Tsinghua Unveils SRHarness and MechBench: AI Agents That Look for Formulas and Mechanisms

Two works from the same Tsinghua lab target a bottleneck of scientific AI agents: they can generate plausible hypotheses, but they are poor at showing that a formula really follows from the data. According to Jiqizhixin, the Data Science and Intelligence Lab at Tsinghua University's Department of Electronic Engineering has introduced SRHarness, a runtime framework for agentic symbolic regression, and MechBench, a benchmark for evaluating the ability to find the mechanisms behind phenomena.

What exactly was proposed

Symbolic regression tackles a problem that sounds simple: take a table of observations and recover an analytical expression rather than fit a black box. You can read such an expression, check it for dimensional consistency and plug it into the next calculation. Classical methods rely on genetic programming and search. The agentic approach replaces search with a loop in which a language model proposes a form, runs the code, looks at the error and revises the hypothesis.

Based on the description, SRHarness operates at the execution level, that is, around the agent itself. It defines how the agent handles data, validates candidates and makes decisions. The goal is stated plainly: make formula discovery more reliable. Anyone who has run agents on long tasks knows the problem. Without a rigid harness, they overfit to noise, forget earlier attempts and produce a beautiful equation that falls apart outside the training sample.

MechBench addresses a related problem. Finding a formula that describes the data well and understanding why the system behaves the way it does are different things. The mechanism benchmark tests the latter: whether an agent can move from a curve to the causal structure of the process.

Why this is harder than it looks

Symbolic regression has two classic traps. The first is non-uniqueness: dozens of expressions with different structures give almost the same error on the available points. The second is noise and limited observations, especially in physics, chemistry and biology, where experiments are expensive. An LLM-based agent adds a third: it tends to pick a formula that resembles one familiar from its training data, even when the real dependence is different.

A separate mechanism benchmark matters for the same reason. Metrics such as approximation error measure fit well and understanding poorly. If an evaluation does not require recovering the mechanism, an agent can win by memorizing standard laws.

What is known so far

The available portion of the announcement contains no comparison results, no model names and no sizes of the task sets. Here is what is known:

• Authors: the Data Science and Intelligence Lab, Department of Electronic Engineering, Tsinghua University.
• SRHarness: a runtime framework for agentic symbolic regression, aimed at reliable derivation of formulas from observational data.
• MechBench: a benchmark that evaluates the mechanism, not only the formula.
• Field: AI for Science (AI4S), scientific AI agents.
• Publication: a Jiqizhixin report dated October 14, 2026; the text references two separate works.
• Quantitative metrics, the number of tasks and the baseline methods for comparison are not given in the source.

How to read releases like this

Until the full papers are out, it makes sense to keep a few questions in mind. Which baselines were chosen for comparison: classical ones such as genetic algorithms, or other agentic designs built on the same LLMs? Does the benchmark separate memorization of known laws from genuine discovery, for example through synthetic or modified systems? What compute budget does the agent spend per task, and how does quality grow with the number of iterations?

For teams building their own scientific agents, the practical takeaway lies elsewhere. Reliability depends not only on the model but also on the harness: hypothesis logging, validation on held-out data, limits on expression complexity. If SRHarness and MechBench are released with open code and data, they could serve as a testbed for such solutions. Keep an eye on the preprint and the repository: that is where the numbers will appear that show how well the claimed reliability holds up.