AutoResearch Lite: a reproducible paper-inspired experiment runner (Adam mechanism reproduction and convergence-speed comparison)
He, Zhen · 2026 · 49999 citationsRead the paper
A dependency-free experiment runner that turns a paper's research question into a configuration-driven, verifiable experiment task. The bundled reproduction maps the first/second moment estimates and bias correction of Adam (Kingma & Ba, arXiv:1412.6980) into runnable Python and compares them against fixed-learning-rate SGD under fixed seeds and epoch budgets. A second experiment measures convergence speed (epochs to reach a target training loss) against SGD with heavy-ball momentum, AdaGrad and RMSProp, with every optimizer family tuned by the same coarse learning-rate sweep. In that experiment Adam is not the fastest; the negative result is reported in full. Every reported number is pinned in an expected-values file with explicit tolerances and checked by one command that also runs the unit tests, and the checking harness is validated against deliberately broken implementations (negative controls). The repository also ships a task contract that expresses the reproduction as a repeatable, scoreable task for an automated research loop.
What the full text shows
Shown in the lab- Existence of a 'sweet-spot' for the number of segments; exceeding this limit impairs the ability to find optimal designs due to increased problem complexity. (Search for digit transformer designs with total length ~8000μm and 100 generations)
- Small length scales limit the device's ability to bend into large angles, restricting outputs to straight or slightly bent shapes. — Score 1s for 800μm length scale (Search with 12 segments and 100 generations)
- Discovery of material designs capable of transforming into multiple different Arabic numerals (0-9) across different actuation states. (Using 4 active materials with orthogonal DNA actuation systems)
- Setup
- Simulation of bilayer hydrogel segments (approx. 500μm height, 250μm width) with 16 possible actuation states; CNN trained on MNIST and 24,000 artificial images.
- How
- A coarse-grained numeric simulation of DNA-co-polymerized hydrogel shape change was integrated with a genetic algorithm for design optimization. A CNN classifier, trained on MNIST and 24,000 manually labeled simulated images, provided autonomous scoring for the genetic algorithm's selection process.
- Limitations
- Algorithm unable to find good designs when system complexity (number of segments) is too high · Simulation assumes little stress along the horizontal axis and between segments · Physical fabrication limits imposed by multistage photolithography
Summarized by Gemma 4 from the open-access full text · 2 Oct 2026. Read it