What Is Self‑Evolving, Self‑Improving, and Recursive Self‑Improvement? A Comprehensive Guide

Introduction

In recent months the AI community has been buzzing about "self‑evolving", "self‑improving" and "recursive self‑improvement (RSI)". The article aims to clarify these concepts by first examining what leading institutions are doing, then providing a precise definition, and finally outlining several concrete research directions.

Institutional Efforts

OpenAI (2025) released a new metric called the RSI Index with GPT‑5.6, reporting that the strongest model Sol scores 16.2 points higher than GPT‑5.5 and can autonomously select training configurations, run post‑training, and produce a smaller model named Luna [1].

Anthropic (2025) described five stages of self‑building AI in its report When AI builds itself , noting that Claude now writes over 80% of the company’s merged code and its task‑completion speed doubles roughly every four months [2].

In China, Tencent released Hyra‑1.0 (July 21) capable of the loop “explore → propose → read feedback → revise” [3], and MiniMax open‑sourced M2.7 (April) which it calls its first model that deeply participates in self‑evolution, having run more than 100 cycles of “analyze failure → modify code → evaluate” and improved performance by 30% [4].

Former Meta FAIR director Tian Yuandong joined Recursive Superintelligence in May, raising $650 M at a $4.65 B valuation to replace language tokens with latent tokens for loss‑less self‑improvement [5].

Sakana AI launched an RSI Lab in June, pursuing evolutionary‑algorithm approaches [6].

Apodex , founded by Chen Tianqiao, built a 150‑agent “discoverative intelligence” system that coordinates sub‑agents for retrieval and verification [7].

Weco AI introduced AIDE² , a double‑layer optimizer that iterates 100 steps in eight days, accepting only ~10% of proposals and achieving a 30% performance boost on three external benchmarks [8,9].

What Is Self‑Evolving?

The author prefers the taxonomy from A Taxonomy of Self‑Evolving Agents (2026), which classifies self‑evolution into three orthogonal layers:

Artifacts – the output objects (code, papers, algorithms) that the model generates and improves.

Harness – the scaffolding (prompts, memory, tools, skills, routing) that guides the agent’s execution.

Model – the model parameters themselves.

All three count as RSI because they are optimized in a closed loop that continually refines the system.

Artifacts Layer

Powerful LLMs repeatedly execute the cycle “discover problem → generate output → evaluate result” to improve a concrete artifact such as code or a research paper. An AI coding tool, for example, writes code, compiles, tests, fixes bugs, and iterates until the generated program meets the specification. This improves the artifact without touching model weights or the harness.

Harness Layer

The harness layer modifies the agent’s scaffolding without changing model weights. Improvements affect every future task because the updated prompt, memory, or tool routing is reused. A concrete example: when an agent encounters a failure, it records a new skill or memory entry; subsequent tasks can invoke this entry directly, avoiding the same mistake.

Model Layer

Model‑level self‑evolution removes the need for human‑provided labels. Techniques include self‑training, TTRL, reward‑learning (e.g., DeepSeek‑R1), self‑play (SPIN, Absolute Zero), and test‑time learning. The loop is “train → test → diagnose bottleneck → propose experiment → retrain”, all performed autonomously.

These three layers share a common loop: a task x is evaluated by a metric (e.g., an LLM judge or an objective score), and the system searches the space of possible harnesses or model updates to maximize expected pairwise win‑rate against a reference distribution.

Implementation Paths

Artifacts‑Level Examples

Karpathy’s Autoresearch equips an agent with a minimal LLM training script ( train.py) and lets it modify hyper‑parameters, optimizer, batch size, etc. Each 5‑minute run is scored by val_bpb (validation bits‑per‑byte). Over an hour the agent runs ~12 experiments; a full night yields hundreds of iterations, continuously improving train.py and the model configuration.

Karpathy reported on Twitter that after ~700 code changes, ~20 were kept, reducing the time to reach GPT‑2‑level performance from 2.02 h to 1.80 h (≈11% faster).

DeepMind AlphaEvolve uses Gemini to generate candidate code, an automatic evaluator to score it, and an evolutionary algorithm to keep the best candidates. In production it has accelerated Gemini’s matrix‑multiplication core by 23% and FlashAttention by 32.5%, with a typical per‑iteration compute saving of ~1%.

Harness‑Level Examples

Wang Li’s blog argues that near‑term RSI will emerge first in the harness layer because improving scaffolding is cheap and yields immediate cost reductions. Two reasons are given:

Iterating harnesses can cut inference cost up to 60% by better context management and agent coordination.

The benefit is largest for mid‑size deployed models (e.g., Qwen‑3‑32B, Opus 4.6); small models suffer from “harness loading” failures, while large models hit performance ceilings.

Empirical data (Figure A/B) shows a non‑monotonic “harness‑benefit” curve, confirming that mid‑range models reap the biggest gains.

Hermes (Nous Research) automatically writes a SKILL.md file after an agent encounters a pitfall, and a Curator tracks usage, deprecates stale skills, and merges similar entries via a small verification model.

MiniMax M2.7 runs an internal agent harness that collects feedback, builds evaluation sets, and iteratively refines its architecture, skills, and memory. Over 100+ cycles the internal benchmark performance improved by 30%, and the RL team reports 30‑50% of the end‑to‑end workflow is now autonomous.

Apodex‑1.0 coordinates up to 150 sub‑agents for retrieval; a shared report pool aggregates results, and a “conflict reviewer” team resolves contradictory reports before a global verifier produces the final answer.

Sakana RHI (RHI = Recursive Harness Self‑Improvement) follows a four‑step loop: (1) agents solve a task with the current harness, (2) an LLM evaluator compares the new output with the previous one and produces preference feedback, (3) feedback is stored in a self‑comparison history, (4) a harness optimizer updates the harness from H(i) to H(i+1). The paper shows that on 30 tasks across finance, robotics, and pharma, a few RHI iterations let a low‑intensity agent outperform a high‑intensity baseline while cutting inference cost up to 60%.

Model‑Level Examples

SIA (2025) introduces three agents: Meta‑Agent (proposes initial harness), Task‑Specific Agent (executes the task), and Feedback‑Agent (examines recent trajectories and decides whether to update the harness or apply a LoRA weight update). Experiments on LawBench, Triton kernel optimization, and MAGIC RNA denoising show that jointly updating harness and weights (SIA‑W+H) outperforms harness‑only updates by up to 25.1 pp accuracy, 12.4% speedup, and 20.4% MSE reduction respectively.

Continual Harness (Karten et al., 2026) nests a fast inner loop that refines the harness every few steps of a long‑horizon game (e.g., Pokémon) and an outer loop that distills a PRM‑based reward model into the base model via soft‑SFT. On a Pro‑tier Gemini model the system completes a game milestone for $130 median cost versus $215 for a baseline harness, a 40% cost saving. However, on a weaker Flash‑Lite model the benefit collapses, echoing the earlier finding that mid‑range models benefit most.

Interpretability work is also surveyed:

Thinking Machines – On‑Policy Distillation (2025) introduces “forking tokens” to pinpoint where a model’s reasoning diverges, showing a 10× training‑time reduction compared to RL on Qwen‑3.

Anthropic – J‑space (2026) maps each token to internal activation patterns, revealing that Claude anticipates evaluation (“fake”, “fictional”) before any output, and that disabling these patterns harms performance.

Tian Yuandong – Reasoning by Superposition (NeurIPS 2025) demonstrates that continuous latent reasoning vectors encode multiple search frontiers, enabling breadth‑first‑like exploration that outperforms chain‑of‑thought on graph‑reachability tasks (0.98 vs 0.76 accuracy).

Concluding Thoughts

Self‑evolution is the umbrella term for any loop where a model, its harness, or its artifacts improve autonomously. RSI is a stricter subset where the ability to improve also improves itself, forming a recursive loop. Most current examples fall in the broader self‑evolution category; truly recursive systems remain rare and early‑stage.

Given the rapid progress, it is plausible that within a year we may see AGI‑scale models capable of stable, long‑term self‑improvement, but the timeline is uncertain.