Framework · Supplied arXiv:2609.11873v1 · 2026-09-10
B0–L5 autonomy framework
Which responsibilities in the improvement loop are controlled by AI, and which changes are inherited?
Non-RSI reference level
In-task AI improvement
Refine output
Changes refine an output during the current task or session. Future independent tasks inherit no accepted system update.
Criteria
- No accepted persistent system change
- Non-RSI reference level
RSI taxonomy level
Improvement execution autonomy
Execute improvements
Humans prescribe the target, update procedure and acceptance criteria. AI executes the procedure and accepted changes persist into later tasks or rounds.
Criteria
- Human-defined what, how and success
- Persistent accepted changes
RSI taxonomy level
Improvement strategy autonomy
Choose how to improve
AI diagnoses weaknesses and chooses interventions and experiments. Humans continue to set objectives, task boundaries and evaluation criteria.
Criteria
- AI chooses improvement strategy
- Objectives and acceptance remain external
RSI taxonomy level
Experience-acquisition autonomy
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.
Criteria
- Learner-conditioned experience acquisition
- Experience retained for later improvement
RSI taxonomy level
Environment adaptation autonomy
Adapt from deployment
Ongoing deployment or environmental interaction determines persistent changes in memory, skills, code, harnesses or parameters reused on later operational tasks.
Criteria
- Persistent adaptation from operational feedback
- High-level goals, access, evaluation and release may remain externally governed
RSI taxonomy level
Recursive inheritance autonomy
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.
Criteria
- Identify the revised mechanism
- Verify creation, retention, inheritance and later invocation
- Evaluate effectiveness separately from structural reuse
Methodology and source authorship
The B0–L5 taxonomy is Duan et al.’s analytical lens, not a settled universal definition. Other taxonomies exist, including Weco’s outcome-focused four-level framework, which is not numerically interchangeable. Assignments depend on system boundary and inherited mechanism. Greater autonomy does not imply greater capability, safety, reliability or efficiency. Bounded mechanism demonstrations do not establish unconstrained autonomy. Capability results and autonomy assessments remain separate. Structural L5 requires a changed improvement mechanism to be created, retained, inherited and invoked later. Effective L5 additionally requires better subsequent improvements under comparable budgets and independent assessment. Repeated gains alone do not establish acceleration.
Extraction review: agent checked · Codex