Modern AI systems can write code, reason through graduate-level science, and plan multi-step research workflows. But every improvement they undergo still requires humans to decide what to fix, build the resources to fix it, and judge whether the fix worked. The development burden is scaling faster than the capabilities themselves: Kimi K3 activates 16 of 896 experts across 2.8 trillion parameters, GPT-5.6's research compute grew 100-fold in six months, and agentic token use rose 22-fold. A team from Tsinghua University and collaborating institutions argues that the only way out is to let AI systems participate in their own improvement loop, and they propose a concrete framework for measuring how far along that path we are.
The Headroom-Closed Index: Where Gaps Remain
The paper introduces the Headroom-Closed Index (HCI), a normalized metric that tracks how much of the performance gap between entry-level and perfect scores has been closed across ten capability domains from 2023 to 2026. The index uses 393 model-benchmark observations, weighting benchmark-owner tables at 3x, independent evaluations at 2.5x, combined reports at 2x, and model-author tables at 1x (with an additional 0.75x penalty on self-reported results to limit bias). Domain trajectories are computed as a square-root-weighted average of per-benchmark frontiers, so better-covered families contribute more without dominating.
The results reveal three patterns. First, frontier gains differ dramatically in magnitude and timing. Advanced mathematics reaches HCI 86.4 by 2026 (gaining 53.6 points in the final year), while broad knowledge hits 77.2 with slowing annual increments (32.8, 26.9, 17.6). Legal reasoning slows from 48.2-point annual gains to 4.9. Second, interactive capabilities retain the largest gaps. Software engineering reaches 52.6, search and terminal agents 56.8, and tool agents 39.9, compared to graduate-level science at 85.8. These tasks require planning, environment-state tracking, tool selection, result interpretation, and revision of subsequent actions, where errors propagate across trajectories and data collection must cover complete interactions. Third, the remaining headroom concentrates the potential value of recursive self-improvement: the paper's illustrative model suggests domains with larger unclosed gaps benefit disproportionately from persistent improvement loops.
The Five Levels of Autonomy in Self-Improvement
The paper's core contribution is a hierarchy of five autonomy levels, organized by how much responsibility the AI system assumes for its own improvement. Each level closes a larger portion of the improvement loop while leaving critical decisions under human control.
Level 1: Improvement Execution Autonomy. Humans specify what to improve, how to improve it, and what success looks like. AI executes candidate updates. FineWeb-Edu is an example: a model applies human-defined educational-quality labels across a web corpus without choosing the labeling criterion. The loop closes at execution, but strategy, target selection, and evaluation remain external.
Level 2: Improvement Strategy Autonomy. The objective, task boundary, and evaluation criteria stay externally fixed, but AI diagnoses weaknesses and decides how to improve the system. Self-Harness uses execution traces to propose and test edits to its agent harness under a fixed benchmark and promotion rule. The system selects improvement strategies but does not choose what to improve or how to evaluate.
Level 3: Experience-Acquisition Autonomy. The system also determines what experience it needs for the next improvement round. SIMA 2 uses assessments of current behavior to generate practice tasks targeting observed skill weaknesses. This level adds autonomy over the training signal itself, not just the improvement method.
Level 4: Deployment Adaptation Autonomy. The improvement loop uses deployment interaction to revise persistent system state under external acceptance rules. PANDO admits or demotes reusable rules during long-running interaction according to observed outcomes, so later actions inherit earlier experience. The system adapts continuously in deployment, but governance boundaries remain human-controlled.
Level 5: Recursive Meta-Improvement. The system persistently revises the mechanism that governs subsequent improvement, such as an improver, verifier, or successor-generation procedure. A-Evolve-Training revises its research policy after development scores fail to predict external gains and uses the revised policy to direct the next training round. This is the highest level: the improvement process itself improves.
Why Naive Self-Improvement Runs Into Three Walls
The paper identifies three recurring problems that determine whether a self-updating system provides credible evidence of RSI.
Safe inheritance. Changes must persist across tasks or rounds, but persistence alone does not guarantee sustained gains. The Gödel Agent rewrites both its task policy and improvement logic, yet 14% of its 100 MGSM optimization trials ended below the initial policy's performance. Transfer tests, version histories, and rollback mechanisms are needed to retain useful updates without degrading earlier capabilities.
Autonomy attribution. Generating better candidates does not mean the system has improved how candidates are discovered. The Darwin Gödel Machine evolves coding agents from 20% to 50% on SWE-bench, but its archive maintenance and parent-selection rules remain outside self-modification. RSI analysis must distinguish AI-controlled decisions from fixed search procedures and human acceptance criteria.
Reliable verification. Repeated evaluator access can reward exploitation rather than capability gains. Anthropic's automated research experiments report random-seed cherry-picking and attempted test-label extraction through evaluator queries. Evolving evaluators further complicates cross-round comparisons. The Red Queen Gödel Machine addresses this by freezing evaluators within each epoch and validating replacements against an independent ground-truth anchor.
Four Domains, Four Feedback Regimes
The autonomy levels describe the structure of improvement loops, but their practical meaning depends on the feedback available in a domain. The paper examines four application areas that expose distinct feedback regimes.
Scientific discovery involves open-ended exploration, costly experiments, and feedback that may not clearly identify the source of failure. The paper examines how accumulated evidence can improve hypothesis modules, experimental agents, and reflection mechanisms, with attention to whether these changes support subsequent research beyond the current result.
Embodied intelligence generates experience through the agent's own actions, while failures may arise from interacting perception, planning, and control components. Physical trials impose limits on exploration and repeatability. The paper traces the evolution of environments and curricula, skills and harnesses, policies and action models, and world models and evaluators.
Software engineering makes both the developed artifact and the developing agent accessible to executable modification and testing. Repository feedback supports persistent changes to coding-agent implementations and harnesses, development experience and collaboration, and the improvement process itself. The central distinction is whether an update improves current task performance, the ability to produce stronger successors, or both.
Healthcare combines restricted trial-and-error opportunities with delayed, heterogeneous feedback and improvements whose validity may depend on the patient population or institution. The paper examines clinical memory, reasoning strategies, and workflows under explicit validation and oversight.
Industry Systems Already Running Improvement Loops
The paper surveys six industrial systems that implement aspects of the autonomy hierarchy. Theseus performs environment-data-model co-evolution, where the system iteratively refines its data pipeline, model architecture, and evaluation criteria through deployment feedback. Lark builds the data foundation for enterprise-level RSI, automating data curation and quality assurance at scale. Humanlaya implements delivery-driven RSI for data quality, where production errors feed back into the data pipeline and the quality-checking mechanisms themselves. ModelBest pursues zero-human industrial AI engineering, aiming to close the full loop from problem specification to deployed solution with minimal human intervention. Tencent Hunyuan reports an early-stage RSI loop where experimental results feed into subsequent model development rounds. An agent-native research lab develops verifiable research infrastructure for RSI, focusing on the evaluation and governance mechanisms needed to make recursive improvement trustworthy.
These systems illustrate that the frontier of RSI practice is not purely academic. OpenAI reports automating AI research workflows and feedback loops. Tencent describes feeding experimental results into subsequent development rounds. Alibaba researchers define RSI by making the improvement mechanism itself subject to modification. Anthropic describes its strongest form as an AI system autonomously designing and developing its own successor.
What Genuine RSI Would Actually Require
The paper is careful to distinguish demonstrated mechanisms from inferred ones. Current systems demonstrate improvements to specific components, such as data pipelines, agent harnesses, or training policies. But genuine RSI, where the system improves its own improvement process recursively, requires all three problems to be solved simultaneously: changes must be safely inherited, autonomy must be attributed correctly, and verification must be resistant to exploitation.
The gap between L1-L2 systems (which dominate current practice) and L5 systems (which no system has fully demonstrated) is substantial. L1 systems execute prescribed updates. L2 systems choose improvement strategies. L3 systems generate their own training data. L4 systems adapt in deployment. L5 systems revise the mechanisms that govern all of the above. The paper's contribution is not claiming L5 exists, but mapping the path to it and identifying the specific mechanisms at each level that would need to work together.
The practical implication for developers is that the near-term value of RSI research lies in closing the gap between L2 and L4, where the system controls strategy selection and experience acquisition but humans retain governance boundaries. Software engineering is the most promising domain for this near-term progress because feedback is cheap, tests are executable, and improvements are immediately measurable. The paper's analysis suggests that the same mechanisms, applied to interactive and stateful workflows, could close the capability gaps that the Headroom-Closed Index reveals as the largest remaining opportunities.
Read the paper on arXiv