Reputation-Gated Cooperation

Agents play a repeated donation game where each decides how much to give a partner based on that partner's observed history of past donations; strategies are written as evolving natural-language prompts, and only strict exclusion of past defectors, not reward of past cooperators, predicts resistance to free-rider invasion.

Properties

family: Generative Exploration
maturity: emerging
diagram: evolution
whyMultiAgent: Reputation only exists as a signal one agent holds about another; a single model has no population of distinct histories to observe or discriminate between. The paper's central result, that culturally evolved LLM strategies collapse to Image Scoring and never reach the more robust Leading-Eight norms, is a property of population-level cultural transmission across many interacting agents, not of any one agent's policy. Distinct from the already-tracked evolutionary-selection pattern, which selects surviving variants via an external fitness function on a task-performance metric: here the selection pressure is endogenous to the population itself (whether a strategy resists exploitation by free-riders in a live donation game), and the paper's finding is about which cooperative norms culturally evolved strategies can and cannot reach, not about optimising a task fitness score. Also distinct from arena-rating, which uses persistent ratings purely to rank competitors for a leaderboard: here reputation directly gates each agent's own cooperate or defect decision toward a specific partner, and the object under study is whether the emergent cooperation norm itself is stable against exploitation, not which agent ranks highest.
useCases: [Multi-agent resource-sharing systems vulnerable to free-riders][Decentralised agent marketplaces without a central reputation authority][Studying norm robustness before deploying autonomous cooperative agents]
source: Emergence of Reputation-Based Cooperation in LLM Agents · arXiv:2608.04507

Provenance

Source: https://arxiv.org/abs/2608.04507

Reputation-Gated Cooperation

Properties

family

Generative Exploration

maturity

emerging

diagram

evolution

whyMultiAgent

Reputation only exists as a signal one agent holds about another; a single model has no population of distinct histories to observe or discriminate between. The paper's central result, that culturally evolved LLM strategies collapse to Image Scoring and never reach the more robust Leading-Eight norms, is a property of population-level cultural transmission across many interacting agents, not of any one agent's policy. Distinct from the already-tracked evolutionary-selection pattern, which selects surviving variants via an external fitness function on a task-performance metric: here the selection pressure is endogenous to the population itself (whether a strategy resists exploitation by free-riders in a live donation game), and the paper's finding is about which cooperative norms culturally evolved strategies can and cannot reach, not about optimising a task fitness score. Also distinct from arena-rating, which uses persistent ratings purely to rank competitors for a leaderboard: here reputation directly gates each agent's own cooperate or defect decision toward a specific partner, and the object under study is whether the emergent cooperation norm itself is stable against exploitation, not which agent ranks highest.

useCases

[Multi-agent resource-sharing systems vulnerable to free-riders][Decentralised agent marketplaces without a central reputation authority][Studying norm robustness before deploying autonomous cooperative agents]

source

Emergence of Reputation-Based Cooperation in LLM Agents · arXiv:2608.04507