairwise interactions, including out-of-distribution scenarios. Our analysis reveals that self-attention acts as a *mutual interaction learner* under minimal assumptions on the diversity of interaction patterns observed during training, thereby encompassing a wide variety of real-world domains. In addition, we validate our theoretical insights through experiments demonstrating that self-attention learns interaction functions and generalizes across both population distributions and out-of-distribution scenarios. Building on our theories, we introduce *HyperFeatureAttention*, a novel neural network module designed to learn couplings of different feature-level interactions between entities. Furthermore, we propose *HyperAttention*, a new module that extends beyond pairwise interactions to capture multi-entity dependencies, such as three-way, four-way, or general $n$-way interactions.