The Self-Improving Algorithm

by

Author: Denis Avetisyan

A new agentic system demonstrates AI’s potential to autonomously drive scientific discovery and refine its own capabilities.

Comparative analyses demonstrate the evolutionary performance of ASI-Evolve and established algorithms like GEPA and OpenEvolve, with shaded regions illustrating the consistency of ASI-Evolve across multiple trials and further investigation revealing its adaptability when paired with large language models such as GPT-5-mini and Qwen3-32B.
Comparative analyses demonstrate the evolutionary performance of ASI-Evolve and established algorithms like GEPA and OpenEvolve, with shaded regions illustrating the consistency of ASI-Evolve across multiple trials and further investigation revealing its adaptability when paired with large language models such as GPT-5-mini and Qwen3-32B.

Asi-Evolve leverages AI-for-AI techniques, including neural architecture search and reinforcement learning, to automate data curation, model design, and training processes.

Despite decades of progress, advancing artificial intelligence remains a costly and iterative process largely dependent on human ingenuity. This limitation motivates research into automating the AI development pipeline, as explored in ‘ASI-Evolve: AI Accelerates AI’, which introduces an agentic framework capable of autonomously conducting AI research. Through a learn-design-experiment-analyze cycle augmented with a cognition base and dedicated analyzer, ASI-Evolve achieves breakthroughs in neural architecture design, pretraining data curation, and reinforcement learning algorithm development-surpassing state-of-the-art results by up to +18 points on challenging benchmarks. Could this closed-loop AI research paradigm represent a fundamental shift, enabling AI to truly accelerate its own evolution and unlock unforeseen capabilities across diverse scientific domains?

Navigating Complexity: The Imperative for Autonomous Discovery

The sheer volume of data now generated across scientific disciplines presents a fundamental challenge to traditional methods of inquiry. Historically, researchers have relied on hypothesis-driven experimentation, carefully designed to test specific, pre-defined questions. However, modern datasets – ranging from genomic sequences and astronomical observations to climate models and social media interactions – are not merely large, but exhibit exponential growth in complexity. This means the number of potential relationships and patterns within the data increases at a rate that quickly overwhelms the capacity for manual analysis, or even analysis guided by pre-existing theoretical frameworks. The combinatorial explosion of possibilities renders exhaustive investigation impractical, and even targeted approaches struggle to identify meaningful signals amidst the noise. Consequently, a shift towards methods capable of navigating and extracting knowledge from these complex datasets is no longer simply desirable, but essential for continued scientific progress.

The escalating volume and intricacy of modern scientific data demand a paradigm shift beyond traditional methodologies. Researchers are increasingly exploring the potential of autonomous agents – artificial intelligence systems designed to independently formulate hypotheses, design experiments to test them, and interpret the resulting data. These agents operate through iterative cycles of prediction and validation, effectively automating the scientific method. This approach doesn’t seek to replace human scientists, but rather to augment their capabilities by rapidly exploring vast datasets and identifying promising avenues of investigation that might otherwise remain hidden. The core innovation lies in imbuing these systems with the capacity for self-directed learning and discovery, potentially accelerating breakthroughs in fields ranging from materials science to drug development and offering a new means to tackle previously intractable scientific challenges.

Successfully deploying autonomous scientific discovery systems hinges on a clear understanding of the challenges embedded within the research process itself. The inherent difficulty isn’t simply about finding answers, but navigating a landscape defined by execution cost – the resources required to perform experiments or simulations – a potentially vast search space of possible hypotheses, and the often-delayed or noisy nature of feedback from those experiments. Quantifying these elements is paramount; a problem with a low execution cost and a constrained search space may be readily solvable by an autonomous agent, while one with high costs and immense complexity demands novel algorithmic approaches. Without a robust characterization of these difficulties, efforts to build self-driving scientists risk being misdirected, focusing on algorithmic improvements where the fundamental limitations lie in the nature of the scientific question itself.

The Asi-Evolve pipeline iteratively improves programs by sampling context, retrieving relevant knowledge via embedding search, executing candidate programs with timeouts, and storing evaluation results for subsequent iterations.
The Asi-Evolve pipeline iteratively improves programs by sampling context, retrieving relevant knowledge via embedding search, executing candidate programs with timeouts, and storing evaluation results for subsequent iterations.

ASI-Evolve: An Agentic Framework for Automated Scientific Inquiry

ASI-Evolve is an agentic framework engineered to automate the complete scientific process through the integrated execution of four core functions: learning, design, experimentation, and analysis. This automation is achieved by iteratively employing these functions; the system learns from existing data and prior experiments, designs new experiments based on learned insights, executes those experiments within a defined environment, and then analyzes the resulting data to refine its understanding and inform subsequent experimental designs. This closed-loop process allows ASI-Evolve to operate with minimal human intervention, effectively functioning as an autonomous scientific investigator capable of generating and evaluating hypotheses.

The ASI-Evolve framework employs a Cognition Base as its primary directive for scientific exploration; this component stores and updates hypotheses, experimental designs, and prior results to inform subsequent actions. This knowledge repository facilitates a focused search for optimal experimental parameters and allows the agent to prioritize potentially fruitful avenues of investigation. Complementing the Cognition Base is the Analyzer module, which processes data generated from experiments. The Analyzer’s function is to identify statistically significant trends, extract relevant information, and quantify the impact of experimental variables, thereby providing the Cognition Base with actionable insights to refine its understanding and guide future experimentation.

ASI-Evolve builds upon the foundation of the AlphaEvolve framework to advance the capabilities of autonomous scientific discovery. Independent evaluations demonstrate that ASI-Evolve achieves state-of-the-art performance across three key areas of AI development: algorithm design, data analysis, and experimental control. Specifically, ASI-Evolve outperforms existing methods in automated algorithm optimization, achieving higher accuracy and efficiency in data interpretation, and enabling more precise and adaptive experimental procedures. These improvements are quantitatively measured by benchmarks demonstrating significant gains in performance metrics compared to previous agentic systems and traditional scientific workflows.

ASI-Evolve, when paired with MAP-Elites, consistently outperforms UCB1 and random sampling on Qwen3-32B, as demonstrated by its more stable and higher-performing evolution curves with minimal run-to-run variability.
ASI-Evolve, when paired with MAP-Elites, consistently outperforms UCB1 and random sampling on Qwen3-32B, as demonstrated by its more stable and higher-performing evolution curves with minimal run-to-run variability.

Refining Predictive Power: Modeling Complex Interactions

Effective prediction of drug-target interactions necessitates prioritizing the most pertinent interaction patterns due to the high dimensionality and sparsity inherent in biological data. Not all potential interactions are equally significant; focusing on relevant patterns reduces noise and computational complexity. This selective approach improves model performance by allowing algorithms to concentrate on the features most indicative of true interactions, as opposed to being diluted by irrelevant data points. Identifying these patterns often involves feature selection techniques, dimensionality reduction, or the application of domain-specific knowledge to filter out unlikely or non-informative interactions, thereby enhancing both the accuracy and efficiency of prediction models.

Top-k Sparse Gating and Domain-Specific Marginalization are techniques employed to enhance signal clarity in drug-target interaction prediction models. Top-k Sparse Gating functions by focusing computational resources on the most salient features – the ‘k’ most impactful interactions – thereby reducing noise and improving efficiency. Domain-Specific Marginalization addresses the challenge of irrelevant information by systematically removing features that are not pertinent to the specific biological domain under investigation. This targeted refinement of input data results in a more focused model, leading to improved prediction accuracy by minimizing the influence of confounding factors and amplifying the signal from genuine interactions.

The Sinkhorn Attention mechanism improves drug-target interaction prediction by leveraging principles from Optimal Transport theory. This mechanism incorporates doubly-stochastic constraints, ensuring that attention weights sum to one across both the drug and target dimensions. This regularization process effectively refines the attention distribution, focusing on the most relevant interactions and mitigating noise. Benchmarking demonstrates a +6.94 Area Under the Receiver Operating Characteristic curve (AUROC) improvement in cold-start generalization scenarios, indicating enhanced predictive performance when limited data is available for new drugs or targets. The implementation utilizes the Sinkhorn algorithm to efficiently compute the doubly-stochastic attention weights, enabling scalable and accurate interaction modeling.

Validating the System: Broad-Spectrum Benchmarking and Impact

ASI-Evolve underwent rigorous testing across a suite of challenging benchmarks designed to assess its capabilities in diverse scientific fields. Evaluations included SWE-bench, a platform for evaluating software engineering solutions; GPQA, focused on complex reasoning in question answering; MMLU, a measure of massive multitask language understanding; and the historically significant Circle Packing problem, a geometric optimization puzzle. This broad spectrum of tests was intentionally chosen to demonstrate the system’s adaptability and generalizability beyond any single domain, revealing its potential for application across a wide range of scientific inquiries and problem-solving tasks. The selection of both established and contemporary benchmarks allowed for a comprehensive assessment of ASI-Evolve’s performance relative to existing state-of-the-art approaches and its ability to tackle problems requiring varied skill sets.

The versatility of ASI-Evolve is powerfully illustrated through rigorous testing across a spectrum of scientific challenges, revealing its capacity to adapt and excel in diverse domains. Performance gains stemming from improved data curation averaged 3.96 points across standard benchmarks, signifying a consistent enhancement in problem-solving ability. Notably, the system achieved over 18 points of improvement on the challenging MMLU benchmark-a measure of multi-task language understanding-demonstrating a substantial leap in complex reasoning capabilities. These results collectively suggest ASI-Evolve is not merely specialized for a narrow task, but possesses a foundational adaptability that positions it as a promising tool for broad application across multiple scientific fields and beyond.

Rigorous benchmarking reveals the potent capabilities of ASI-Evolve’s agentic framework, demonstrating not only adaptability across diverse scientific challenges but also tangible performance gains. In the complex domain of geometric problem-solving, ASI-Evolve achieved a score of 2.63597 on the ‘Circle Packing’ problem in just 17 steps, equalling the performance of current state-of-the-art methods. Furthermore, the system significantly outperformed existing approaches on challenging reasoning benchmarks; notably, ASI-Evolve registered a 12.5-point improvement over the GRPO baseline on the AMC32 benchmark, and a 0.97-point gain over DeltaNet in the refinement of model architectures, collectively illustrating the effectiveness of its core components and its potential to advance automated scientific discovery.

Ablation studies on circle packing demonstrate that the full method consistently outperforms variants lacking key components, as evidenced by the more stable and higher-performing evolution curves shown with shaded regions representing run-to-run variability.
Ablation studies on circle packing demonstrate that the full method consistently outperforms variants lacking key components, as evidenced by the more stable and higher-performing evolution curves shown with shaded regions representing run-to-run variability.

The Asi-Evolve framework embodies a holistic approach to scientific discovery, mirroring the interconnectedness of complex systems. This research demonstrates that improvements in one area – such as neural architecture search – inevitably influence others, including data curation and training methodologies. As Bertrand Russell observed, “To be happy, one must be able to change.” Similarly, Asi-Evolve’s agentic system is not static; it continuously adapts and refines its processes, illustrating how a flexible, evolving structure is crucial for sustained progress in artificial intelligence. The system’s ability to autonomously iterate and optimize highlights the principle that structure dictates behavior, and a well-designed system can drive significant advancements.

The Road Ahead

The presentation of Asi-Evolve suggests a shift, not towards artificial intelligence as a destination, but towards artificial evolution. The system’s capacity for autonomous research, while promising, merely highlights the inherent fragility of current scientific infrastructure. One does not rebuild the entire city to repair a single pothole; similarly, future work must focus on seamlessly integrating these agentic systems into existing workflows, rather than demanding wholesale replacement. The true limitation isn’t the speed of discovery, but the capacity of the established order to absorb it.

A critical unresolved question concerns the nature of the ‘cognition base’. Simply accumulating data, even through automated curation, does not guarantee wisdom. The system reflects the biases of its initial conditions – the digital equivalent of a city planned on flawed geographical surveys. Future iterations must address how to instill, or perhaps evolve, principles of robust reasoning and error correction, moving beyond mere pattern recognition.

Ultimately, the success of this paradigm will not be measured in algorithmic breakthroughs, but in its ability to yield unanticipated discoveries. If Asi-Evolve simply refines existing knowledge, it remains a sophisticated tool. True progress demands the generation of genuinely novel insights – evidence that the system has transcended its origins and begun to explore the unknown, charting a course independent of its creators.

Original article: https://arxiv.org/pdf/2603.29640.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-04 01:59