If it feels like AI is developing too fast to keep up with, a group of Chinese researchers have some bad news – because they've developed a model that "evolves" on its own, creating better versions of itself with each self-analytical loop.
ASI-Evolve, built by researchers at Shanghai Jiao Tong University, works by running a continuous loop that mirrors how humans would put this type of technology through its paces. Essentially, it creates variations of AI models, alters how they're trained and adjusts the data they learn from. It then runs its own experiments to see which clone performs better, using those results to guide what it tries next.
"ASI-Evolve augments standard evolutionary agents with two key components: a cognition base that injects accumulated human priors into each round of exploration, and a dedicated analyzer that distills complex experimental outcomes into reusable insights for future iterations," the researchers wrote. "To our knowledge, ASI-Evolve is the first unified framework to demonstrate AI-driven discovery across three central components of AI development: data, architectures, and learning algorithms."
While that means very little to a lot of people, there's a reason this development has attracted a lot of buzz in the industry. By generating ideas, testing them and refining the results in a self-improving loop, ASI-Evolve mirrors the trial-and-error process of not just AI model building, but also science and math research. As such, it raises the possibility of accelerating discoveries in fields where progress is slow due to human researchers testing many possible outcomes.
"What if you could run a tireless AI researcher on your hardest problem – one that reads the literature, designs experiments, runs them, and learns from every failure? That's ASI-Evolve," the researchers noted on GitHub where the model's assets are hosted. "It is a general agentic framework that closes the loop between knowledge → hypothesis → experiment → analysis – and repeats it autonomously, round after round, until it finds something that works.
"We built it for AI research," they added. "But the loop doesn't care about domain. A financial analyst, a biomedical engineer, a climate scientist, or a game developer can all plug their own problem into ASI-Evolve and let it search for better solutions than any human has time to manually explore."
ASI-Evolve was able to improve a specific function – its attention mechanism – by 0.97 points on a standard benchmark test, compared to 0.34 points achieved by a human. The “points” refer to scores on that test, where even small increases are considered meaningful. So while this was only one test on one aspect of the AI build, it was nearly three times faster at improving itself.
What's more, when used as a drug discovery model, ASI-Evolve outperformed existing systems, demonstrating its promise that goes beyond AI. If you want to crunch the numbers, this video is an excellent recap of the paper's findings:
And, no, this system is not going to take anyone's job – it still requires humans to oversee its "evolution" – there's a reason why it's an exciting development.
"In ASI-Evolve, we introduced a large amount of human prior experience," researcher Xu Weixian told China's 36Kr, the country's TechChrunch equivalent. "We don't pursue 'blind evolution' without human guidance because the initial experimental purpose and core ideas are always proposed by humans. The real value of the system lies in using AI's strong exploration ability to iterate rapidly in the direction guided by humans. It is more like an extremely efficient collaborative system rather than a cold substitute. ASI-Evolve promotes people to shift from problem-solving and repair to problem definition."
It's worth noting that the researchers haven't detailed energy costs of running ASI-Evolve, but its speed and efficiency, and closed-loop self-learning, suggests it's nowhere near as power-hungry as leading models trained on enormous datasets. AI agents are expected to drive China's next stage of development – one where new data centers are also mandated to be powered by green tech.
The research has been published on arXiv.
Source: Shanghai Jiao Tong University via GitHub
