Training Generalizable Collaborative Agents via Strategic Risk Aversion

• URL: http://arxiv.org/abs/2602.21515v2
• Date: Fri, 27 Feb 2026 04:56:38 GMT
• Title: Training Generalizable Collaborative Agents via Strategic Risk Aversion
• Authors: Chengrui Qu, Yizhou Zhang, Nicolas Lanzetti, Eric Mazumdar,
• Abstract summary: We study the concept of strategic risk aversion and interpret it as a principled inductive bias for generalizable cooperation with unseen partners.<n>We develop a multi-agent reinforcement learning (MARL) algorithm that integrates strategic risk aversion into standard policy optimization methods.
• Score: 14.968945672756854
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Many emerging agentic paradigms require agents to collaborate with one another (or people) to achieve shared goals. Unfortunately, existing approaches to learning policies for such collaborative problems produce brittle solutions that fail when paired with new partners. We attribute these failures to a combination of free-riding during training and a lack of strategic robustness. To address these problems, we study the concept of strategic risk aversion and interpret it as a principled inductive bias for generalizable cooperation with unseen partners. While strategically risk-averse players are robust to deviations in their partner's behavior by design, we show that, in collaborative games, they also (1) can have better equilibrium outcomes than those at classical game-theoretic concepts like Nash, and (2) exhibit less or no free-riding. Inspired by these insights, we develop a multi-agent reinforcement learning (MARL) algorithm that integrates strategic risk aversion into standard policy optimization methods. Our empirical results across collaborative benchmarks (including an LLM collaboration task) validate our theory and demonstrate that our approach consistently achieves reliable collaboration with heterogeneous and previously unseen partners across collaborative tasks.

Related papers

• Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia [100.74015791021044]
  Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction.<n>Existing evaluation methods fail to measure how well these capabilities generalize to novel social situations.<n>We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains.
  arXiv  Detail & Related papers  (2025-12-03T00:11:05Z)
• Adaptively Coordinating with Novel Partners via Learned Latent Strategies [19.014669675808133]
  We introduce a strategy-conditioned cooperator framework that learns to represent, categorize, and adapt to a broad range of potential partner strategies in real-time.<n>Our approach encodes strategies with a variational autoencoder to learn a latent strategy space from agent trajectory data.<n>We leverage a fixed-share regret minimization algorithm that dynamically infers and adjusts the partner's strategy estimation during interaction.
  arXiv  Detail & Related papers  (2025-11-16T19:45:35Z)
• Learning "Partner-Aware" Collaborators in Multi-Party Collaboration [12.287537011305497]
  Large Language Models (LLMs) are increasingly bring deployed in agentic settings where they act as collaborators with humans.<n>This paper builds on the AI alignment and safe interruptability literature to offer novel theoretical insights on collaborative behavior.<n>We propose Interruptible Collaborative Roleplayer (ICR)-a novel partner-aware learning algorithm to train CG-optimal collaborators.
  arXiv  Detail & Related papers  (2025-10-26T00:05:48Z)
• Modeling Latent Partner Strategies for Adaptive Zero-Shot Human-Agent Collaboration [8.829018359718086]
  We introduce TALENTS, a strategy-conditioned cooperator framework that learns to represent, categorize, and adapt to a range of partner strategies.<n>Our approach utilizes a variational autoencoder to learn a latent strategy space from trajectory data.<n>In order to adapt to previously unseen partners, we leverage a fixed-share regret minimization algorithm.
  arXiv  Detail & Related papers  (2025-07-07T17:53:13Z)
• Corrupted by Reasoning: Reasoning Language Models Become Free-Riders in Public Goods Games [87.5673042805229]
  How large language models balance self-interest and collective well-being is a critical challenge for ensuring alignment, robustness, and safe deployment.<n>We adapt a public goods game with institutional choice from behavioral economics, allowing us to observe how different LLMs navigate social dilemmas.<n>Surprisingly, we find that reasoning LLMs, such as the o1 series, struggle significantly with cooperation.
  arXiv  Detail & Related papers  (2025-06-29T15:02:47Z)
• Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination [37.90912492084769]
  We study how reinforcement learning on a distribution of environments with a single partner enables learning general cooperative skills.<n>We introduce two Jax-based, procedural generators that create billions of solvable coordination challenges.<n>Our findings suggest that learning to collaborate across many unique scenarios encourages agents to develop general norms.
  arXiv  Detail & Related papers  (2025-04-17T07:41:25Z)
• Multi-agent cooperation through learning-aware policy gradients [53.63948041506278]
  Self-interested individuals often fail to cooperate, posing a fundamental challenge for multi-agent learning.<n>We present the first unbiased, higher-derivative-free policy gradient algorithm for learning-aware reinforcement learning.<n>We derive from the iterated prisoner's dilemma a novel explanation for how and when cooperation arises among self-interested learning-aware agents.
  arXiv  Detail & Related papers  (2024-10-24T10:48:42Z)
• Tackling Cooperative Incompatibility for Zero-Shot Human-AI Coordination [36.33334853998621]
  We introduce the Cooperative Open-ended LEarning (COLE) framework to solve cooperative incompatibility in learning. COLE formulates open-ended objectives in cooperative games with two players using perspectives of graph theory to evaluate and pinpoint the cooperative capacity of each strategy. We show that COLE could effectively overcome the cooperative incompatibility from theoretical and empirical analysis.
  arXiv  Detail & Related papers  (2023-06-05T16:51:38Z)
• PECAN: Leveraging Policy Ensemble for Context-Aware Zero-Shot Human-AI Coordination [52.991211077362586]
  We propose a policy ensemble method to increase the diversity of partners in the population. We then develop a context-aware method enabling the ego agent to analyze and identify the partner's potential policy primitives. In this way, the ego agent is able to learn more universal cooperative behaviors for collaborating with diverse partners.
  arXiv  Detail & Related papers  (2023-01-16T12:14:58Z)
• Conditional Imitation Learning for Multi-Agent Games [89.897635970366]
  We study the problem of conditional multi-agent imitation learning, where we have access to joint trajectory demonstrations at training time. We propose a novel approach to address the difficulties of scalability and data scarcity. Our model learns a low-rank subspace over ego and partner agent strategies, then infers and adapts to a new partner strategy by interpolating in the subspace.
  arXiv  Detail & Related papers  (2022-01-05T04:40:13Z)
• On the Critical Role of Conventions in Adaptive Human-AI Collaboration [73.21967490610142]
  We propose a learning framework that teases apart rule-dependent representation from convention-dependent representation. We experimentally validate our approach on three collaborative tasks varying in complexity.
  arXiv  Detail & Related papers  (2021-04-07T02:46:19Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.