AI R&D works through a cycle of hypothesis, experimentation, and analysis—each step requiring significant manual engineering effort. A new framework by SII-GAIR researchers aims to close this bottleneck by automating the complete optimization loop for training data, model architectures, and learning algorithms.
It’s called a new frame ASI-EVOLVEDeveloped by researchers at the Generative Artificial Intelligence Research Laboratory (SII-GAIR), it aims to solve this bottleneck. Designed as an agent system for AI-for-AI research, continuous use "study-design-experiment-analysis" cycle to automate core AI stack optimization.
In experiments, this self-improvement loop has autonomously discovered new designs that have vastly outperformed state-of-the-art human bases. The system created new language model architectures, improved preprocessing data pipelines to increase benchmark scores by more than 18 points, and developed highly efficient reinforcement learning algorithms.
For enterprise teams performing iterative optimization cycles on AI systems, the framework offers a way to reduce manual engineering costs while meeting or exceeding human-engineered baselines.
Information and design bottleneck
Engineering teams can only explore a small portion of the vast design space possible for AI models at any given time. Running experimental workflows requires expensive manual efforts and frequent human intervention. And the insights gained from these expensive cycles are often written off as individual intuition or experience, making it difficult to systematically preserve and transfer this knowledge to future projects or different teams. These limitations fundamentally limit the speed and scale of artificial innovation.
Artificial intelligence has made incredible progress in scientific discoveries AlphaFold solving discrete biological problems into agent systems that answer fundamental scientific questions. However, existing frameworks still struggle with open AI innovation and are mostly limited to narrow optimization within very specific constraints.
Developing basic AI capabilities is more complex. This requires modifying large interdependent codebases, running computationally intensive experiments consuming tens to hundreds of GPU hours, and analyzing multidimensional feedback from training dynamics.
“Existing frameworks have yet to demonstrate that AI can work uniformly effectively in this mode, or that it can make significant progress across the three main pillars of AI development,” the researchers write.
How ASI-EVOLVE learns to conduct research
To overcome the limitations of manual R&D, ASI-EVOLVE works on a continuous loop between prior knowledge, hypothesis generation, experimentation and improvement. The system learns relevant knowledge and historical experience from existing databases, develops candidate software that represents its next hypothesis, runs experiments to obtain evaluation signals, and parses the results back into its knowledge base into reusable, human-readable lessons.
There are two main components that drive ASI-EVOLVE. “Cognitive base” acts as the main domain expertise of the system. To speed up the search process, the system is preloaded with human knowledge, task-related heuristics, and known pitfalls extracted from the existing literature. This directs exploration in promising directions from the first iteration.
The second component is the “Analyzer”, which resolves complex, multidimensional feedback from experiments. It processes raw training records, benchmark results, and efficiency traces, turning them into concise, actionable insights and cause-and-effect analyses.
Several other complementary modules bring the framework together. An “exploratory” agent reviews prior knowledge from the cognitive base and past experimental results to generate new hypotheses, either to suggest localized code changes, or to write new programs.
The “Engineer” component performs the actual experiments. Because AI training tests are incredibly expensive, Engineer is equipped with efficiency measures such as wall-clock limits and early-failure quick tests to filter out flawed candidate programs before they consume excessive GPU hours.
Finally, the “Data Base” serves as the persistent memory of the system, storing code, research motifs, raw results, and Analyzer final reports for each iteration, enabling systematic integration of insights over time.
By combining these components, ASI-EVOLVE enables an artificial intelligence agent to systematically learn from complex, real-world experimental feedback without requiring constant human intervention.
While previous frameworks were designed to improve candidate solutions, “ASI-EVOLVE evolves cognition itself,” the researchers write. “Gathered experience and distilled insights are continuously stored and retrieved to inform future exploration, enabling the system to grow not only in the quality of solutions, but also in its ability to think about where to look next.”
ASI-EVOLVE is in action
Researchers have shown in their experiments that ASI-EVOLVE can successfully improve data curation, model architecture, and learning algorithms to build better AI systems.
For real-world enterprise applications, high-quality data is a persistent bottleneck. When tasked with developing category-specific cleanup strategies for a mass production corporation, ASI-EVOLVE examined data samples and diagnosed quality issues such as HTML artifacts and formatting inconsistencies. The system autonomously designed custom curation rules, finding that systematic cleaning combined with domain-aware protection rules was more effective than aggressive filtering.
In benchmarks, 3D parameter models trained on AI-generated data saw an average improvement of about 4 points over models trained on raw data. Performance on Massive Multitask Language Understanding (MMLU), an LLM standard that includes tasks in STEM, humanities and social sciences, increased by more than 18 points, with the highest gains in knowledge-intensive tasks.
In addition to data, the system has proven to be highly adept at designing neural architectures. In 1,773 rounds of autonomous exploration, it created 105 new linear attention architectures that outperformed the highly efficient human-designed DeltaNet. To achieve these results, ASI-EVOLVE developed multiscale routing mechanisms that dynamically adjust the model’s computational budget based on the specific content of the input.
Finally, in reinforcement learning algorithm design, ASI-EVOLVE discovered new optimization mechanisms. He has developed algorithms such as AMC32 and AIME24 that outperform the competitive GRPO database on complex mathematical reasoning criteria. He invented one successful option "Dynamic Radius on a Budget" keeps model updates within a specified budget, effectively stabilizing training on noisy data.
What this means for enterprise AI
Enterprise AI workflows constantly require optimization of existing systems, from fine-tuning open-source models on private data to making minor changes to architecture and algorithms. Typically, the computational resources and engineering hours required to implement such efforts are large and beyond the reach of most organizations. As a result, many are left to run unoptimized versions of standard AI models.
The research team says the framework is designed so that enterprises can integrate their own domain knowledge into a cognitive store and allow the autonomous circuit to replicate in-house AI systems.
There is a research team open source ASI-EVOLVE codemaking the core framework available to developers and product builders.
