Benchmark
CoBench
2.1
Diagnose historical internal R&D failures from infrastructure snapshots.
53.2%
Claude Opus 5
An evidence tracker
Tracking AI’s ability to research, improve, and build AI systems.
What has been measured, what changed, and what the evidence tells us about recursive self-improvement.
- Measures
- 14
- Public results
- 48
- Papers & reports
- 16
Where are we? ·
How much of its own improvement does AI control?
Duan et al. (Sep. 2026) assessment
L2
Most domains have reached L2: choose how to improve
3 of 4 domains. Lower and intermediate levels have broad evidence; L3–L4 are more domain-dependent.
- What L2 means
- AI diagnoses weaknesses and chooses interventions and experiments. Humans continue to set objectives, task boundaries and evaluation criteria.
- Not yet common: L3, choose what to learn
- AI chooses or generates subsequent learning experience based on the evolving learner’s state. What it learns from changes as the learner changes.
Autonomy levelB0Refine outputL1Execute improvementsL2Choose how to improveL3Choose what to learnL4Adapt from deploymentL5Improve the improvement mechanism
ScienceBelow frontierBelow frontierEvidencedEmergingNo evidenceNo evidence
Embodied intelligenceBelow frontierBelow frontierEvidencedEvidencedEmergingNo evidence
Software engineeringBelow frontierBelow frontierEvidencedEmergingNo evidenceEmerging
HealthcareBelow frontierBelow frontierEvidencedEmergingEmergingNo evidence
What each level means
- B0 Refine output
- Changes refine an output during the current task or session. Future independent tasks inherit no accepted system update.
- L1 Execute improvements
- Humans prescribe the target, update procedure and acceptance criteria. AI executes the procedure and accepted changes persist into later tasks or rounds.
- L2 Choose how to improve
- AI diagnoses weaknesses and chooses interventions and experiments. Humans continue to set objectives, task boundaries and evaluation criteria.
- L3 Choose what to learn
- AI chooses or generates subsequent learning experience based on the evolving learner’s state. What it learns from changes as the learner changes.
- L4 Adapt from deployment
- Ongoing deployment or environmental interaction determines persistent changes in memory, skills, code, harnesses or parameters reused on later operational tasks.
- L5 Improve the improvement mechanism
- The procedure governing future improvement is itself revised, retained and invoked in subsequent improvement rounds. It may be an improver, search or research policy, evaluator or successor generator.
Evidenced frontierEmergingBelow frontierNot yet
The top rung, L5: AI improves how it improves
Structural L5Bounded prototypes and emerging research or industrial systems.
Effective L5Some bounded meta-improvement and transfer; reliable multi-generation accumulation remains unresolved.
Recursive accelerationNot established by this survey’s reviewed evidence under comparable resources.
Read the evidence in context
How we interpret results →- No single score
No single benchmark or percentage captures progress toward RSI.
- Results keep their context
Each score belongs to one system, test version, and setup.
- Capability ≠ autonomy
Doing research tasks well is not the same as controlling the improvement loop.
Selected evidence
Browse all measures85%Estimated requirement
Benchmark
Internal Research Debugging Evaluation
Astra card snapshot, September 2026
Resolve real bugs in internal research experiments.
78.05%
GPT-6 Astra
Benchmark
AI4AI-Bench
Paper v1
Rewrite training algorithms in frozen research repositories.
0.250 normalized score
Claude Opus 5 / Claude Code effort aggregate
Evidence categories
AI-R&D capability
Research, debugging, and engineering work involved in building AI.
6 measuresAI-system improvement
Changes to training, post-training, algorithms, data, and tools.
5 measuresResearch autonomy
How independently and reliably systems complete extended work.
1 measureObserved R&D automation
Reported contribution in real organizational work.
3 measuresRecursive improvement evidence
Evidence that improved systems contribute to later improvements.
Explore evidenceEvaluation integrity
Controls for contamination, gaming, overfitting, and transfer.
Explore evidence