Multi-Agent Reinforcement Learning

Also known as MARL

Overview

Multi-agent reinforcement learning studies how multiple decision-makers learn when each agent’s actions affect both the environment and other agents, across cooperative, competitive, and mixed-motive settings. Common approaches include centralized training with decentralized execution, independent or shared actor-critic policies, value factorization such as QMIX, self-play and game-solving methods, and graph-based communication or hierarchical policies. Researchers test these methods in benchmarks such as SMAC, MPE, Overcooked, and multi-agent MuJoCo, and apply them to robotics, traffic, wireless networks, energy systems, and language-agent workflows. A central challenge is scaling coordination and credit assignment while keeping learning stable under non-stationarity and policies robust, safe, and fair with unfamiliar partners and conditions.

Latest papers

402 papers37 papers in the last four weeks, down 27% on the four weeks before. 0.3% of all new papers.

Nov 13, 2026

When Cooperation Must Change: Semantic Strategy Unlearning in Multi-Agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) supports coordinated decision-making in distributed Web services, multi-robot control, and collaborative games through learned coordination conventions. When deployment requirements change, a previously useful convention may become obsolete or unsafe. For example, revised cache-sharing permissions in Web content delivery may require caches to abandon a learned division of storage and serving roles while maintaining efficient delivery. Teams must therefore remove the specified convention while preserving useful cooperation. This is challenging because the same relational pattern can be reproduced by different agents across different histories. Runtime shielding leaves the underlying coordination preference intact and can impair task progress, while existing reinforcement unlearning and trajectory-deletion methods do not directly target role-invariant conventions. This paper formalizes this problem as semantic coordination-strategy unlearning and proposes CARVE.CARVE uses causal coordination attribution to identify influential interactions and retain-aware subspace updates to limit interference with unrelated behaviors. Counterfactual re-coordination helps agents learn productive alternatives, while adversarial auditing tests whether the target convention can still be recovered. Accepted updates are merged into decentralized actors for direct deployment. Evaluation across five cooperative MARL benchmarks shows that CARVE reduces mean nominal convention-completion and worst-case elicitation rates by 91.4% and 83.3%, respectively.The median utility loss is 5.9%, and measured inference latency matches that of the original policies.
Nov 12, 2026

From Delegation to Coordination: Learning Efficient Multi-Agent Deep Research

Multi-agent deep-research systems use delegation to parallelize search and keep long tool traces out of the main agent's context. Yet delegation helps only when task allocation and subtask execution are coordinated: the main agent must issue useful subtasks, while subagents must return reliable evidence. Existing training either optimizes only the main agent or broadcasts the same team advantage to all agents, which can reward misleading subagent reports and penalize useful ones. We introduce LEAD-RL (Learning Effective Agent Delegation with Reinforcement Learning), a joint main-agent-subagent training framework with contribution-aware credit assignment. A training-time rubric evaluates the research utility of each subagent report and gates its inherited team advantage, while final-answer correctness remains the only team reward; no token, step, or speedup reward is used. Across five benchmarks and two model families, LEAD-RL improves answer quality. On BrowseComp, it improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit. During validation, accuracy increases by 6.78 points at approximately constant critical-path cost. These results show that contribution-aware credit assignment improves the quality-efficiency trade-off of multi-agent deep research.
Nov 12, 2026

HALight: Heterogeneous Multi-Agent Reinforcement Learning for Traffic Signal Control under Lane-Level Missing Observations

Multi-agent reinforcement learning has shown great promise for traffic signal control. However, most existing methods rely on two key assumptions: (i) agents are assumed to be homogeneous with unified observation/action spaces; (ii) full and reliable lane-level traffic observations are available to all agents. In practice, urban road networks exhibit substantial structural heterogeneity (e.g., lane configurations and phase designs), while sensor failures and incomplete detector coverage often result in missing incoming lane measurements. Such issues can severely impair coordination and destabilize learning. To address these challenges, we propose HALight, a communication-enhanced heterogeneous multi-agent reinforcement learning framework for traffic signal control under lane-level missing observations. HALight supports intersection-specific observation and action modeling to accommodate heterogeneous intersections, and incorporates a lightweight one-hop neighborhood communication module to enrich spatial context when lane-level observations are partially unavailable. Moreover, HALight adopts a sequential policy update scheme to improve coordination among heterogeneous agents. Experimental results on two real-world road networks demonstrate that HALight outperforms representative homogeneous and heterogeneous baselines in terms of traffic efficiency and convergence behavior. The code and datasets used in this work are available at https://github.com/Huchenkaii/HALight.
Nov 9, 2026

A Modular Perspective for Modeling MARL Systems

The growing complexity of Multi-Agent Reinforcement Learning (MARL) has fragmented the literature, obscuring underlying architectural choices behind diverse implementations. This fragmentation makes it difficult to evaluate critical modeling dimensions in isolation, identify how components affect the resulting system dynamics, and compare, reuse, or combine contributions across the literature. This paper proposes a modular perspective for modeling, analyzing, and implementing MARL systems. We decompose them into eight functional modules: agents, environments, learning, reward modeling, training and execution modes, communication, modeling of other agents, and evaluation criteria. For each module, we clarify its role and review representative approaches. We instantiate this decomposition in MARLKIT, a framework whose modules are explicit and interchangeable. MARLKIT demonstrates how the proposed perspective supports transparent experimental assumptions and the controlled variation and combination of MARL mechanisms. Experiments illustrate its ability to prototype, compare, reuse and combine modules while making differences between configurations explicit. Overall, the paper provides a structured perspective on MARL modeling and implementation, offering a common language for making modeling choices explicit, supporting clearer descriptions of MARL approaches, and supporting more controlled experimentation through the systematic isolation and combination of MARL mechanisms.
Nov 7, 2026

A Gamified Framework for Conflict-Driven Design: Discovering Strategies for Collaborative Resilience with Multi-Agent Learning

Conflicting stakeholder objectives make it difficult to coordinate investments in urban resilience. We model conflict-driven urban design as a non-zero-sum digital board game in which stakeholder agents allocate scarce resources and balance individual payoffs against shared urban development and resilience objectives. We use multi-agent reinforcement learning (MARL) to train policies and examine their behavior across three settings. On a 4 x 4 board, learned policies reach about 84% of the total return found by a simulated annealing benchmark. On an 8 x 8 board with spatial and resource constraints, trained agents favor projects that improve resilience in self-play. In a 12 x 12 simulation initialized from historical maps of Tatsumi, Tokyo, trained agents achieve a mean total score of 452.1, compared with 357.3 for a random-policy baseline across 10,000 simulation runs. Gains in adaptability, sustainability, flood resistance, and greenery come with reductions in vitality and density. A human-in-the-loop interface supports replay and lets participants take over stakeholder decisions, providing a testbed for exploring these trade-offs in participatory urban planning.
Nov 7, 2026

M3FT: Micro-Macro-Micro Reasoning for Multi-Agent Cryptocurrency Perpetual Futures Trading

Cryptocurrency perpetual futures form a large-scale online financial market in which automated agents manage interacting asset positions under changing conditions. Policies relying solely on micro-level numerical states may overlook cross-asset relationships and become vulnerable to regime shifts, while macro-level linguistic judgments alone cannot ensure precise measurement or optimize positions under trading costs and risk constraints. Connecting market interpretation to coordinated multi-asset decisions remains challenging. We propose M3FT, a micro-macro-micro framework for multi-agent perpetual futures trading. The first micro stage constructs time-aligned numerical states from exchange and account observations through deterministic processing. The macro stage transforms these measured states into explicit market interpretations without adding external market information. Four asset-specific teams, each comprising a Market Interpreter, an Asset Communicator, and a Risk Reviewer, refine these interpretations through experience retrieval, cross-asset communication, and review of joint-position risks. The final micro stage combines encoded linguistic representations with the original numerical states to learn continuous target positions. We implement this framework with a language-conditioned multi-agent policy: asset actors use global and private numerical states together with linguistic context, while centralized critics evaluate joint actions using portfolio-wide information and shared-account rewards. Experiments on real cryptocurrency perpetual futures data, including comparisons across language models and policy-learning algorithms and component ablations, show that the best-performing M3FT configuration achieves a higher mean test return than the evaluated direct LLM, numerical RL, and rule-based baselines.
Nov 6, 2026

A Separation Principle for Cooperative Multi-Agent Reinforcement Learning

We study cooperative multi-agent reinforcement learning (MARL) problems in which agents evolve under decoupled noisy dynamics but share a population objective. A canonical example is a mobility operator that orchestrates a fleet of vehicles to meet customer demand. The stochastic and time-varying nature of these problems calls for learning approaches, yet existing MARL algorithms do not scale: learning assignment and routing jointly over the full state-action space becomes prohibitive as the number of agents grows. Formulating the problem over the space of probability measures, we prove a separation principle: the population cost-to-go function is upper bounded by an optimal transport problem whose transportation cost is the target-conditioned single-agent cost-to-go function. This simplifies the multi-agent problem to learning a single-agent policy and coordinating the agents via optimal transport. Building on this principle, we introduce SALT (Separation-based Assignment and Learning via optimal Transport), a family of algorithms that learn a target-conditioned single-agent policy using any standard reinforcement learning algorithm, and then coordinate the agents via optimal transport. Since learning occurs at the single-agent level, SALT scales to fleet sizes where MARL baselines become intractable and transfers without retraining to fleet sizes unseen during training. In a vehicle-delivery experiment on the southern Manhattan road network, where MARL baselines are computationally prohibitive, SALT improves travel time by 15-21% over a shortest-path routing heuristic. In stochastic grid worlds, SALT reduces average travel times by 23-61% relative to MARL baselines and scales to population sizes where these methods fail.
Nov 6, 2026

CARGO: Causal Attribution and Responsibility-Guided Preventive Safe Multi-Agent Reinforcement Learning for Cooperative Driving

This study presents Causal Attribution and Responsibility-Guided Optimization (CARGO), a multi-agent reinforcement learning framework for cooperative on-ramp merging in mixed traffic. A shared latent-context posterior and an action-conditioned response model estimate directed behavioral influence, with posterior uncertainty propagated into interaction-edge selection. Complementary attribution mechanisms couple signed influence with marginal team utility for cooperative reward shaping and assign responsibility from positive excess tail risk relative to a safe reference policy. Conditional diffusion generates future responses for action-dependent risk estimation. Responsibility-guided preventive optimization calibrates agent-specific risk-reduction requirements to local capacity within a policy trust region, using reserved safety margins, adaptive requirement scaling, and joint feasibility verification. Theoretical analysis establishes risk monotonicity and sufficient local conditions for non-decreasing environmental return under explicit assumptions. Experiments in highway-env across three traffic densities, together with controlled graph-recovery and attribution studies, support the effectiveness of the framework. Under high-density traffic, CARGO increases the success rate by 66.04% and reduces the crash rate by 56.23% relative to the best evaluated baseline for each metric.
Nov 5, 2026

Discovering Update Rules for Opponent Shaping in General-Sum Games

In general-sum games, self-interested learners can reach outcomes that are poor for everyone, which has motivated methods that shape the opponent's learning. Methods like Learning with Opponent-Learning Awareness (LOLA), which differentiates through the opponent's learning process, and Model-Free Opponent Shaping (M-FOS), which uses model-free optimisation methods for long-horizon opponent shaping, have enabled mutual cooperation and exploitation of learners. However, these methods require access to the opponent's parameters, policy, rewards or learning rule during deployment. In single-agent reinforcement learning, update rules discovered by meta-learning have achieved state-of-the-art results (e.g. DiscoRL). These rules are separate from the policy they train and read only the agent's own experience. We hereby extend this framework to general-sum settings by introducing Discovered Update Rule for Opponent Shaping (DUOS). To the best of our knowledge, this is the first method to discover an update rule for an agent in general-sum games. DUOS meta-trains the rule to maximise the agent's expected return over its lifetime, and updates an inner network using the agent's own experience. We show that our discovered rule exhibits behaviour consistent with opponent shaping without reading opponent policies or parameters, and ranks 1st in normalised mean score and 1st in self-play among tournament entrants.
Nov 5, 2026

SEER: A Scalable Event-Driven Runtime for Parallel Execution of Multi-Agent Reinforcement Learning Environments

Multi-agent reinforcement learning (MARL) is increasingly used in systems where many entities interact in a shared environment. In these workloads, policy inference can be batched on GPUs, but the CPU-side environment path still updates entity states, resolves local interactions, and builds observations inside each env.step\mathrm{env.step}. This path can dominate latency in interaction-intensive workloads after policy batching. Existing vectorized environment interfaces and general event-driven runtimes do not directly address the fine-grained cross-entity work inside one bounded MARL step while preserving the step-level semantics expected by learning systems. We present SEER (Scalable Event-Driven Environment Runtime), a runtime for MARL environment execution under an explicit event-locality contract. SEER models each environment step as a bounded stream of timestamped events. If direct state updates are entity-local and cross-entity effects are represented as ordered future events, SEER can run independent events in parallel while preserving serial-equivalent boundary outputs. To make this contract practical, SEER uses a CAS-based dual-layer event queue: local event queues preserve per-entity causal order, while a parallel executive list exposes ready work to global worker threads. SEER further uses STPA, a preflight-guided thread-pool selection mechanism that combines short runtime traces with static workload and hardware features. We evaluate SEER on two parameterized MARL workloads: compute-intensive UAV swarm control (Aircraft) and communication-intensive wireless management (PCS). On Aircraft, SEER reduces inference latency by 16.2% over the baseline. On PCS, SEER reduces environment-execution latency by 46.9%, 19.5%, and 64.9% over ROSS-T, NBSE, and ROOT-Sim, respectively.
Nov 4, 2026

H^2MARL: Hierarchical Heterogeneous Multi-agent Reinforcement Learning for Humanoid Robot Control

Coordinated whole-body control of humanoid robots remains challenging due to high-dimensional coupled dynamics and substantial physical heterogeneity across joint degrees of freedom (JDoFs). Existing multi-agent reinforcement learning methods either employ coarse agent partitions or directly apply generic heterogeneous-agent algorithms, while hierarchical controllers commonly optimize high-level planning and low-level execution independently, leading to uncoordinated motion. In this paper, we propose H^2MARL, a hierarchical heterogeneous multi-agent reinforcement learning framework with two key mechanisms. We first formulate humanoid whole-body control using a JDoF-level physical-heterogeneity-aware policy factorization, where each torque-actuated JDoF is conditioned on an explicit descriptor of its actuation, kinematic, and dynamic properties. Moreover, we develop a closed-loop policy-executor co-optimization mechanism that jointly adapts the high-level pose planner and the low-level joint executor using shared interaction trajectories, thereby reducing the distribution mismatch between planning and execution. Extensive experiments in MuJoCo and Isaac Gym demonstrate that H^2MARL consistently improves locomotion stability, gait coordination, and motion quality over homogeneous, coarse-grained, and independently optimized alternatives.
Nov 3, 2026

Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning

Deep search and long-horizon reasoning unfold as branching trajectories that pursue parallel sub-queries, abandon dead-end detours, and cross-check sibling findings. Yet modern serving engines, harnesses, and RL trainers execute episodes as flat, append-only sequences: stretching a context window leaves failed detours that dilute attention, in-place edits break KV caching and recurrent states, and all-layer cache splicing overwrites hybrid recurrent working memory and triggers hijacking. We introduce Octopus, a co-designed serving runtime, memory manager, and RL framework executing episodes as a GPU-resident DAG over radix cursors (89.6% KV hit rate). Agents coordinate via five primitives: fork shares parent KV prefixes in O(1)O(1) with isolated 32K budgets, rewind restores recurrent checkpoints, read_thread mounts sibling KV pages in zero copy, and wait and done synchronize results. Octopus splices sibling KV pages only into full-attention layers under monotonic position offsets while recurrent layers skip mounted spans, retaining 92.2% of full-text answer information and cutting hijacking 21×21\times (57.9%→2.7%57.9\% \to 2.7\%) with zero new KV pages. During RL, our packed DAG kernel and DAG-GRPO objective deduplicate shared prefixes via loss masks while keeping mounted KV edges differentiable, cutting rollout-training KL divergence 32×32\times and routing reader gradients into producer KV projections. Trained on Qwen3.5-4B with binary rewards, Octopus self-organizes a commander-worker hierarchy and latent KV mesh, lifting BrowseComp-Plus accuracy under BM25 from 26.0% (32K ReAct) and 32.1% (143K GRPO) to 49.6% (+17.5 pts; 51.2% on OpenResearcher), surpassing 9B context-editing RL (42.5%), matching o3 (49.3%), and outperforming dense-retrieval agents up to 250×250\times larger: Tongyi-DeepResearch (44.5%), DeepSeek-R1 (16.4%), Kimi-K2 (35.4%), GPT-4.1 (36.4%), and Claude 4 Opus (36.8%).
Nov 3, 2026

Safeguarding Multi-Agent Reinforcement Learning for Connected and Automated Vehicles with Recursively Feasible Control Barrier Functions

Multi-Agent Reinforcement Learning (MARL) shows promise in coordinating connected and automated vehicles but offers no safety guarantee. Although Control Barrier Functions (CBFs) can add this guarantee, their optimization problem can become infeasible in practice due to control bounds. Moreover, they often overapproximate vehicle and road geometry, which makes them conservative. This paper proposes the Backup-Horizon CBF (BH-CBF), which evaluates safety along a horizon until all vehicles stop. BH-CBF is recursively feasible under control bounds, and it considers the exact vehicle and road geometry. Based on BH-CBF, we derive a fully decentralized safety filter that only needs a single broadcast of vehicle states and routes per step and stays real-time in dense traffic. We use BH-CBF to safeguard a MARL policy over its life cycle: as a reward during pre-training, as a run-time filter at deployment, and as a filter in the training loop that keeps fine-tuning on unseen maps collision-free. In simulations on real-world maps of a roundabout and an intersection, BH-CBF avoids all collisions of unsafe MARL policies and keeps its safety guarantee at every step. Existing filters either collide, lose the guarantee, or fall back to full braking far more often. In recorded roundabout traffic with 1264 human-driven vehicles over 115 km, our safeguarded policies shorten the median travel time by 14% compared with the human drivers, without any collision.
Nov 2, 2026

Finite-Sample Convergence in Networked Average Reward MARL: Decentralization Pitfalls and Entropy Remedies

We study the problem of average reward Multi-Agent Reinforcement Learning (MARL) where agents interact within a network. Each agent in the network conducts local updates with information within its kk-hop neighborhood and collaboratively maximizes the overall reward of the entire network. We first provide impossibility results in achieving convergence to the global optimal joint policies in our setting. These impossibility results highlight the challenges of decentralized learning with local information in multi-agent systems, namely the issues of multiagency, where each agent independently optimizes its policy, and partial observability, where agents have limited visibility into the global state. Given these challenges which indicate that decentralized policy optimization can generally be certified only up to stationarity, we provide finite-sample convergence guarantees for a decentralized actor-critic algorithm with linear function approximation that converges to an approximate stationary point with small gradient norm. To overcome such impossibility results, we further show that adding fixed-state entropy regularization reshapes the objective landscape which leads to even stronger global convergence guarantees for our proposed actor-critic algorithm.
Nov 2, 2026

An Axiomatic Framework for Reinforcement Learning in Open Multi-Agent Systems

Open multi-agent systems, where only a subset of agents are controllable, are central to real-world domains such as smart grids, yet lack a principled foundation for reinforcement learning algorithm design. We study this setting through n-agent ad hoc teamwork (NAHT), a recent framework for learning in open multi-agent environments. We propose an axiomatic framework that derives value function learning from the Shapley axioms (Efficiency, Symmetry and Linearity), treating them as structural constraints rather than heuristic design choices. This formulation builds on a novel NAHT decomposition, which characterises the fundamental difference between open and fully controllable multi-agent systems, and establishes a cooperative game-theoretic foundation for learning. We further show that enforcing these axioms on learning individual value functions recovers the Shapley value for Dec-POMDPs, providing a principled bridge between reinforcement learning and classical cooperative game theory. Building on this framework, we introduce Shapley Machines and Banzhaf Machines, two generic algorithmic templates that can be instantiated with standard reinforcement learning methods. When applied to strong base algorithms such as IPPO and POAM, our methods consistently improve learning efficiency and generalisation to unseen agent types and conventions across NAHT benchmark tasks. The code of experiments is published on https://github.com/hsvgbkhgbv/axiomatic-rl-naht.
Nov 2, 2026

KlaskTron: An Open-Source Platform for Physical Adversarial Multi-Agent RL

Progress in robot learning, particularly physical multi-agent reinforcement learning (MARL), is currently challenged by a limited availability of accessible, standardized benchmarks. While simulation-based MARL has produced remarkable emergent behaviors, from coordinated team play to complex tool use, translating these advances to physical systems remains difficult. A key barrier is infrastructural: physical platforms are often prohibitively expensive, require specialized facilities, or lack open mechanisms and designs for reproducibility. We introduce KlaskTron, an open-source, low-cost (<$2500 USD), desktop-scale robotic testbed for adversarial MARL based on the dynamic tabletop game KLASK. The platform features dual CoreXY gantries with high-torque brushless motors, enabling the high-speed, precise manipulation required for competitive play. Crucially, we provide a complete ecosystem: hardware designs (CAD, BOM), a GPU-native digital twin in NVIDIA Isaac Lab for high-fidelity modeling and simulation, and a validated sim-to-real baseline using a learned neural actuator model. We demonstrate that this baseline enables zero-shot policy transfer, with trained agents exhibiting emergent adversarial behaviors. To ensure full reproducibility of our results, all project assets are released publicly: hardware CAD/BOM and assembly documentation at [https://idscethzurich.github.io/klask_hardware], inference software stack at [https://github.com/IDSCETHZurich/klask_software] and the simulation and MARL setup at [https://github.com/IDSCETHZurich/klask_rl].
Nov 2, 2026

AGiR: Mitigating Gift Over-Reliance in Mixed-Motive Games

In this paper, we consider mixed-motive games (MMGs) which model various real-world multi-agent scenarios where self-interested agents must both cooperate and compete. We identify an issue with existing gifting methods, where agents become overly dependent on incoming gifts and fail to learn reward-yielding behaviors, a phenomenon we call gift over-reliance. To address this, we propose a refusal-augmented gifting framework that balances cooperation and competition in MMGs. Our approach introduces an adaptive refusal mechanism that allows agents to selectively reject a portion of received gifts from their policy-update signals, thereby mitigating gift over-reliance and preventing the agents from being exploited by other agents. This mechanism is gifting-method-agnostic, serving as a plug-in for existing schemes while enabling decentralized learning. We demonstrate that across diverse MMG environments and gifting schemes, our mechanism consistently reduces gift over-reliance and prevents lock-in to sacrificing policies, improving αα-fairness as a result, without necessitating major changes to existing training pipelines.
Nov 1, 2026

From Single-Agent Self-Discovery to Multi-Agent Adaptation: Scaling Laws of Partner Feedback and Convention Rigidity in Meta Reinforcement Learning Algorithms

A key milestone toward Recursive Self-Improvement (RSI) is replacing hand-crafted optimizers with meta-learned update rules (such as Disco103 and OPEN) that discover how agents learn. Yet when a self-discovered RL model trains with adapting teammates, does co-evolution sustain capability gains or trap agents in brittle conventions? Across 1,800+ runs in continuous collective robotics (VMAS) and discrete coordination (Overcooked), we establish five laws of multi-agent adaptation. First, four-cell counterfactual partner swaps expose an evaluation illusion: apparent team progress often masks focal regression (+1.066 team gain concealing -0.119 focal loss) and inflates transfer by +0.649 to +1.210. Second, a 288-run physical-coupling sweep (α∈[0,1.5]α\in [0, 1.5]) reveals a linear scaling law (β=+0.255β= +0.255, p=0.000488p = 0.000488, Spearman ρ=1.0ρ= 1.0): live reciprocal feedback keeps Disco103 invariant (AUC 0.297--0.322), whereas severing feedback triggers a monotonic collapse (-0.172 to -0.490), with OPEN establishing a sensitivity spectrum (Disco103 > OPEN > MAPPO). Third, extending navigation training eightfold (1.05M steps) enables Disco103 to overtake MAPPO (+0.087 team return, +0.133 bank gain), while decoupled training boosts cross-partner completion by +10.16% (56.25% vs. 46.09%). Fourth, in Overcooked, MAPPO's self-play conventions induce a 28.13% cross-partner collision rate, whereas Disco103 cuts collisions to 12.77% and achieves 3.33x higher bank return (9.96 vs. 2.99; net gain +6.43), further improved by decoupled meta-adaptation (+1.39 with scripted partners). Finally, perturbation pulses (p=0.00098p=0.00098), auxiliary entropy diagnostics (8.485 nats), and 48 multi-lifetime meta-evolved rules (L2∈[2.11,4.91]L_2 \in [2.11, 4.91]) trace these dynamics to partner-gradient blindness and auxiliary drift, informing multi-agent self-discoverd algorithms.
Oct 30, 2026

OmniPiano: Diverse Dexterous Piano-Playing Challenges for Standard, Robust, Safe, and Multi-Agent RL

Dexterous robotic manipulation remains a major challenge in robotics, exemplified by piano playing, which requires coordinated control of fingers and provides a demanding testbed for reinforcement learning (RL). Although previous work established a benchmark for robotic piano playing, it is restricted to two hands and does not support unified evaluation. To address this gap, we introduce OmniPiano, a benchmark that provides one to up to five Shadow Hands and supports standard, robust, safe, and multi-agent RL within a shared piano-playing task family. OmniPiano is developed for scalable and diverse piano-playing tasks with configurable perturbations, explicit safety constraints, and decentralized cooperation settings. Particularly, to facilitate usability and extendability, OmniPiano adopts a highly modular design and provides comprehensive tasks to support RL study of task performance, robustness, safety, and cooperation in dexterous control. With at least 912 task settings and 36 baseline algorithms, OmniPiano further incorporates LLM-based agents to broaden the evaluation scope and reveal new insights. Extensive evaluations show that state-of-the-art RL algorithms and frontier LLM-based agents still struggle on challenging piano-playing tasks, even under standard RL settings, highlighting substantial room for improvement in dexterous control. The code, dataset, tutorial, and demo are available at https://sail-research-lab.github.io/omnipiano/.
Oct 29, 2026

EpiAgent: Epistemic Self-Play for LLM Agents

Epistemic game theory explains an agent's strategic behavior via the beliefs it holds about others and how these beliefs are updated as the interactions unfold. Modern LLM-based agents, however, often lack an explicit mechanism for learning and updating such beliefs to guide decisions under incomplete information. To bridge this gap, we propose EpiAgent, a framework that equips LLM-based agents with an explicit, learnable epistemic layer, which maintains beliefs over others' private information and combines these beliefs with the agent's available information into an epistemic state for action selection. Concretely, the agent's policy is decomposed into two coupled stages: Epistemic Update, which updates these beliefs from interaction histories, and Action Planning, which selects actions conditioned on the updated epistemic state. By formulating their dependency as a bi-level optimization problem, we introduce Epistemic Self-Play, an algorithm that alternately optimizes the two stages through self-play, utilizing an intrinsic prediction reward to evaluate how well inferred beliefs predict subsequent observations. Empirical results across both mixed and competitive scenarios show that EpiAgent consistently outperforms prompting-based and RL-based baselines. Further analyses reveal the benefits of dedicated epistemic training, improvements in belief inference accuracy, and the contribution of learned beliefs to action planning. These results demonstrate the value of an explicit, learnable epistemic layer for effective multi-agent interactions.
Oct 28, 2026

Reinforcement Learning for the Game of Balatro: A Fast Simulator and Two-Policy Training Approach

Balatro, a 2024 roguelike deckbuilder, presents three difficulties that have historically derailed reinforcement learning methods: a 3,000-action state-dependent action space, horizons of several hundred decisions with sparse rewards, and pervasive stochasticity in both card draws and shop generation. Published RL results on the game remain scarce. We introduce an end-to-end pipeline comprising (i) a high-throughput Rust simulator that is automatically state-diffed against the live game client via BalatroBot, (ii) a condensed observation-action interface with dynamic per-step masking of legal moves, and (iii) a multi-agent PPO architecture that decomposes play into a tactical blind policy and a strategic shop policy that share the full game state but act in disjoint phases. Long-horizon credit assignment is partially addressed by per-agent reward routing with explicit cross-agent credit. On the main joint run, training metrics show a sharp phase transition in which the first-blind clear rate converges above 99% and mean episode return increases by roughly an order of magnitude, plateauing at a mean maximum ante of 3 out of 8; single-agent baselines under identical rewards oscillated between greedy and hoarding failure modes and did not stably clear the first blind. These results establish the two-policy decomposition as a strong reactive baseline for Balatro and localise the remaining gap to long-horizon strategic planning rather than tactical blind play.
Oct 28, 2026

CoSTA: Co-Evolving Self-Play Timeline Agents for Complex Temporal Knowledge Graph Question Answering

Complex Temporal Knowledge Graph Question Answering (TKGQA) involves more than retrieving facts: the timestamp or validity interval of one fact can constrain the search for another, while multiple intervals may jointly determine the answer. Existing single-agent TKGQA methods typically investigate these facts within a single reasoning context, where evidence from different event histories can become interleaved and distractors can misdirect subsequent retrieval. Distributing investigations across agents can separate event-specific evidence into distinct contexts, but effective collaboration must still be learned because question decomposition, subtask investigation, and evidence synthesis are interdependent. To address these limitations, we propose Co-Evolving Self-Play Timeline Agents (CoSTA), which organizes facts into navigable timelines, delegates investigations to workers, and uses a coordinator to synthesize their original timestamped observations. CoSTA initializes a shared policy through role-conditioned distillation and jointly optimizes both roles through co-evolving self-play with final-answer rewards. Across three benchmarks, CoSTA achieves state-of-the-art overall performance, with particularly strong gains on complex multi-fact questions, while its compact variants outperform the evaluated closed-source LLM agents.
Oct 28, 2026

Does 3D Pose Improve Action Valuation in Football? An Empirical Study of Pose Representations and Learning Objectives

Football action valuation estimates players' contributions to future scoring. Existing multi-agent deep reinforcement learning (MADRL) models use 2D positions and velocities, omitting posture information relevant to imminent actions. We investigate whether adding 3D pose to a SARSA-based model improves observed-action values relative to attack returns: goals, conceded goals or final expected possession value. Using UEFA Champions League 2022/23 event, tracking and 3D pose data, we compare six representations (no pose, raw joints, torso and head orientation, and spatial-temporal graph convolutional network (ST-GCN) embeddings) across five learning settings and three seeds. The settings vary temporal-difference targets and action-prediction loss, with Monte Carlo return regression as a reference. Across 50 on-ball and off-ball comparisons, pose robustly improved return prediction in only three, all using torso and head orientation, while worsening it in twelve, mostly with ST-GCN. Training choices had greater effects: removing the action-prediction loss raised on-ball skill of position-only values from 0.003 to 0.040 and their correlation with expected goals from 0.23 to 0.60; the improvement in return prediction persisted at matched training length. Separately, pose improved prediction of the on-ball action 0.2 s ahead (macro AUC 0.914 to 0.934 with raw joints), while the passer's pose during execution improved pass-success prediction (AUC 0.844 to 0.856). Within this training protocol, pose informed imminent actions but rarely improved observed-action values.
Oct 27, 2026

Multi-Agent Reinforcement Learning for Global Climate Policy Cooperation

International climate policy requires coordination among countries whose mitigation decisions impose local costs but generate global benefits. Classical game-theoretic and integrated assessment approaches can study these incentives, but typically rely on prescribed behavioural assumptions or equilibrium concepts. Recent work has therefore explored Multi-Agent Reinforcement Learning as a way to model climate-policy cooperation as an emergent outcome of repeated strategic interaction. We extend an existing MARL climate framework with a multi-stage negotiation mechanism in which agents propose agreements, accept or reject commitments, and face penalties for non-compliance. This allows us to study how agreement design affects learned mitigation behaviour, compliance, and climate outcomes. In an idealized setting with four homogeneous agents, we find that non-binding agreements do little to overcome free-riding, while moderate non-compliance penalties improve cooperation and reduce warming. However, excessively strong penalties discourage ambitious commitments, revealing a trade-off between enforcement and participation. We position the framework as a methodological contribution for testing agreement design rather than a predictive model of real negotiations.
Oct 27, 2026

Multi-Agent Reinforcement Learning of Co-Adaptive Hip Exoskeleton Control (MARL-Exo) for Metabolic Energy Reduction

Objective: This study aims to determine whether a hip-exoskeleton policy trained in simulation, by co-adapting with a musculoskeletal human model under a metabolic-energy reward, reduces the metabolic cost of walking in new users without per-user tuning, and how mechanical-power shaping affects the learned assistance. Methods: We developed MARL-Exo, a multi-agent reinforcement learning framework in which a full-body human model (304 muscle actuators; randomized anatomy, gait, and speed) and a bilateral hip-exoskeleton policy are trained concurrently in two stages. The exoskeleton observes only delayed hip kinematics and its past commands; its reward includes muscle metabolic power from a physiologically based energetics model, alone (Metabolic) or with mechanical-power shaping (Metabolic + Power). The Metabolic policy was deployed and tested without retraining in eight healthy adults walking at 1.25 m/s, using indirect calorimetry. Results: In simulation, the Metabolic policy reduced modeled muscle metabolic power by 13.8% relative to no-device walking, versus 8.8% for Metabolic + Power. On hardware, MARL-Exo reduced net metabolic rate by 18.3 ±\pm 6.6% relative to Passive in all eight participants. Relative to No-Exo, the mean reduction was 7.4 ±\pm 10.0%; six participants more than offset the 13.1% device-wearing penalty, walking up to 24.2% below their No-Exo cost. Conclusion: A metabolically informed policy co-adapted in simulation transferred to users without per-user tuning and consistently reduced the metabolic cost of walking with the device, while explicit power shaping reduced the simulated benefit. Significance: Unlike per-user human-in-the-loop optimization, this approach yields a single population-level controller and allows assistance objectives to be screened in simulation before human testing.
Oct 26, 2026

Preference-Guided Cooperation in Multi-Objective Multi-Agent Reinforcement Learning

Multi-objective multi-agent reinforcement learning (MOMARL) addresses agent coordination and objective trade-offs simultaneously while optimising overall system performance. Existing work tends to treat multi-agent cooperation and multi-objective optimisation as largely independent components of a learning problem, which amounts to an exorbitant complexity except in trivial cases. We study here whether the structure provided by multiple objectives can itself be exploited to facilitate cooperation, rather than being treated as an additional optimisation challenge. We formulate MOMARL at both the team and individual levels and introduce objective-wise counterfactual credit assignment to evaluate each agent's contribution across the team objectives. We propose Meta-Gradient Informed Multi-Objective Multi-Agent Reinforcement Learning (MIRL), with static (MIRL-S) and dynamic (MIRL-D) variants. By incorporating meta-gradient information, MIRL enables agents to learn how to prioritise objectives individually while collectively optimising the system-level objective, allowing the multi-objective structure to directly guide cooperation. Experiments on a multi-objective multiple particle environment show that our methods improve Pareto-front quality over baselines, with the dynamic variant achieving the strongest performance.
Oct 26, 2026

Belief-Conditioned Multi-Agent Actor Critic for Decentralized LLM Collaboration

Recent work has explored optimizing LLM collaboration through Multi-Agent Reinforcement Learning (MARL), but many approaches still rely on predefined workflows, limiting efficiency and scalability. Decentralized LLM collaboration is challenging because agents execute independently with limited communication, share a joint objective, and must infer useful information from communication traces. In this paper, we propose BCMAAC, a Belief-Conditioned Multi-Agent Actor-Critic method that maintains beliefs over contextual information and uses them to facilitate training. Experiments on Minecraft game-playing, web searching and tool-using tasks show that BCMAAC improves task performance and collaborative behavior compared to naive actor-critic baselines. The improvements are particularly pronounced in multi-turn tasks with richer contextual information, where belief modeling provides a clear advantage.
Oct 25, 2026

BE-CTDE: Blockchain-Enabled Multi-Agent Reinforcement Learning with Anomaly Detection and Token Economics

Multi-agent reinforcement learning supports cooperative decision-making in Internet of Things systems, but strategic deviations can disrupt coordination and undermine confidence in the training process. This paper proposes Blockchain-Enabled Centralized Training and Decentralized Execution (BE-CTDE), which uses blockchain as an active training-management layer. A behavioral verifier compares executed trajectories with shared-policy recommendations; its flags drive a token mechanism combining slashing, burning, rewards, and compensation. Retained balances determine each agent's contribution to the global policy update, connecting auditable behavior to learning influence. We derive a one-round incentive condition that accounts for action-dependent compensation and the payoff from deviation, and a finite expected exclusion-time bound under explicit conditional detection and accounting assumptions, showing perfect detection is unnecessary. An exact finite-population capacity bound characterizes the limits of the population-relative detector. In multi-agent path finding experiments, BE-CTDE outperforms a state-of-the-art learning-based path planner with substantially higher success rates under attack, even when adversaries form the majority. It also outperforms partially defended baselines in all configurations, with collision rates up to 35% lower than under detection without economic consequences and up to 43% lower than under conventional stake slashing. Together, these results characterize how behavioral verification and token feedback can support accountable cooperative learning.
Oct 21, 2026

SpecMAGIC: Specification-Guided Multi-Task Generalization in Multi-Agent Reinforcement Learning

Multi-task generalization is an important yet underexplored problem in multi-agent reinforcement learning (MARL). Many real-world applications require multi-agent systems to generalize across complex, temporally extended tasks with diverse inter-agent coordination requirements. We introduce a specification language that combines temporal-logic task specifications with operators for expressing such requirements. We propose SpecMAGIC, a hierarchical framework that leverages the automaton structure of task specifications for subtask decomposition. A centralized high-level planner identifies and allocates compatible subtask sequences across agents, while decentralized low-level policies execute individual subtasks. This enables zero-shot generalization to new tasks by composing reusable subtask policies according to new specifications, without retraining. We establish soundness and probabilistic completeness of the high-level planner. Experiments on multi-agent navigation tasks show that SpecMAGIC substantially outperforms existing approaches on unseen task specifications and maintains strong performance as the number of agents increases.
Oct 20, 2026

JaxMARL-BC: Multi-Agent Reinforcement Learning for Macroeconomic Modelling, in Seconds

Multi-agent reinforcement learning (MARL) offers a principled way of populating simulated economies with many heterogeneous decision makers. Its adoption, pioneered in finance, has nonetheless been slow in macroeconomics because of two main limitations: the lack of a shared frame of reference connecting MARL systems to dynamic stochastic general equilibrium (DSGE) models the profession already trusts, and the very long training times. The first limitation was recently addressed by the Multi-Agent Reinforcement Learning Business Cycle (MARL-BC), which places deep RL households in a real business cycle economy and reduces to classic DSGE models in specific limits. Here we address the second limitation. We present JaxMARL-BC, a JAX-accelerated simulator that trains MARL-BC economies in seconds (between 75x and 724x faster than the reference implementation), and scales to 200,000 independent learning households. We verify its correctness against the known DSGE limits, characterise its scaling through a simple cost law, and demonstrate its value by a proof-of-concept calibration of an extended model with heterogeneous returns on capital, in a two-hour experiment that would previously have taken over a month. We release JaxMARL-BC as open source.