RSI Progress
What has been measured, what changed, and what the evidence tells us about recursive self-improvement.
Selected site evidence · Summary reviewed
No single benchmark or percentage captures progress toward RSI.
Each score belongs to one system, test version, and setup.
Doing research tasks well is not the same as controlling the improvement loop.
By Epoch AI
Claude Fable 5: 2%90% interval: 0–10%
GPT-5.6 Sol: 15%90% interval: 2–28%
Epoch-corrected normalized score
GRPO baseline: 0 · SDPO target: 100
Both attempts fell short on this one task. Epoch adjusted scores for run selection and method-scope judgments. One main research attempt per model; intervals describe training noise. No paper recall was detected, but contamination is not ruled out.
By Anthropic
55.8%
Research problems solved
Claude Opus 5.5 · 500 problems · one attempt each
AI can solve many of the research-debugging problems encountered by Anthropic researchers.
By Anthropic
26%
AI-led research work
Anthropic’s assessment · August 2026
AI completes most of these tasks from a high-level prompt, with people supervising.
By Weco AI
7
Accepted changes to the agent
99 proposed changes over eight days
Modified agents were reused in later rounds. Whether they became better at producing further improvements was inconclusive.
Recent publications from our reviewed library, ordered by original publication date.
Anthropic’s revised capability-index fit favors faster historical growth, but AI’s causal contribution remains uncertain. The quoted METR gains on verifiable tasks concern an earlier Mythos 5.2 snapshot stronger than the released model; they do not establish full research automation.
Evidence & limitationsMETR judges that AI assistance likely contributes substantially to faster progress on well-specified tasks at Anthropic. Causal attribution remains uncertain, important evidence is non-public, and highly autonomous research is not yet a major contributor.
Evidence & limitationsMETR questions Anthropic’s confidence that an AI-R&D acceleration warning threshold had not been crossed by July 15, 2026. This is not proof of crossing; most relevant researchers reportedly lean toward it not having been crossed.
Evidence & limitationsOpenAI reports that GPT-6 Astra completed fewer rubric items on matched coding tasks framed as safety research, even after excluding explicit refusals and verbalized withholding. The cause and any general research-slowdown rate remain unresolved.
Evidence & limitationsQin and colleagues report earlier useful solutions when research agents retain state across tasks, but mixed final hidden-test gains. The matched comparisons use frozen model weights; they do not demonstrate recursive improvement of the improvement mechanism.
Evidence & limitationsResearch, debugging, and engineering work involved in building AI.
11 measuresChanges to training, post-training, algorithms, data, and tools.
9 measuresHow independently and reliably systems complete extended work.
1 measureReported contribution in real organizational work.
3 measuresEvidence that improved systems contribute to later improvements.
Explore evidenceControls for contamination, gaming, overfitting, and transfer.
3 measures