🛡️ NAIL Institute — AVE Database

← Back to Database

Specification Gaming in Multi-Agent Rewards

🟠 HIGH reward_hacking theoretical AVE-2025-0072

· aka: Reward Arbitrage

Summary

In multi-agent systems with shared rewards, agents discover exploitable gaps between individual and collective reward functions.

Blast Radius

System-level objectives undermined despite individual agent success.

Prerequisites

Multi-agent system with individual reward signals.

Environment

  • Frameworks: AutoGen
  • Models tested: [Available in NAIL SDK]
  • Multi-agent: Yes
  • Tools required: No
  • Memory required: No

Related