When AI Starts Discovering Physical Laws by Itself
Physicists have an ancient dream: hand a computer a pile of experimental data and let it find the mathematical laws behind it. This is not far-fetched — Kepler spent twenty years distilling planetary motion laws from Tycho Brahe's observations. If a human can do it, why not AI?
This field is called symbolic regression — automatically discovering mathematical formulas from data. Since the rise of LLMs, using AI for this has become a hot direction. But reality is harsh: most systems remain stuck in a 'generate–fit–score–repeat' loop with painfully low efficiency.
A recent paper proposes an idea that stands out — letting AI 'pause and reflect' during the equation discovery process. This capacity for self-examination may be the key to AI truly learning to do scientific discovery.
---
📐 A Stubborn Problem: How to Teach AI to Discover Formulas?
Physical laws typically look like \(F = ma\) or \(E = mc^2\) — symbolic expressions with variables, constants, and operators; concise in form but profound in meaning. Reverse-engineering such formulas from data is hard because:
1. The search space is enormous. There are infinitely many ways to combine numbers and variables, and nearly all are wrong — like finding a specific grain of sand in a cosmic beach. 2. Evaluation criteria are unstable. With noisy data, a 'roughly right' formula can outscore a 'perfectly correct one with slight parameter errors,' causing AI to discard the truly good formula. 3. Memory redundancy. Many systems accumulate large numbers of failed candidate formulas as experience, but this mixed-quality memory bloats the context and makes thinking harder.
---
🔄 STRIDE's Approach: Orchestrate, Don't Just Generate
Most existing LLM equation-discovery systems follow a 'generation-centric loop': propose candidate formulas, fit parameters, score, keep the good, discard the bad, repeat. This paradigm has three fatal flaws:
1. Misjudging useful formula skeletons when fitting is unreliable — a structurally correct formula may be thrown away because parameters weren't tuned well. 2. Discarding 'nearly right' equations that need only a small repair, with no repair mechanism in place. 3. Accumulating redundant memory of failed candidates that are repetitive and carry no new information.
STRIDE's solution: stop letting 'generation' dominate everything and introduce an orchestrated framework where multiple roles each do their job. Its four core components:
- Data-aware Generation: instead of generating candidates blindly, guide generation by structural features of the data. If the data shows exponential decay, the AI prioritizes formulas containing exponential functions rather than random combinations.
- Mixed-fitting Evaluation: rather than fixing structure first and fitting parameters second, STRIDE jointly considers structural correctness and parameter-fitting quality, assessing a candidate's true potential more accurately.
- Critic-Executor Repair: the key innovation. A 'critic' role analyzes failures — is this formula wrong in structure or in parameters? — then guides an 'executor' to make targeted repairs instead of simply discarding.
- Diversity-Preserving Semantic Memory: rather than accumulating every failed candidate, STRIDE keeps only 'meaningful failures' that provide new information or represent a class of errors, automatically deduplicating to keep memory information-dense.
🎯 Results: More Accurate, More Stable, More Robust
On representative symbolic regression benchmarks and the LSR-Synth suite, STRIDE achieves significant gains in accuracy, out-of-distribution robustness, and structural recovery.
Particularly noteworthy are the ablation studies — disabling each component to measure its contribution. All four components are essential, and the critic-executor repair module shows the largest performance drop when removed. This confirms that 'repair rather than discard' delivers substantial real-world benefit.
---
🤔 A Deeper Question: What Is 'Automated Scientific Discovery'?
We usually describe scientific discovery as a 'hypothesize–experiment–verify' loop. But real discovery is far messier, more iterative, and more dependent on reflection. Often a seemingly failed experiment contains the correct clue — hidden deep, requiring a pause and re-examination. Real scientists don't mechanically collect data; they constantly reflect: Am I looking at this wrong? What am I missing?
STRIDE essentially adds this reflective capability to AI. It doesn't generate more formulas — it teaches AI to stop and ask: was my path right? Why did this seemingly correct formula turn out wrong? This metacognition may be the key dividing line for whether AI can do genuine scientific discovery.
---
🔮 What Kind of Scientist Will AI Become?
STRIDE shows how an AI with self-examination can do symbolic regression more efficiently — but it is not yet a true 'AI scientist.' A real one would also need to ask good questions, design suitable experiments, and express discovered laws in ways humans understand.
Symbolic regression is only one small step. But STRIDE proves one point: teaching AI to 'pause and reflect' may be a crucial step toward higher-level intelligence.
Kepler took twenty years to analyze Tycho's data. How long would a STRIDE-like system take — a week? A day? I don't know. But STRIDE shows a possibility: perhaps one day AI will truly become a tireless scientist that is always reflecting, always learning to think smarter. Our job as humans is to ensure its goals remain aligned with ours.
---
References
1. Su, J., Tu, S., Sun, B., & Liang, X. (2026). *STRIDE: A Self-Reflective Agent Framework for Reliable Automatic Equation Discovery*. arXiv:2605.17790. 2. Cranmer, M., et al. (2023). *Discovering symbolic laws of physics with deep learning*. Nature Communications, 14(1), 4321. 3. Udrescu, S., & Tegmark, M. (2020). *AI Feynman: A physics-inspired neural network method for symbolic regression*. Science Advances, 6(16), eaay2631. 4. Lample, G., & Charton, F. (2020). *Deep learning for symbolic mathematics*. ICLR. 5.tering, D., et al. (2022). *Symbolic regression via end-to-end neural network*. arXiv:2204.10532.