How ASI-EVOLVE Automates AI R&D to Outperform Human-Designed Models

Artificial intelligence research and development has long been constrained by a manual, resource-intensive cycle of hypothesis generation, experimentation, and analysis. A new framework developed by researchers at the Generative Artificial Intelligence Research Lab (SII-GAIR) aims to automate this process, potentially accelerating AI innovation while reducing engineering overhead. Called ASI-EVOLVE, the system autonomously optimizes three foundational pillars of AI development: training data, model architectures, and learning algorithms.

Automating the AI research loop

ASI-EVOLVE operates on a continuous “learn-design-experiment-analyze” cycle, designed to tackle the complex, interdependent challenges of AI development. Unlike previous AI tools that focused on narrow optimization tasks, this framework addresses the full research pipeline in a unified manner. The system generates hypotheses, designs experiments, runs evaluations, and distills outcomes into reusable insights—all with minimal human intervention.

The framework incorporates two key components: a “Cognition Base” and an “Analyzer.” The Cognition Base serves as the system’s foundational domain expertise, pre-loaded with human knowledge, task-relevant heuristics, and known pitfalls from existing literature. This steers the exploration toward promising directions from the first iteration. The Analyzer processes raw training logs, benchmark results, and efficiency traces, distilling them into compact, actionable insights and causal analyses that inform future iterations.

Additional modules bring the framework together. A “Researcher” agent reviews prior knowledge from the cognition base and past experimental results to generate new hypotheses, proposing localized code modifications or writing new programs. The “Engineer” component runs the actual experiments, equipped with efficiency measures like wall-clock limits and early rejection tests to filter out flawed candidates before they consume excessive computational resources. A “Database” acts as the system’s persistent memory, storing code, research motivations, raw results, and the Analyzer’s reports for every iteration, ensuring insights compound systematically over time.

Outperforming human-designed baselines

In experiments, ASI-EVOLVE demonstrated the ability to autonomously discover novel designs that significantly outperformed state-of-the-art human baselines across multiple domains. The framework’s impact was particularly notable in three key areas: data curation, neural architecture design, and reinforcement learning algorithm development.

For data curation, ASI-EVOLVE was tasked with designing category-specific cleaning strategies for massive pretraining corpora. The system inspected data samples, diagnosed quality issues such as HTML artifacts and formatting inconsistencies, and autonomously formulated custom curation rules. The result was a systematic cleaning approach combined with domain-aware preservation rules, which proved far more effective than aggressive filtering. Models trained on the AI-curated data saw an average performance boost of nearly 4 points over models trained on raw data. The gains were most pronounced in knowledge-intensive tasks, with performance increasing by over 18 points on the Massive Multitask Language Understanding (MMLU) benchmark, which covers STEM, humanities, and social sciences.

In neural architecture design, ASI-EVOLVE generated 105 novel linear attention architectures across 1,773 autonomous exploration rounds, all of which surpassed DeltaNet, a highly efficient human-designed baseline. The system developed multi-scale routing mechanisms that dynamically adjust the model’s computational budget based on input content, demonstrating its ability to innovate beyond static architectural designs.

For reinforcement learning algorithm design, ASI-EVOLVE discovered novel optimization mechanisms that outperformed the competitive GRPO baseline on complex mathematical reasoning benchmarks. One successful variant invented a “Budget-Constrained Dynamic Radius” that keeps model updates within a defined budget, effectively stabilizing training on noisy data. The discovered algorithms achieved performance gains of up to 12.5 points on AMC32, 11.67 points on AIME24, and 5.04 points on OlympiadBench.

The data and design bottleneck

AI research and development has traditionally been limited by the sheer complexity of exploring the vast design space for models. Engineering teams can only test a fraction of possible configurations due to the high costs of manual effort, computational resources, and the siloed nature of insights gained from experiments. These constraints fundamentally limit the pace and scale of AI innovation.

While AI has made significant strides in scientific discovery—from specialized tools like AlphaFold solving discrete biological problems to agentic systems answering basic scientific questions—current frameworks still struggle with open-ended AI innovation. Most existing systems are limited to narrow optimization within very specific constraints. Advancing core AI capabilities requires modifying large interdependent codebases, running compute-heavy experiments that consume tens to hundreds of GPU hours, and analyzing multi-dimensional feedback from training dynamics.

Existing frameworks have not yet demonstrated that AI can operate effectively in this regime in a unified way, nor that it can generate meaningful advances across the three foundational pillars of AI development rather than within a single narrowly scoped setting.

Implications for enterprise AI workflows

Enterprise AI workflows often require constant optimizations to existing systems, from fine-tuning open-source models on proprietary data to making incremental changes to architectures and algorithms. The computational resources and engineering hours required for such efforts are typically immense, placing them beyond the reach of most organizations. Many enterprises are left running unoptimized versions of standard AI models.

The ASI-EVOLVE framework offers a potential solution to this challenge. The research team designed the system so enterprises can integrate proprietary domain knowledge into the cognition repository, allowing the autonomous loop to iterate on internal AI systems. This could enable organizations to achieve performance gains without the need for extensive manual engineering effort.

The researchers have open-sourced the ASI-EVOLVE code, making the foundational framework available for developers and product builders. The system’s ability to systematically learn from complex experimental feedback without constant human intervention could mark a turning point in how AI systems are developed and refined. While previous frameworks focused on evolving candidate solutions, ASI-EVOLVE evolves cognition itself, continuously storing and retrieving accumulated experience to inform future exploration.

As AI continues to advance, frameworks like ASI-EVOLVE may play a crucial role in bridging the gap between the rapid pace of innovation and the practical constraints of manual research and development. By automating the optimization loop, the system not only accelerates AI progress but also democratizes access to cutting-edge improvements, potentially leveling the playing field for organizations with limited resources.