MedRSI: Recursive Self-Improvement for Medical Agents via Clinically Aligned Self-Evolution
Junde Wu, Jiayuan Zhu, Minghao Hu, Fenglin Liu, Jiazhen Pan
cs.AI
2026-09-22
MedRSI invents clinical tools from diagnostic failures and registers them after three cohort trials. Glaucoma balanced accuracy rises from 56.2% to 94.1%, beating MedAgent-Pro.
Medical agents today bolt a general reasoning model onto a fixed kit of specialist tools: segment the disc, compute cup-to-disc ratio, estimate ejection fraction. Each tool is conceived by a clinician, trained by an engineer, and wired in by a developer. After deployment the agent logs thousands of failures and has no way to turn them into new clinical operations.
Recursive self-improvement hands that loop to the agent. In medicine the naive version fails in two places. Error frequency is a poor proxy for patient harm: missing advanced glaucoma is not the same mistake as calling a borderline eye a suspect. And a tool that looks good on the cases that motivated it can poison later diagnostic traces if it is admitted on the same day.
MedRSI, from Oxford and Stanford, keeps the reasoning model and the diagnostic workflow frozen. Self-improvement only grows a stable tool registry. Each round diagnoses a patient batch, ranks failures by clinical consequence, clusters them into a missing capability, builds candidates by code generation, tool composition, or model training, then trials them on later cohorts before registration.
Clinical-cost-aware prioritization scores each error from 0 to 4 on a frozen, task-specific rubric. Priority is the error indicator times (1+λ times normalized cost), with λ=2 by default. A discharged advanced case outranks a minor borderline disagreement. Setting λ=0 recovers uniform error weighting.
Fast discovery, slow registration: a positive discovery-set delta is enough to enter an experimental pool. Default K=3 later cohorts must then show mean gain above δ (0.5 balanced-accuracy points for classification) with no rise in clinical cost. Candidates are frozen with a version hash. The planner collects new experience only with the stable registry; experimental tools are visible only to evaluator copies.
The starting harness is OpenHands. Diagnosis uses GPT-4o at temperature 0; reflection and tool building use 0.7. Public tasks run 20 rounds with up to six candidates and 12 A100 hours per round.
On REFUGE2 glaucoma, balanced accuracy moves from 56.2% to 94.1% over 20 rounds; on MITEA heart disease, from 56.5% to 82.3%. A matched MedAgent-Pro rerun sits at 89.6% and 77.1%. Glaucoma sensitivity rises from 28.4% to 92.0%, expected calibration error falls from 0.29 to 0.05. Five independent trajectories finish between 93.4%–94.7% and 81.2%–83.1%.
Of 118 proposed tools, 41 enter the pool and 16 register. Cup/disc Dice scores are 0.87 and 0.95; vertical cup-to-disc ratio correlates with reference at r=0.91. Left-ventricular Dice is 0.90, with 5.8-point mean absolute error on ejection fraction.
On two private tasks the same loop starts from MedAgent-Pro. Multimodal glaucoma balanced accuracy goes from 79.2% to 92.1% against a hand-built fusion network at 87.8%. Contrast-echo normalized error falls from 0.85 to 0.55 (MAE 5.1 points) against a task-specific regressor at 0.66, after the agent trains a noncontrast-to-contrast translator and retrains on real plus synthetic studies.
| Setup | Round-20 bAcc | Clinical cost /100 | Round-30 bAcc |
| Generic RSI | 77.4% | 9.2 | 72.7% |
| Uniform priority | 88.7% | 6.0 | 88.3% |
| Immediate registration | 82.4% | 7.6 | 76.9% |
| MedRSI | 94.1% | 2.5 | 94.4% |
Uniform priority spends 63% of early reflection on borderline suspects, hits 97.8% train accuracy, and stalls at 88.7% test. Immediate registration leads at round 8 (89.6%), then falls to 76.9% by round 30 with 57 tools. A rim-texture tool that gained +6.1 on discovery posted -2.3, -5.4, and -3.1 on trials; the full pipeline discarded it in the pool.
The transferable lesson is not "let the agent write tools." It is that medical RSI has to weight what it improves by clinical consequence, and has to delay what it keeps until later cohorts agree. A locally pretty discovery, once in the stable table, becomes the environment that later reflection learns from.
The run is expensive: about 212 accelerator hours, 1.9 billion reasoning tokens, and 4.5 days for one public-task trajectory. Across four tasks: 651 GPU hours, 3.2 billion tokens, 27 registered tools.
Clinical cost is still scored by a frozen GPT-4o judge; clinician weighted-κ is reported only on a calibration set. The mechanism assumes trustworthy labels and disjoint trial cohorts. Private data come from two tertiary centres; there is no prospective study of clinician use or patient outcomes. Swapping the backbone to Claude Sonnet 4.5, Gemini 2.5 Pro, or Qwen2.5-VL-72B yields round-20 glaucoma accuracy of 89.8%–93.6%, with 11 of 16 tool functions recovered by all four, still on the same endpoints. At λ=8, effort concentrates on a few severe cases and accuracy drops to 92.4%.