Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents

Key points

• The central challenge is therefore to construct a reward from existing experience that exposes the current policy's failures while remaining stationary during optimization.1
• We introduce Reward Engine, which, like a self-motivated student, first works out what success requires and then optimizes toward it.1
• The frozen RewardPack then guides the agent to revise and select its outputs, and can also serve as a reward for reinforcement learning.1
• Used to revise executions, it raises three-run consistency ($Pass^{3}$) over Baseline from 45.0% to 54.0% and from 78.4% to 83.7%.1

Sources (1)

• [1]Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 04:25 AM

  Open-ended tool agents can complete the same objective through different tool paths, and their quality depends jointly on the interaction trace, environment state, and final artifacts, so such tasks rarely admit an executable ground-truth reward.

  The central challenge is therefore to construct a reward from existing experience that exposes the current policy's failures while remaining stationary during optimization.

Extractive summary: sentences quoted from the sources.