Aster: Autonomous Scientific Discovery over 20x Faster Than Existing Methods

AI summary

Overview

Research area: Autonomous scientific discovery — AI agents built on Large Language Models that iteratively write, test, and refine code to solve open research problems.

Technical level: Intermediate. The paper is short and readable, but comfortably assumes familiarity with benchmarks like NanoGPT Speedrun, GPU kernel optimization, and scRNA-seq denoising.

Scope: The paper presents Aster, an autonomous discovery agent, and reports a speedup benchmark plus five applied results spanning mathematics, kernel engineering, biology, and machine learning.

What This Paper Is About

Existing autonomous discovery systems can produce genuine scientific breakthroughs, but they typically need hundreds to thousands of iterations to get there. That makes them impractical for problems where a single evaluation takes hours — such as training a machine learning model. Aster is an agent that takes an initial program, an evaluator script, and a prompt, then iteratively refines the program, with the central claim that it reaches discovery in far fewer iterations than prior frameworks.

Key Contributions

1. A >20x iteration speedup. On the circle packing problem (packing 26 circles), Aster matched a threshold OpenEvolve needed 115 iterations to reach in just 5 iterations — a 23x figure in Figure 1. Both systems were configured with the same model distribution (80% Gemini 2.0 Flash, 20% Claude 3.7 Sonnet) for a fair comparison.
2. Four state-of-the-art results and one match across five domains. Aster reports SOTA on the Erdős minimum overlap problem, the TriMul GPU kernel, single-cell denoising, and the NanoGPT Speedrun, and matches best human performance on ZAPBench.
3. Demonstration on a multi-hour evaluation task. ZAPBench training runs take hours to evaluate, and Aster reached a matching result in roughly two and a half days — a setting the paper argues would have taken over a month without the speedup.
4. A deployable system. Aster is offered via a web interface and API at asterlab.ai, with all generated programs listed in Appendix A.

Main Findings

• Erdős minimum overlap: Aster surpassed the TTT-Discover record (0.380876) to reach 0.380874 in 40 iterations, starting from a highly basic initial program. The best human result cited is 0.380927. Aster's construction used 8192 pieces, versus 600 pieces in the previous state-of-the-art solution.
• TriMul kernel optimization: Over 70 iterations, Aster reduced TriMul forward-pass runtime on an NVIDIA H100 to 1114 µs, beating the TTT-Discover benchmark of 1161 µs. The best human figure listed is 1371 µs. The initial program was the best result from a prior 94-iteration run performed on a different GPU image.
• Single-cell denoising: On the OpenProblems benchmark, Aster improved on TTT-Discover (0.709) to reach 0.711 in 30 iterations. On the PBMC dataset, Aster's MSE was 0.150 and Poisson 0.049, versus 0.190 / 0.050 for the best human method (MAGIC) and 0.154 / 0.048 for TTT-Discover.
• NanoGPT Speedrun: In 8 iterations, Aster shaved 1.6 seconds off the record, going from 96.8 to 95.2 seconds. The competition trains a language model to under 3.28 cross-entropy loss on the FineWeb validation dataset on a node of 8 NVIDIA H100 GPUs. The gains came from refinements to Triton kernels, optimizing memory load-ins and avoiding unnecessary recomputation.
• Prior AI contributions to the same record: Locus contributed a 0.9% speedup and Hiverge 1.3%, making Aster's 1.6% the largest of the three; Aster is the third AI system to contribute to the record after those two.
• ZAPBench: The best human model is a UNet trained for 36 hours on 16 A100s, achieving an MAE of 0.0182. After 34 iterations (one 20-iteration run with a one-hour timeout, and a 14-iteration run with a three-hour timeout), Aster produced a model with the same MAE of 0.0182, matching human performance with 190x less compute. The evaluation script ran on an NVIDIA T4 GPU.
• ZAPBench caveat and context: Aster was still improving when the run was cut off. A separate tree-search system by Aygün et al. reached an MAE of 0.0176 under similar runtime constraints — lower than Aster's matched result.
• Circle packing details: OpenEvolve attempted the problem 460 times but generated programs 4x in parallel, counted as 115 iterations, reaching a score of 2.634. Aster surpassed this with 2.6353 in 5 iterations, and — using its default model configuration — reached the known state-of-the-art of 2.635983 in 6 iterations.

Methodology in Plain English

Aster is an autonomous agent in a loop. The user supplies three things: a starting program, a script that evaluates how good a program is, and a prompt describing the task. Aster proposes an improved version of the program, the evaluator scores it, the result is stored in a database, and the process repeats. The database of past attempts is what the agent draws on when proposing the next version.

The paper's argument is that this loop is far more iteration-efficient than existing frameworks, so the total wall-clock cost stays manageable even when a single evaluation is expensive. The authors test that claim directly on circle packing against OpenEvolve, deliberately holding the model mix constant so the comparison reflects the agent rather than the underlying LLM.

For the applied tasks, the evaluators and prompts were largely adapted from the TTT-Discover repository, which the authors credit. Initial programs varied by task: a basic script for the overlap problem, the best human script for single-cell denoising, and a prior run's best program for the TriMul kernel. For ZAPBench, the task was framed as the short-context benchmark predicting one step ahead, with Mean Average Error as the objective.

Why This Matters

If the number of iterations needed for a discovery drops by more than 20x, the set of problems an autonomous system can realistically attack expands. Tasks that were previously excluded purely because each evaluation takes hours become tractable. That is the paper's distinctive claim: not just faster convergence, but access to a different class of problems.

Real-world applications:

• Protein structure prediction — the TriMul kernel is a core computational primitive in the AlphaFold architecture, so kernel-level speedups propagate to that workload.
• Genomics and precision medicine — denoising single-cell RNA sequencing data recovers true gene expression profiles from sparse, dropout-heavy measurements, improving the value of expensive sequencing experiments.
• Neuroscience — cellular-resolution forecasting of neural activity across a larval zebrafish brain, the setting of ZAPBench.
• Large language model training — the NanoGPT Speedrun advances are directly relevant to training efficiency; the paper notes the Muon optimizer emerged from this benchmark lineage and was used to train Kimi K2.

Industry relevance: the system is offered as a product via web interface and API, and the kernel-optimization and training-speedup results point at direct cost reductions for AI infrastructure.

Future Directions

• Push ZAPBench further. Aster was still improving when its run was terminated, and the authors state it is highly likely a longer runtime would lower the MAE below 0.0182. A separate system (Aygün et al.) already reported 0.0176.
• Establish the limits of the speedup claim. The 20x+ figure is demonstrated on a single benchmark, circle packing. Whether comparable iteration efficiency holds across problem types is not established by the reported evidence.
• Extend to other long-evaluation domains. The paper's core motivation is that long-evaluation tasks are now reachable; only ZAPBench among the reported tasks tests that premise directly.
• Compare against the broader landscape of discovery frameworks. The head-to-head comparison is with OpenEvolve, and baseline results elsewhere are drawn from TTT-Discover rather than reproduced.

Target Audience

Researchers and engineers working on AI-for-science systems, LLM-based agents, and automated program search will find the closest match to their interests. It is also relevant to practitioners in GPU kernel engineering, computational biology, and AI training infrastructure who care about benchmark performance in those specific areas, and to anyone evaluating whether autonomous discovery tooling is mature enough for problems with expensive evaluation loops. The paper is short and accessible, though the Appendix A program listings are referenced rather than reproduced in the available content.

Authors’ abstract

We introduce Aster, an AI agent for autonomous scientific discovery capable of operating over 20 times faster than existing frameworks. Given a task, an initial program, and a script to evaluate the performance of the program, Aster iteratively improves the program, often leading to new state-of-the-art performances. Aster's significant reduction in the number of iterations required for novel discovery expands the domain of tractable problems to include tasks with long evaluation durations, such as multi-hour machine learning training runs. We applied Aster to problems in mathematics, GPU kernel engineering, biology, neuroscience, and language model training. More specifically: the Erdos minimum overlap problem, optimizing the TriMul kernel, a single-cell analysis denoising problem, training a neural activity prediction model to perform well on ZAPBench, and the NanoGPT Speedrun Competition. Aster attains SOTA results in every task, except for ZAPBench, where it matches the performance of the best human solution with less than 1/190th of the compute. Aster is accessible via a web interface and API at asterlab.ai.