Researchers from the Laboratory of Data Science and Intelligence at the Department of Electronic Engineering, Tsinghua University, have introduced two interconnected AI projects for science: SRHarness—a runtime framework for symbolic regression of agents, and MechBench—a benchmark that assesses an AI's ability to reconstruct the mechanisms underlying a formula, rather than just the formula itself.
The team justified its work with a historical example: Kepler derived the laws of planetary motion from observations by describing the movement, whereas Newtonian mechanics explained the cause of that movement. Since modern large language models and scientific agents are capable of analyzing data, using tools, and iterating hypotheses, the researchers decided to test both these stages.
SRHarness does not dictate the search strategy. Instead, it transforms the execution environment used by agents into a reusable infrastructure consisting of three components: reusable scientific actions that wrap analysis, fitting, searching, and evaluation tools under a unified interface; a persistent scientific state that saves formula candidates and proofs outside the chat context; and trajectory lifecycle management that handles the continuation, branching, restarting, and stopping of long searches. The code is available open source on GitHub and installable via PyPI, and an online demonstration is also provided.
According to the presented papers, using the same base model DeepSeek-v4-flash, SRHarness achieved a symbolic recovery rate of 93.69% on the LSR-Transform benchmark, significantly higher than the previous best method SR-Scientist's 62.16%. Even when the scientific context and variable values are removed from the tasks, SRHarness still recovers 72.97% of the target equations, compared to 39.64% for SR-Scientist and 20.72% for Codex running on the same base model. The authors note that this result surpasses Codex coupled with GPT-5.5, which showed 68.47%.
MechBench separates two capabilities: recovering the observed law and recovering the mechanism underlying it. It uses mechanism probes to check internal quantities that the formula itself cannot determine, and it filters out competing mechanisms capable of yielding identical observations. The study results show that even Codex with GPT-5.6-sol only recovers mechanisms in 13.75% of the main set, while the recovery rate of the laws themselves is 35%.
The researchers conclude that this gap demonstrates that today's scientific agents are much better at finding patterns than they are at understanding them, positioning MechBench as a benchmark for the next step in development.