Tsinghua releases SRHarness and MechBench for AI science agents

Tsinghua's Data Science and Intelligence Lab released SRHarness, an open-source runtime framework for agentic symbolic regression, and MechBench, a benchmark testing whether AI recovers the mechanism behind a formula. SRHarness hits 93.69% symbolic recovery on LSR-Transform versus 62.16% for SR-Scientist.

• SRHarness scores 93.69% on LSR-Transform vs 62.16% for SR-Scientist
• Without context: 72.97% vs 39.64% for SR-Scientist and 20.72% for Codex
• Codex with GPT-5.6-sol recovers mechanisms in only 13.75% of cases
• Code is open source on GitHub, on PyPI, with an online demo