Shanghai Jiao Tong University's groundbreaking ASI-Evolve framework marks a pivotal moment in artificial intelligence research, where AI now autonomously drives its own evolution. Developed by a team of researchers led by Weixian Xu at the university's GAIR-NLP lab, this self-improving system mimics the human process of trial-and-error learning to accelerate discoveries not just in AI but potentially across scientific domains. By automating the full research cycle—reading literature, hypothesizing, experimenting, and analyzing—ASI-Evolve has demonstrated superior performance over human-designed methods, outperforming baselines in neural architecture search, data curation, and reinforcement learning algorithm design.

In the competitive landscape of global AI advancement, Chinese universities like Shanghai Jiao Tong University (SJTU) are positioning themselves at the forefront. SJTU, one of China's elite C9 League institutions, has long been a hub for cutting-edge AI research, boasting state-of-the-art facilities and collaborations with industry giants. The release of ASI-Evolve on March 31, 2026, via arXiv underscores the university's commitment to open-source innovation, making this powerful tool freely available on GitHub for researchers worldwide to build upon.

Understanding the ASI-Evolve Framework

At its core, Artificial Superintelligence-Evolve (ASI-Evolve) is an agentic framework that closes the loop on AI-for-AI research. Traditional AI development relies on human researchers painstakingly iterating through ideas, but ASI-Evolve automates this with a structured learn-design-experiment-analyze cycle. This human-like trial-and-error process begins with sampling historical data from a persistent database and retrieving relevant human priors from a cognition base—a repository of distilled insights from literature and past experiments.

The Researcher agent, powered by large language models, generates candidate programs or code artifacts conditioned on this context. These are then executed by the Engineer in real environments, with early stopping for inefficient ideas. Finally, the Analyzer distills outcomes into actionable reports, feeding back into the cognition base for future rounds. This self-reflective loop ensures continuous improvement without human intervention beyond initial setup.

Breakthroughs in Neural Architecture Design

One of ASI-Evolve's most striking achievements is in neural architecture search, particularly for linear attention mechanisms. Starting from the DeltaNet baseline, the framework ran 1,773 exploration rounds, generating 1,350 candidates. It discovered 105 state-of-the-art (SOTA) architectures, with the top performer—PathGateFusionNet—surpassing DeltaNet by 0.97 points on benchmarks. This gain is nearly three times the 0.34-point improvement from recent human-designed models like Mamba2, highlighting AI's edge in exhaustive exploration.

Key innovations include adaptive multi-scale routing and content-aware gating, which enhance efficiency across small (20M params), medium (340M), and large (1.3B) models. Development benchmarks improved by 57.28%, and generalization by 45.40%, proving the architectures' robustness. At SJTU, this builds on the university's strong computer science program, where faculty and students collaborate on scalable AI systems.

For more details on the paper, see the full arXiv preprint.

Revolutionizing Data Curation for Pretraining

ASI-Evolve also evolved a pretraining data curation pipeline for the Nemotron-CC dataset (672B tokens), producing Nemotron-CC ASI+ (504B tokens). By iteratively refining category-specific cleaning strategies—addressing issues like HTML artifacts and PII—the system achieved a 3.96-point average benchmark uplift. Standout gains included +18.64 on MMLU (multiple-choice language understanding) and +18.80 on CSQA (commonsense QA), outperforming human-curated datasets like FineWeb-Edu.

This step-by-step process—identifying flaws, proposing filters, validating on subsets, and analyzing coverage—mirrors how human data scientists operate but at superhuman speed. In Chinese higher education, where massive datasets fuel national AI initiatives, such tools could democratize high-quality data preparation across universities.

Advancing Reinforcement Learning Algorithms

In reinforcement learning for mathematical reasoning, ASI-Evolve targeted improvements to Group Relative Policy Optimization (GRPO). After exploration on a 4B model and verification on 14B, it yielded algorithms outperforming GRPO by up to 12.5 points on AMC32, 11.67 on AIME24, and 5.04 on OlympiadBench. Innovations like pairwise asymmetric optimization and percentile normalization with global update budgets stabilized training while boosting performance.

These results, derived from 10 post-GRPO papers in the cognition base, showcase AI's ability to synthesize and extend human knowledge mathematically—a boon for SJTU's math-AI interdisciplinary programs.

Beyond AI: Transfer to Other Sciences

ASI-Evolve's generality shines in non-AI domains. On circle packing—a combinatorial optimization benchmark—it hit SOTA (2.63597) in just 17 rounds, faster than competitors like OpenEvolve (460 rounds). In biomedicine, an evolved architecture improved drug-target interaction prediction by +1.91 AUROC and +6.94 in cold-start generalization over DrugBAN.

Researchers note its plug-and-play potential for finance, climate modeling, or game design, positioning Chinese universities like SJTU as leaders in versatile AI tools. The open-source release fosters global collaboration, amplifying China's soft power in higher ed.

SJTU's Role in China's AI Higher Education Landscape

Shanghai Jiao Tong University, founded in 1896, exemplifies China's push toward AI supremacy. Ranked among the world's top 50 for computer science, SJTU hosts labs like GAIR (Generative AI Research), where the ASI-Evolve team—Weixian Xu, Tiantian Mi, Yixiu Liu, and others—thrives. Government backing via the 14th Five-Year Plan fuels such projects, with billions invested in university AI infrastructure.

This breakthrough aligns with national goals for self-reliance in tech, training thousands of AI specialists annually. Other C9 unis like Tsinghua and Peking echo this, but SJTU's focus on agentic systems sets it apart, potentially reshaping curricula toward AI-augmented research.

Explore SJTU's innovations further via their official site.

Challenges and Ethical Considerations

While promising, ASI-Evolve raises questions. Compute demands for large-scale runs remain high, though efficient sampling (e.g., MAP-Elites) mitigates this. Ethical oversight is crucial—human validation prevents biases in the cognition base. In China, regulations like the 2023 AI Law ensure safe deployment, with universities emphasizing responsible AI education.

Stakeholders, including ethicists at SJTU, stress hybrid human-AI teams: AI handles iteration, humans define goals. This balances acceleration with accountability.

Photo by Mehyar Belal on Unsplash

Future Outlook for Self-Evolving AI in Academia

ASI-Evolve heralds an era where AI researchers become orchestrators, not laborers. In Chinese higher ed, expect widespread adoption: imagine Tsinghua using it for quantum simulations or Fudan for drug discovery. Globally, it could narrow gaps, enabling smaller labs to compete.

Projections: By 2030, self-evolving agents may double research output in universities. For students, this means curricula blending AI tools with critical thinking. China's lead, via open-source like ASI-Evolve, invites collaboration—check the GitHub repo to experiment yourself.

For more on New Atlas coverage, visit their article.