For people relying on an artificial intelligence agent to write code, use software tools or carry out a complex task, an important question is whether the system can learn from its failures without waiting for developers to retrain it. Recent academic work describes a route towards that goal through recursive self-improvement, or RSI: an autonomous, closed-loop process in which an AI identifies its limitations, creates and tests changes, then uses the resulting capabilities to improve how it improves.
The research sets out a five-stage roadmap. It begins with improvement-execution autonomy, followed by improvement-strategy autonomy and experience-acquisition autonomy. The later stages are environment-adaptation autonomy and recursive meta-improvement. The framework is intended to cover work in scientific discovery, embodied intelligence and software engineering.
Most systems remain in the first two stages. At L1, an AI carries out a human-defined improvement process. Results that meet the required standard are retained and used in later tasks or further rounds of improvement. At L2, the system reviews evidence from previous attempts, identifies a weakness and chooses which proposed intervention to try next.
Those limits matter because the objectives and underlying improvement methods are still set by people. At L4, the framework envisages an AI drawing on feedback from live use to make lasting changes, including rules, skills or revisions to its operating harness. Its goals, access limits, protected evaluation and final authority over release would remain outside the system. L5 would allow a system to revise the mechanism through which it improves itself. The research does not establish that such systems have been deployed or demonstrated at scale.
Several teams have built early examples of the approach. RSIAgent is described as a training-free, multi-agent framework that uses autonomous memory construction and a broad-then-deep approach to exploration. Applied to Kimi-K3 and GLM-5.3, it recorded 78.98% Partial Score on OSWorld 2.0, compared with 72.60% for GPT-6 Astra. It also scored 84.82% on Agents' Last Exam (Near-term), against 82.26% for GPT-6 Astra.
Other projects focus on narrower parts of the cycle. Self-Harness uses records of an agent's actions to suggest and test changes to the harness that supports the agent, working under a fixed benchmark and promotion rule. SIMA 2 assesses present behaviour and creates later practice tasks aimed at identified weaknesses.
NeoHorse-1, a family of agent-native models, explores the approach through agentic post-training. It treats use in deployment as the beginning of further training. Its Routing Harness records each real request and the full trail of interactions made by the agent, feeding an evaluate, select and update loop.
Across eleven benchmarks covering harness-based agents, tool use, coding and instruction following, post-training on NeoHorse-1 lifted the macro-average from 58.94 to 64.87 at 4B parameters. Across the same eleven benchmarks, it raised the macro-average from 65.60 to 69.04 at 9B parameters. The results show measurable gains from a feedback loop, though they do not resolve whether the same method can operate equally well across every field.
The research also points to uneven capability across areas where AI systems interact with tools and tasks. The Headroom-Closed Index is proposed as a way to measure current performance and the remaining room for improvement. By 2026, graduate science reached Headroom-Closed Index 85.8. Software engineering reached 52.6 and tool agents 39.9. There was a 45.9-point gap between the highest and lowest interactive-capability domains by 2026.
That variation leaves open major practical questions. Researchers have yet to show how quickly systems at L1 and L2 could progress to greater autonomy, or whether computing demands and safety constraints will slow that progress. Earlier evidence suggested that rapid recursive improvement could produce superhuman capabilities, while the same work identified limits in compute and safety that could hold development back.
For now, the projects amount to working pieces of a larger design rather than a fully autonomous self-improving system. They show that AI agents can preserve lessons from experience, select some changes and gain on defined benchmarks. The harder task is proving that such systems can safely acquire experience, adapt in live environments and revise their own improvement process while remaining within externally set boundaries.
Comments (0)