ASI-Evolve: Autonomous AI Evolution

• ASI-Evolve is a framework that integrates evolutionary computation, cognitive prior injection, and automated experimentation to drive autonomous AI self-improvement.
• It implements a cyclic process of learning, designing, experimenting, and analyzing to iteratively enhance neural architectures, data curation, and algorithmic discovery.
• Empirical results across neural architecture search, data curation, and RL algorithm discovery validate its approach, marking significant steps towards post-AGI superintelligence.

ASI-Evolve refers to a class of agentic, closed-loop frameworks and methodologies by which artificial intelligence systems autonomously accelerate the research, design, and improvement of AI itself. These systems integrate evolutionary principles, cognitive prior injection, and automated experimentation with analytic feedback mechanisms, operating across neural architecture, data curation, and algorithmic discovery. ASI-Evolve addresses the challenge of AI-driven progress in open-ended domains, supporting steps toward post-AGI superintelligence by evolving not only models but also aspects of the AI research loop itself (Xu et al., 31 Mar 2026, Genewein et al., 10 Jun 2026, Pedersen et al., 18 Dec 2025).

1. Conceptual Foundations and Definitions

The term ASI-Evolve is rooted in efforts to characterize and operationalize the transition from artificial general intelligence (AGI) to artificial superintelligence (ASI). AGI is formally defined via the Legg–Hutter universal intelligence metric as an agent $\pi$ for which $U(\pi) \geq \theta_{\text{AGI}}$, where

$U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$

and $K(\mu)$ is the Kolmogorov complexity of environment $\mu$ and $V^\pi_\mu$ the expected discounted reward for agent $\pi$ in environment $\mu$ (Genewein et al., 10 Jun 2026). ASI, in turn, is a system that “exceeds the capabilities of large, well-coordinated human-expert collectives across virtually all domains,” with $U(\pi_{\text{ASI}}) \gg U(\pi_{\text{AGI}})$ and $\pi_{\text{ASI}}$ outperforming human collectives $U(\pi) \geq \theta_{\text{AGI}}$0 on tasks of human interest.

ASI-Evolve specifically names research and engineering programs aiming to realize open-ended improvement in AI agents and frameworks by recursively automating the learn–design–experiment–analyze loop, thus accelerating their own progress across foundational stages of AI development (Xu et al., 31 Mar 2026). The conceptual underpinnings incorporate formal models of recursive self-improvement, evolutionary computation, and multi-agent intelligence scaling.

2. ASI-Evolve Framework: Closed-Loop Evolutionary AI-for-AI

ASI-Evolve implements a closed-loop, evolutionary pipeline for AI-for-AI research. The workflow operates iteratively over a space of candidate programs or scientific artifacts $U(\pi) \geq \theta_{\text{AGI}}$1, enhanced by two principal innovations: a cognition base containing aggregated human priors, and an analyzer module that structures experimental outcomes into actionable insights.

Framework Cycle

At each round $U(\pi) \geq \theta_{\text{AGI}}$2:

• Learn: Context nodes $U(\pi) \geq \theta_{\text{AGI}}$3 (motivation, code, results, etc.) are sampled from a persistent database $U(\pi) \geq \theta_{\text{AGI}}$4 of prior experiments; relevant human priors $U(\pi) \geq \theta_{\text{AGI}}$5 are retrieved from a cognition store $U(\pi) \geq \theta_{\text{AGI}}$6 using embedding-based semantic search on $U(\pi) \geq \theta_{\text{AGI}}$7.
• Design (Researcher): An LLM, guided by $U(\pi) \geq \theta_{\text{AGI}}$8 and $U(\pi) \geq \theta_{\text{AGI}}$9, generates a new candidate program $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$0 and a natural-language motivation. Optionally, edits (diffs) of existing programs can be proposed.
• Experiment (Engineer): $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$1 is executed under a task-specific evaluation script, returning primary fitness $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$2 and auxiliary metrics; early rejection is supported via quick tests and LLM-based judging.
• Analyze (Analyzer): Full logs and results are ingested. Automated causal and statistical analyses distill a decision-oriented report encompassing motivation, failure modes, and recommendations.

This loop enables structured accumulation of knowledge, reuse of experimental experience, and injection of domain-specific human expertise—anchoring AI-driven exploration in established priors while promoting autonomous discovery (Xu et al., 31 Mar 2026).

3. Mathematical Formulation and Search Dynamics

ASI-Evolve formalizes search within a program space $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$3 by integrating context-aware candidate sampling and composite multi-objective fitness. The database $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$4 preserves all experimental records, and the cognition base $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$5 encodes domain priors.

• Sampling: $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$6; $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$7 selects priors according to embedding similarity with current context.
• Candidate Generation: $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$8.
• Fitness Objectives: Each $U(\pi) = \sum_{\mu \in E} 2^{-K(\mu)}\, V^\pi_\mu$9 is scored by $K(\mu)$0 (task-specific benchmark), $K(\mu)$1 (LLM-judged novelty/quality). The composite fitness function is

$K(\mu)$2

with $K(\mu)$3 tuned per domain.

Task-specific fitness functions are layered for:

• Neural Architecture Search (NAS): Perplexity, benchmark scores, and judged code quality are combined, with explorations scaling from 20M to 1.3B parameter models across up to 16 language benchmarks.
• Data Curation: Quality criteria include coverage of known issues, clarity, and consistency; composite metrics drive selection of text-filtering strategies.
• RL Algorithm Discovery: Fitness scored as improvements over strong baselines (e.g., GRPO) on leading RL and mathematical problem suites.

The analyzer uses regression, mutual information, and qualitative diagnosis via LLM to generate concise, novel insight for subsequent iterations (Xu et al., 31 Mar 2026).

4. Key System Components: Cognition Base and Analyzer

Cognition Base

The cognition base encodes structured human priors (heuristics, design rules, known pitfalls) obtained from extensive literature review (e.g., $K(\mu)$4150 papers on linear attention, 100 on DTI modeling). Each prior $K(\mu)$5 is embedded and indexed for fast retrieval using contextual queries $K(\mu)$6.

Retrieval occurs by embedding $K(\mu)$7's analysis and motivation, selecting top-$K(\mu)$8 similar priors. These priors are injected verbatim into the prompt for the Researcher agent, thus grounding LLM-driven hypothesis generation and accelerating exploration and cross-domain transfer (Xu et al., 31 Mar 2026).

Analyzer

The analyzer module transforms raw, high-dimensional experimental data into structured knowledge. Processing includes:

• Statistical summaries (means, variances, early-stop curves)
• Causal analysis (feature–performance correlation)
• Qualitative diagnosis (LLM-prompted explanations of root causes) Structured outputs highlight drivers of performance, failure modes, and actionable recommendations, driving knowledge accumulation and exploration–exploitation balance.

5. Empirical Achievements and Research Domains

ASI-Evolve demonstrates empirical benefits across core AI research domains:

Research Domain	Experiment Scale	SOTA Gains
Neural architecture	1,350 candidates, up to 1.3B params, 16 benchmarks	+0.97 perplexity over DeltaNet (3× Mamba2 improvement)
Data curation	504B cleaned tokens	Avg +3.96 on benchmarks; MMLU +18.64 points
RL algorithm discovery	300 rounds, 14B models	AMC32: +12.5, AIME24: +11.67, OlympiadBench: +5.04

Innovations include PathGateFusionNet (hierarchical gating), ContentSharpRouter (learnable temperature content routing), and Pairwise Asymmetric Optimization (RL advantage, gradient dropout, and budget constraints) (Xu et al., 31 Mar 2026).

Transfer experiments show effectiveness in mathematics (e.g., circle packing: SOTA in 17 rounds), drug–target interaction prediction (AUROC increases up to +6.94 for unseen drugs), and demonstrate principles generalizing beyond the AI stack.

6. Theoretical Context: Pathways Toward ASI and Open-Endedness

ASI-Evolve is situated against four technical pathways from AGI to ASI (Genewein et al., 10 Jun 2026):

1. Scaling AGI: Empirical scaling laws predict power-law improvements in benchmark scores with compute and model size ($K(\mu)$9), raising questions concerning smooth continuation to ASI-level capability.
2. AI Paradigm Shifts: Involve test-time scaling, adaptive retrieval, continual learning, and potentially radical shifts (neuromorphic architectures, AutoML-Zero).
3. Recursive Self-Improvement: Modeled as $\mu$0; rates of improvement governed by recursive automation of algorithm design, data curation, and hardware search.
4. Multi-Agent Collectives: Conjectured group-intelligence scaling laws ($\mu$1), and organizational strategies (centralized versus decentralized), with open questions on coordination efficiency.

ASI-Evolve operationalizes recursive self-improvement by enabling AI to automate the research loop and distill knowledge, contributing to open-endedness and the prospect of superhuman capability without external optimization (Xu et al., 31 Mar 2026, Pedersen et al., 18 Dec 2025).

7. Challenges, Bottlenecks, and Research Outlook

Six principal frictions shape the ASI-Evolve trajectory (Genewein et al., 10 Jun 2026):

• Data wall (insufficient human-generated data)—mitigated by simulation, self-play, and synthetic data.
• Resource/economic limits (compute, energy scaling)—offset by efficiency optimizations.
• Neural paradigm ceiling (architectural limitations)—addressed by paradigm evolution.
• Research difficulty—countered by automation of research itself.
• Abstraction barrier (limitations due to human abstractions)—requiring unsupervised or embodied discovery.
• Deliberate slow-down/governance—balanced by competitive pressures and regulatory responses.

Open research directions include modeling capability–resource coupling, designing ASI-relevant benchmarks, formalizing recursive improvement bounds, advancing multi-agent theory, and addressing governance and safety at scale.

The mechanisms instantiated in ASI-Evolve are positioned as a domain-general blueprint for autonomous, open-ended, and recursive AI development, providing a substrate for progress along the AGI–ASI continuum, with implications for theoretical AI, benchmark design, and societal adaptation strategies (Xu et al., 31 Mar 2026, Genewein et al., 10 Jun 2026, Pedersen et al., 18 Dec 2025).