PostTrainBench⁰: Can LLM Agents Automate LLM Post-Training Without Gradients?

Yuqiao Tan · Minzheng Wang · Shizhu He · Jun Zhao · Kang Liu
Institute of Automation, Chinese Academy of Sciences

PostTrainBench⁰ is a research preview for evaluating whether an autonomous LLM agent can improve a frozen language model using only score feedback and zeroth-order optimization.

Benchmark setup

Each four-hour run uses one read-only base checkpoint, eight inference workers, and seven tasks: Countdown, GSM8K, MATH-500, OlympiadBench, MBPP, ROCStories, and USPTO-50K. Gradients, evaluator edits, model replacement, and inference-time ensembles are forbidden.

Current evidence

Nearby checkpoints can improve the joint score, and all 51 pilot runs observed a complete seven-task candidate above their base checkpoint. The traces reveal distinct search programs and trajectories, while repeat variation makes finite-search reliability a concrete evaluation question. PostTrainBench⁰ releases this lower-cost, replayable AI-for-AI loop so the community can inspect it and build better search algorithms and protocols.

Code and benchmark materials