AnswersPublished16 sources
What did researchers at Tsinghua University’s Data Science and Intelligence Lab in the Department of Electronic Engineering introduce with SRHarness and MechBench, as reported on October 14, 2026, how does the Kepler–Newton distinction illustrate their goals of discovering observ
Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, as reported on Octobe Tsinghua Univers...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, as reported on Octobe
- Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, a
- ## Goals and agent support
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did researchers at Tsinghua University’s Data Science and Intelligence Lab in the Department of Electronic Engineering introduce with S. Article summary: Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whethe. Topic tags: general web, openai, chatgpt, llm, agents. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts wi
Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, as reported on October 14, 2026. 10
11 The central distinction is between finding an observed regularity and explaining why it occurs.
11
Goals and agent support
- Kepler versus Newton: The analogy separates describing how a system behaves—like Kepler’s orbital laws—from explaining that behavior through an underlying theory—like Newtonian gravity. This illustrates the projects’ stated move from fitting observations toward understanding mechanisms.
11
- SRHarness: The repository describes a domain-specific runtime in which language-model agents prepare scientific data, invoke composable analysis and fitting tools, retain evaluated hypotheses, and refine interpretable formulas over long searches.
16
- Reusable actions: The paper’s retrieved description identifies composable scientific actions as one of its three core mechanisms.
2 However, the available excerpts do not fully establish the requested details about strategy independence, persistent formula-and-evidence state outside chat context, or lifecycle controls for continuing, branching, restarting, and stopping searches. Insufficient evidence to verify those details precisely.
Availability
- GitHub: The verified repository is
tsinghua-fib-lab/SRHarness. - PyPI and online demo: I don’t know their exact package name or demo address from the available evidence; providing them would risk inventing details.
Symbolic recovery results
The authors report the following LSR-Transform results using DeepSeek-v4-flash-0731. 1
2
| Test condition | SRHarness | SR-Scientist | Codex |
|---|---|---|---|
| Scientific context and variable meanings retained | 93.69% |
62.16% |
Not established in the retrieved excerpts |
| Scientific descriptions and variable meanings removed | 72.97% |
39.64% |
20.72% with the same DeepSeek backbone |
- The October 14 report describes SRHarness’s anonymized-test performance as comparable to Codex using GPT-5.5, but the retrieved evidence does not supply GPT-5.5’s exact recovery rate.
10
- These are reported symbolic-recovery results, not evidence that the agents have recovered the underlying physical mechanisms.
8
MechBench and the interpretation
- Separate targets: MechBench asks whether an agent can recover the mechanism behind a formula, rather than merely recover the formula itself.
- Evaluation design: Insufficient evidence in the retrieved excerpts to verify precisely how its internal-quantity probes and screening of observationally equivalent mechanisms work.
- Core-set result: Codex with GPT-5.6-sol recovered laws in 35.00% of cases but mechanisms in only 13.75%.
- Implication: The reported gap supports a narrower conclusion than “AI understands science”: recovering observable patterns remains substantially easier than recovering their generating mechanisms on this benchmark.
The available evidence supports the main contributions and headline results, but not every implementation detail, distribution endpoint, or requested Codex comparison.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What did researchers at Tsinghua University’s Data Science and Intelligence Lab in the Department of Electronic Engineering introduce with SRHarness and MechBench, as reported on October 14, 2026, how does the Kepler–Newton distinction illustrate their goals of discovering observ"?
Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, as reported on Octobe
What are the key points to validate first?
Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, as reported on Octobe Tsinghua University’s Data Science and Intelligence Lab introduced two complementary contributions: SRHarness, a runtime for agents that discover mathematical formulas from data, and MechBench, a benchmark testing whether agents can recover the mechanisms behind those formulas, a
What should I do next in practice?
## Goals and agent support