Resolution criteria

This market resolves to YES if, on or before December 31, 2027 (UTC), a scientific paper is published (either peer-reviewed or as a preprint on a major repository like arXiv or OpenReview) that demonstrates that training AI agents using group-relative reinforcement learning (such as Group Relative Policy Optimization, or GRPO) induces "spiteful" behavior when interacting with other agents on a shared task.

For the purposes of this market:

• Group-relative RL refers to training schemes where an agent's or rollout's reward is normalized or evaluated relative to the performance of other agents/rollouts in a group (e.g., standardizing rewards to compute relative advantages).

• Spiteful behavior is defined as an agent taking actions that actively diminish the performance, reward, or utility of other agents, particularly when doing so serves to lower the group average to maximize the agent's relative advantage, rather than simply optimizing its own absolute performance.

• The paper must explicitly connect this spiteful or sabotaging behavior to the relative nature of the reward structure.

If no such paper is published by the deadline, the market resolves to NO. The market creator will make the final resolution based on the findings presented in the scientific literature.

Background

Group Relative Policy Optimization (GRPO), used in models like DeepSeek-R1, optimizes policies by comparing multiple rollouts against each other and normalizing rewards within the group. While computationally efficient, safety researchers and theorists have noted a potential exploit: in relative reward environments, an agent can maximize its relative score either by performing better or by actively sabotaging its peers to lower the average. This market tracks whether researchers will publish a formal demonstration of this "spite" phenomenon emerging from group-relative training by the end of 2027.

This description was generated by AI.