The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

TL;DR

Position survey mapping the recursive-self-improvement (RSI) space by introducing two organizing tools: a Headroom-Closed Index (HCI) that quantifies how much of the human–LLM performance gap current systems have actually closed, and a five-stage autonomy roadmap — improvement-execution → improvement-strategy → experience-acquisition → environment-adaptation → recursive meta-improvement. The staging is designed to make RSI claims falsifiable: for any paper claiming a “self-improving” agent, ask which stage it actually automates. The survey then compares RSI requirements across three application scenarios — scientific discovery, embodied intelligence, and software engineering — and connects academic research threads to industry practice, identifying key gaps to reach “genuine” (stage-5) RSI. [§Abstract]

Key claims

• RSI can be decomposed into five progressively more autonomous stages: improvement-execution autonomy (agent executes improvements someone else designed), improvement-strategy autonomy (agent chooses its own strategy), experience-acquisition autonomy (agent collects its own experience), environment-adaptation autonomy (agent adapts to new environments), and recursive meta-improvement (agent improves the process of improvement itself) [§Abstract].
• The Headroom-Closed Index (HCI) is proposed as the common metric for cross-system comparison, revealing where current LLMs fall short across scenarios [§Abstract].
• RSI development speeds and requirements differ substantially across scientific discovery, embodied intelligence, and software engineering — a single recipe is unlikely to serve all three [§Abstract].
• Genuine RSI (stage 5) requires the loop to close through both capability improvement and improvement-process improvement; no filed system yet demonstrably operates at this stage [§Abstract, framing].

Method

Position/survey paper. The two organizing tools are the HCI metric and the five-stage autonomy roadmap; the paper uses them to classify existing RSI attempts, draw contrasts across scientific-discovery / embodied-agent / software-engineering scenarios, and highlight open challenges. Only the abstract is captured here — the full taxonomy of surveyed systems and the per-scenario requirements analysis are in the paper body.

Results

Not an empirical paper — no headline benchmark numbers. The Headroom-Closed Index is proposed as the measurement tool; per-system HCI values across the surveyed scenarios are in the paper body but not extracted here.

Why it’s interesting

This survey formalizes what AI-for-AI Research has been informally tracking as “granularity” — the five-stage roadmap is a sharper vocabulary than the concept page currently uses to distinguish (say) Karpathy autoresearch (stage 1: agent executes a fixed-scope hyperparameter search) from AIDE² (AIDE²: First Evidence of Recursive Self-Improvement) which explicitly tests stage-5 recursion via ignition and rejects it as unsupported. It also directly complements History May Repeat Itself: RSI Seen from a Previous AI Era (Yuandong Tian’s ICLR 2026 RSI-workshop keynote): where Tian frames RSI as model + harness and enumerates four ingredients (data, memory, search, cost) each system must solve, this survey provides the orthogonal autonomy-stage axis and the HCI comparison metric. Third, the paper’s per-scenario split (scientific discovery vs embodied vs SWE) matches the wiki’s own observation that the filed AI-for-AI-research systems already cluster by scenario — Kosmos (Kosmos: An AI Scientist for Autonomous Discovery) and Hyra (Hyra-results — Hunyuan Research Agent (results bundle)) target scientific discovery; ENPIRE (ENPIRE: Agentic Robot Policy Self-Improvement in the Real World) and GRID Auto-Engineering (Introducing Auto-Engineering for Robotics) target embodied; ASI-ARCH (AlphaGo Moment for Model Architecture Discovery) and Kimi K3 kernel autoresearch (Kimi K3 — Open Frontier Intelligence (2.8T MoE with KDA + AttnRes)) target SWE-adjacent architecture/kernel work. Providing a unified metric would let the wiki compare across these scenarios rather than only within-scenario.

See also

• AI-for-AI Research — the concept page this survey is most directly a map of; the five-stage roadmap is a candidate vocabulary for tightening the existing “granularity” framing
• History May Repeat Itself: RSI Seen from a Previous AI Era — Yuandong Tian’s complementary RSI-workshop keynote; that talk provides the ingredient axis (data/memory/search/cost), this survey provides the autonomy-stage axis
• AREX: Towards a Recursively Self-Improving Agent for Deep Research — AREX brands its within-task refinement “RSI”; the five-stage roadmap here would classify it as stage 1–2 (execution + strategy), not the harness-improves-harness sense
• AIDE²: First Evidence of Recursive Self-Improvement — AIDE² explicitly tests the highest autonomy stage (agent improves the outer-loop agent) and reports the ignition experiment as inconclusive, one of the few filed empirical touches on stage 5
• Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation — an example of self-generated experience acquisition (stage 3) in the pretraining phase