Research map / Multi-agent systems and games / LLM agents

Talk, Judge, Cooperate: Gossip-Driven Indirect Reciprocity in Self-Interested LLM Agents

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

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 ALIGN, a framework where self-interested LLM agents sustain indirect reciprocity through open-ended, hierarchically toned public gossip, evaluated across eight LLMs in donation, IR, investment and market testbeds.

Problem

Indirect reciprocity is hard to sustain among decentralized, self-interested LLM agents without reliable reputation systems, and prior LLM reputation work either seeds altruists or assumes centralized observability.

Approach

Each agent has a gossip module that broadcasts open-ended evaluative messages in five hierarchical tones (praising to criticism) and an action module that conditions on public gossip logs plus private memory, with reflection steps; game theory shows public gossip can sustain cooperation when discount factor is at least c/b.

Key results

• ALIGN agents with DeepSeek-V3.1 Reasoner reach 100% cooperation ratio and discounted return of 15.44 in the infinite-horizon donation game, whereas non-gossiping reasoning models get zero return.
• Non-gossiping chat models such as GPT-4o Mini reach 0.36 cooperation ratio in the donation game, while reasoning models like o4-mini and Kimi-K2 defect fully (0.00 cooperation).
• Gemini 2.5 Flash-Lite with ALIGN achieves only 0.60 cooperation ratio in the donation game, showing lower gains for some chat models.

Surprising findings

• Some chat LLMs (e.g., GPT-4o Mini) cooperate without gossip even when cooperation is strictly suboptimal, while reasoning models consistently defect.
• Reasoning-focused LLMs are not inherently less cooperative; they cooperate when incentive-compatible under public gossip.

Limitations

• Results depend on a fixed temperature-0 setting and five seeds across simulated environments.
• Gossip can be untruthful or noisy, and some weaker models still defect or misreport in self-report extension.
• Collusive attackers reduce utility for some models, so robustness is not uniform.

Experimental setup

Benchmarks: repeated donation game indirect reciprocity game sequential investment scenario transaction market

Metrics: cooperation ratio image score reward per round discounted return Gini coefficient utility gap vs collusive agents

Baselines: non-gossiping agents binary reputation signals centralized-observability LLM reputation setups

Builds on: image scoring leading eight social norms grim trigger Tit-for-Tat

Closest papers by meaning

• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.87)
• Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.86)
• From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.86)
• A Unified Model of Direct and Indirect Reciprocity in Multichannel Games (AAAI 2025 · similarity 0.86)
• Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.86)
• To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)