New AI framework autonomously optimizes training data, architectures and algorithms — outperforming human baselines

```json { "title": "ASI-Evolve: AI Framework That Optimizes Itself", "metaDescription": "Researchers at SJTU, SII, and GAIR have built ASI-Evolve, an agentic AI framework that autonomously optimizes training data, architectures, and algorithms.", "content": "<h2>New AI Framework ASI-Evolve Autonomously Optimizes Training Data, Architectures, and Algorithms — Outperforming Human Baselines</h2>\n\n<p>Researchers at Shanghai Jiao Tong University (SJTU), the Shanghai Innovation Institute (SII), and the Generative AI Research lab (GAIR) have released a new agentic framework called <strong>ASI-Evolve</strong> that automates the full optimization loop across three of the most labor-intensive pillars of AI development: training data curation, neural architecture design, and reinforcement learning algorithm design. The paper, titled <em>ASI-Evolve: AI Accelerates AI</em>, was submitted to arXiv on March 31, 2026 (arXiv:2603.29640), and the codebase has been fully open-sourced on GitHub under GAIR-NLP. Benchmark results reported in the paper show substantial performance gains over human-designed baselines across all three domains.</p>\n\n<h2>What Is ASI-Evolve and How Does It Work?</h2>\n\n<p>At its core, ASI-Evolve is structured around a <em>learn-design-experiment-analyze</em> cycle. The system is composed of three specialized agents — a Researcher, an Engineer, and an Analyzer — working in concert. A cognition base stores accumulated human priors and past experimental insights, while a dedicated experiment database logs every trial with its motivation, code, result, and analysis. Parent selection across trials uses strategies including UCB1, greedy, random, and MAP-Elites island sampling.</p>\n\n<p>This architecture is designed to address a well-recognized bottleneck in AI research and development: the manual engineering effort required to run hypothesis, experiment, and analysis cycles repeatedly at scale. Rather than automating one slice of that pipeline, ASI-Evolve targets the entire loop simultaneously.</p>\n\n<p>According to the paper abstract, attributed to lead author Weixian Xu and colleagues: <em>"ASI-Evolve is an agentic framework that demonstrates AI-driven discovery across key AI development components, achieving superior performance in neural architecture design, data curation, and reinforcement learning algorithm design."</em></p>\n\n<p>The authors — Weixian Xu, Tiantian Mi, Yixiu Liu, Yang Nan, Zhimeng Zhou, Lyumanshan Ye, Lin Zhang, Yu Qiao, and Pengfei Liu — are affiliated with SJTU, SII, and GAIR. GAIR (Generative Artificial Intelligence Research) is part of SII, an institution dedicated to innovation in education and research.</p>\n\n<h2>Benchmark Results: How ASI-Evolve Performs Against Human-Designed Systems</h2>\n\n<h3>Neural Architecture Design</h3>\n\n<p>In neural architecture search, ASI-Evolve autonomously discovered 105 state-of-the-art linear attention architectures. The best-performing model surpassed DeltaNet by +0.97 points — a gain the authors describe as nearly 3x that of recent human-designed improvements. This result is notable not only for the margin of improvement, but for the scale of novel architectures generated without manual intervention.</p>\n\n<h3>Pretraining Data Curation</h3>\n\n<p>In the domain of pretraining data curation, ASI-Evolve's evolved pipeline improved average benchmark performance by +3.96 points. Gains on the MMLU benchmark exceeded 18 points. Data curation is typically one of the most time-consuming and expertise-intensive phases of large model development, making automated improvement in this area particularly significant from a development-cost perspective.</p>\n\n<h3>Reinforcement Learning Algorithm Design</h3>\n\n<p>In reinforcement learning algorithm design, ASI-Evolve's discovered algorithms outperformed GRPO — a widely used reinforcement learning baseline — by up to +12.5 points on AMC32, +11.67 points on AIME24, and +5.04 points on OlympiadBench. These are competitive mathematical reasoning benchmarks, and margins of this size represent meaningful leaps in automated algorithm discovery.</p>\n\n<h2>Efficiency Gains: Circle Packing and the 17-Round Result</h2>\n\n<p>Beyond the three core AI development domains, the paper provides evidence of transferability to non-AI tasks. In a combinatorial optimization benchmark — circle packing — ASI-Evolve achieved a state-of-the-art score of 2.63597 in just 17 rounds of iteration. By comparison, OpenEvolve, a competing system, required 460 rounds to reach a comparable result. This efficiency gap suggests the framework's search and learning mechanisms generalize beyond the AI research stack, at least in preliminary tests.</p>\n\n<h2>Biomedical Applications: Drug-Target Interaction Prediction</h2>\n\n<p>The paper also reports results in biomedical drug-target interaction prediction, where ASI-Evolve's evolved architecture improved performance by +1.91 AUROC and +6.94 AUROC in cold-start generalization over DrugBAN, an established baseline in that domain. Cold-start generalization — the ability to perform well on novel compounds or targets with limited prior data — is a particularly challenging and clinically relevant problem in drug discovery.</p>\n\n<p>These results are framed cautiously in the paper as preliminary evidence of transferability rather than definitive proof of broad applicability. Readers should interpret them accordingly until independent replication is available.</p>\n\n<h2>Why This Matters: The Broader Context of Automated AI Research</h2>\n\n<p>The AI research community has seen a wave of systems aimed at automating portions of the scientific and engineering pipeline. Prior frameworks have targeted specific slices: some automate publication workflows, others focus on iterative solution improvement through coding agents. ASI-Evolve's stated distinction is that it addresses all three foundational components of modern AI development — data, architectures, and learning algorithms — within a single unified system.</p>\n\n<p>The authors write: <em>"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."</em></p>\n\n<p>If this claim holds up under independent scrutiny, it represents a meaningful architectural advance in the field of AI-for-AI research — a domain concerned with using AI systems to accelerate the development of other AI systems. The commercial and research implications are significant: organizations that can compress months of manual architecture search or data pipeline tuning into automated cycles stand to reduce both time-to-deployment and engineering overhead substantially.</p>\n\n<p>At the same time, the results reported in arXiv preprints are self-reported by the research team and have not yet undergone the scrutiny of peer review. The benchmark gains are striking, but independent replication on diverse tasks and model scales will be necessary before the broader claims can be treated as settled.</p>\n\n<h2>Open Source and Accessibility</h2>\n\n<p>The decision to fully open-source ASI-Evolve on GitHub under GAIR-NLP is noteworthy. Open availability means the research community can inspect the pipeline architecture, replicate experiments, and probe the system's limitations directly — a degree of transparency that strengthens the credibility of the claims while also inviting adversarial testing. The experiment database design, which stores every trial with its motivation, code, result, and analysis, is also structured in a way that supports reproducibility.</p>\n\n<h2>What the Authors Say: Looking Ahead</h2>\n\n<p>The paper concludes with measured language about what ASI-Evolve represents at this stage. The authors write: <em>"Together, these results suggest that ASI-Evolve represents a promising step toward enabling AI to accelerate AI across the foundational stages of development, offering early evidence for the feasibility of closed-loop AI research."</em></p>\n\n<p>The framing — "promising step," "early evidence," "feasibility" — is deliberately cautious, and that caution is appropriate given the scope of the ambition. Closed-loop AI research, in which AI systems autonomously design, test, and refine the next generation of AI systems with minimal human intervention, remains an aspirational goal. ASI-Evolve appears to represent a credible early implementation of that vision across a broader surface area than prior systems, but the path from early evidence to reliable, production-grade automation is long.</p>\n\n<p>What the paper does establish, based on the benchmarks reported, is that a unified agentic system can generate competitive or superior results compared to human-designed baselines across multiple technically demanding AI development tasks simultaneously. Whether those gains hold at larger scales, across more diverse problem types, and under conditions not cherry-picked by the original authors remains an open question for the field to answer.</p>\n\n<p>For more tech news, visit our <a href=\"/news\">news section</a>.</p>\n\n<h2>Moccet Take: Why This Matters for Productivity and Personal Performance</h2>\n\n<p>Frameworks like ASI-Evolve signal an accelerating shift in how AI tools are built and improved — and that acceleration has direct downstream effects on the productivity and health tech products that millions of people use daily. As AI systems become capable of optimizing themselves across data, architecture, and learning, the pace at which smarter, more personalized wellness and performance tools reach consumers is set to increase sharply. Staying informed about these foundational shifts is not just for researchers — it's for anyone who wants to use the best tools available to manage their health, focus, and output. <a href=\"/#waitlist\">Join the Moccet waitlist to stay ahead of the curve.</a></p>", "excerpt": "Researchers at SJTU, SII, and GAIR have released ASI-Evolve, an open-source agentic framework that autonomously optimizes training data, neural architectures, and reinforcement learning algorithms within a single unified system. Benchmark results reported in the arXiv preprint show the framework outperforming human-designed baselines across all three domains, including gains of up to +12.5 points over GRPO on mathematical reasoning tasks. The paper describes ASI-Evolve as early evidence for the feasibility of closed-loop AI research.", "keywords": ["ASI-Evolve", "agentic AI framework", "automated AI research", "neural architecture search", "AI-for-AI"], "slug": "asi-evolve-ai-framework-optimizes-training-data-architectures-algorithms" } ```