Researchers at the Data Science and Intelligence Lab in Tsinghua University's Department of Electronic Engineering have released two linked pieces of AI for Science work: SRHarness, a runtime framework for agentic symbolic regression, and MechBench, a benchmark that asks whether AI can recover the mechanisms behind a formula rather than only the formula itself.
The team frames the work with a historical example. Kepler derived the laws of planetary motion from observations, describing how planets move; Newton's mechanics explained why. Large language models and scientific agents can now analyze data, call tools and iterate on hypotheses, so the researchers set out to test both steps.
SRHarness does not prescribe a search strategy. Instead it turns the runtime that agents rely on into reusable infrastructure with three parts: reusable scientific actions that wrap analysis, fitting, search and evaluation tools behind one interface; persistent scientific state that stores candidate formulas and their evidence outside the chat context; and trajectory lifecycle management that handles continuing, branching, restarting and stopping long searches. The code is open source on GitHub, installable from PyPI, and has an online demo.
According to the papers, with the same DeepSeek-v4-flash base model, SRHarness reaches a 93.69% symbolic recovery rate on the LSR-Transform benchmark, compared with 62.16% for the previous best method, SR-Scientist. When scientific context and variable meanings are removed from the problems, it still recovers 72.97% of target equations, versus 39.64% for SR-Scientist and 20.72% for Codex running on the same base model. The authors report that this also exceeds Codex paired with GPT-5.5, at 68.47%.
MechBench separates two abilities: recovering an observed law and recovering the underlying mechanism that produces it. It uses mechanism probes to check internal quantities a formula alone cannot determine, and screens competing mechanisms that could produce identical observations. In the paper's results, even Codex with GPT-5.6-sol recovers mechanisms in only 13.75% of core-set cases, against a 35% rate for recovering the laws themselves.
The researchers say the gap shows today's scientific agents are much better at finding patterns than at understanding them, and position MechBench as a yardstick for that next step.
SEE ALSO: Tsinghua AIR Zeva Lifts Embodied Manipulation Success From 26% to 73% Without Retraining