SimpleTES discovers better solutions by automating scientific trial-and-error loops

Who: Posted by @KevinQHLin, sharing work announced by @haotian_yeee. Kevin Lin appears to be amplifying the research; Haotian Ye is the lead voice on SimpleTES, a system for automating scientific discovery using AI.

What's new: A system called SimpleTES automates the loop that scientists run manually — propose a solution, evaluate it, refine based on results, repeat — and does this at scale across 21 different scientific problems. Most prior AI research tools focus on generating more text or running more parallel agents. SimpleTES instead treats the evaluation step as the engine of progress, letting the system learn from scored outcomes rather than just producing more output.

How it works: SimpleTES drives an to propose candidate solutions, then passes each proposal through a domain-specific evaluator that scores it objectively. That score feeds back into the next round of proposals — the same way a student revises an essay after seeing a grade, except the loop runs thousands of times automatically. The key insight is that science is bottlenecked not by idea generation but by rigorous evaluation, so hardwiring evaluation into the loop is what unlocks real progress.

The numbers: On regression, SimpleTES produced a solver 2.17x faster than . On quantum circuit routing for , it cut overhead by 24.5% versus . It also set new records on the and on denoising data from the Tabula Muris mouse-tissue dataset, generalizing to tissue types it had never seen during training.

Why it matters: Beating decades-old hand-engineered solvers and open math problems across domains that have nothing in common — statistics, quantum computing, combinatorics, genomics — suggests the evaluation-driven loop is a genuinely general scaffold, not a trick tuned to one field. If it holds up under scrutiny, it points toward a future where the bottleneck in scientific research shifts from human iteration speed to compute budget.