PostTrainBench

PostTrainBench is a benchmark that evaluates whether CLI agents can enhance the performance of pre-trained large language models through post-training. Each agent receives access to evaluation scripts and a fixed compute budget on an H100 GPU, then attempts to boost model scores across reasoning, math, knowledge, health, and code tasks. The benchmark naturally assesses an agent's capacity to conduct AI research and development work autonomously.

The leaderboard tracks multiple agents including those built with Claude Code, Codex CLI, and other frameworks. Performance is measured across six evaluation tasks spanning domains like AIME mathematics, graduate-level science questions, and code generation. Results include mean performance scores with standard deviations across multiple seeds, providing a standardized way to compare different agent approaches.