Recursive, a newly founded AI research startup, has shown that its automated research system achieved new state-of-the-art results in small language-model training and GPU kernel optimization. According to the company, the system automates the research loop: it proposes ideas, implements them, runs experiments, validates outcomes, and uses what it learns to select the next experiment.
What they achieved
Recursive reported successful use of its system on three benchmarks:
- NanoChat Autoresearch: "Train a small language model to highest performance given a small compute budget." The startup says it set a new state-of-the-art on this task.
- NanoGPT Speedrun: "Train a small language model to a certain performance as fast as possible." Recursive reports improved results here as well.
- SOL-ExecBench: "Optimize GPU kernels toward hardware limits." The company states its system pushed performance on this benchmark too.
These results are presented as demonstrations of the system’s practical operation and its ability to automate parts of the research process.
Why this matters
The team frames these outcomes as early evidence that recursive self-improvement approaches can advance AI training and infrastructure work when objectives are well-defined, measurable, and fast to evaluate. The crucial open question is whether similar automated-research methods will generalize to domains where goals are less clear, metrics are harder to obtain, or evaluations are expensive and slow.
Summary
While the demonstrations do not prove that recursive self-improvement will succeed across all problem classes, Recursive’s benchmark results serve as meaningful early indicators that an automated research loop can push frontiers in well-scoped training and optimization tasks.
