Research map / Multi-agent systems and games / Games and equilibria
Learning to Balance Altruism and Self-interest Based on Empathy in Mixed-Motive Games
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LASE, a decentralized MARL algorithm that gifts a portion of rewards to co-players, with gift weights set by counterfactual social-relationship estimates computed via a perspective-taking module, balancing altruism and self-protection in mixed-motive games.
Problem
Agents in mixed-motive games must balance altruistic cooperation with self-protection against exploitation, but decentralized self-interested learners often converge to defection and existing gifting methods ignore social relationships.
Approach
Empathic gifting: each agent infers a continuous social relationship with each co-player by comparing the Q-value of the actual joint action to a counterfactual baseline that marginalizes that co-player's action, using a perspective-taking network to predict the co-player's policy from a simulated observation; the inferred value sets the fraction of reward gifted.
Key results
- In IPD, two LASE agents converge to cooperate with probability around 0.93, escaping the (D,D) equilibrium that A2C falls into.
- In Cleanup, LASE achieves Equality E ≈ 0.802 versus 0.496 for GO, indicating higher fairness.
- In extended Cleanup (8 agents), LASE self-play total reward is 56.513 versus 20.798 for IA and 0.135 for A2C.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LASE instead of IA | self-play on extended Cleanup · Cleanup.Extn | total reward: 20.798 -> 56.513 | Helped |
| use LASE instead of IA | self-play on extended SSG · SSG.Extn | total reward: 227.762 -> 232.564 | Helped |
| replace inferred co-player policy with a uniform policy (ablation) | LASE · Cleanup | convergence speed and performance | Hurt |
| replace inferred co-player policy with a uniform policy (ablation) | LASE · Cleanup | mean gifting weight: 0.346 -> 0.457 | Mixed |
| use LASE instead of GO | self-play in SSH · SSH | collective reward | Helped |
| use LASE instead of GO | Cleanup · Cleanup | Equality: 0.496 -> 0.802 | Helped |
| use LASE instead of A2C | self-play in IPD · IPD | collective reward | Helped |
| use LASE instead of A2C | self-play in SSG · SSG | collective reward | Helped |
Surprising findings
- The gifting-based GO (group optimal) agents fall into a lazy problem in SSH, underperforming LASE despite optimizing collective reward.
- Replacing the perspective-taking prediction with a uniform policy (LASE w/o) only modestly degrades performance but yields higher gifting weights, i.e., more gifting and less efficiency.
Limitations
- Gifts are not allowed to be refused; negative social relationships are clipped to zero, so exploitation handling is limited.
- Only discrete action spaces are studied; extension to continuous actions is left to future work.
- Social relationship estimates have notable uncertainty under partial observability.
Experimental setup
Datasets: Iterated Prisoner's Dilemma Coingame Cleanup Sequential Stag-Hunt Sequential Snowdrift Game
Benchmarks: Iterated Prisoner's Dilemma Coingame Cleanup Sequential Stag-Hunt Sequential Snowdrift Game
Metrics: collective reward cooperation probability Equality (fairness) index gifting weights
Baselines: A2C LIO LOLA IA SI GO LASE w/o
Builds on: LIO LToS SOM A2C LOLA
Closest papers by meaning
- Fair Cooperation in Mixed-Motive Games via Conflict-Aware Gradient Adjustment (NeurIPS 2025 · similarity 0.89)
- Beyond Mandatory Federations: Balancing Egoism, Utilitarianism and Egalitarianism in Mixed-Motive Games (AAAI 2025 · similarity 0.89)
- Aligning Individual and Collective Objectives in Multi-Agent Cooperation (NeurIPS 2024 · similarity 0.88)
- A Unified Model of Direct and Indirect Reciprocity in Multichannel Games (AAAI 2025 · similarity 0.87)
- Reciprocal Reward Influence Encourages Cooperation From Self-Interested Agents (NeurIPS 2024 · similarity 0.86)
- Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.