World Feedback for Clinical Agents: the new MedAgentBench findings

An arXiv paper by Ananya Mantravadi, Harshit Rajgarhia, Prasanna Desikan and Abhishek Mukherji, submitted 1 Jul 2026, audits prior MedAgentBench versions and releases MedAgentBench-v3 (MAB-v3). MAB-v3 contains 508 tasks, and the authors report that earlier MedAgentBench v1/v2 exhibited a 41.7% silent-finish ceiling, while MAB-v3 reduces that to an 8.9% silent-finish ceiling. The paper trains Qwen3-8B and measures learning barriers and pass rates under pure reinforcement learning and rule-based supervised fine-tuning.

What did the paper add to MedAgentBench?

The paper constructs MedAgentBench-v3 with 508 tasks and explicitly quantifies the silent-finish failure mode across benchmark versions. The direct findings are: MedAgentBench v1/v2 had a 41.7% silent-finish ceiling that made inaction the dominant RL strategy, while MAB-v3 reduces that ceiling to 8.9%, creating a feedback channel more suitable for RL from world feedback.

The authors position MAB-v3 as a diagnostic benchmark for clinical protocol-execution tasks that require actions such as checking lab values, applying thresholds, and placing correctly structured FHIR orders. The paper argues these tasks are natural candidates for reinforcement learning from world feedback, provided the feedback channel and base capability are adequate.

How did Qwen3-8B perform and what blocked RL?

Training Qwen3-8B exposes two structural barriers: a capability ceiling and a format-knowledge barrier. The capability ceiling appears as zero base performance on half the task types, specifically 10 out of 20 task types showing 0% base performance and producing zero gradient for RL training. The format-knowledge barrier affects 3 out of 20 task types that require exact clinical codes which the agent cannot discover by exploration.

Quantitatively, pure RL achieved 18.2% pass@1, while a rule-based supervised fine-tuning (SFT) approach reached 34.1% pass@1. The authors attribute the full 15.9 percentage-point gap between pure RL and rule-based SFT entirely to the two barriers above. They propose a taxonomy (decision/format-knowledge/lookup) that predicts RL learnability and prescribes remedies: use SFT to inject exact codes and use RL to learn conditional decision logic.

Why it matters

The paper demonstrates concrete limits that arise when applying reinforcement learning to clinical protocol tasks in FHIR environments. The 41.7% silent-finish ceiling in earlier benchmarks, and the remaining 8.9% ceiling in MAB-v3, show that naive RL can be rewarded for inaction unless the environment and verifier are designed to penalize silent finishes. The measured gap between pure RL (18.2% pass@1) and rule-based SFT (34.1% pass@1), and the identification of 10/20 task types with 0% base performance, clarify where model capability and data-format knowledge must be supplied before RL can improve outcomes.

These results matter to teams trying to use world feedback rather than costly per-episode annotation: the verifier approach can grade unlimited rollouts, but only if the agent has sufficient base capability and access to precise format knowledge. The paper's prescription, to inject format knowledge via SFT while using RL to learn conditional logic, gives a concrete path for practitioners working on clinical agents that interact with FHIR orders and coded clinical data.

What to watch

Watch for experiments that combine SFT injection of exact clinical codes with RL for conditional decision-making, and for follow-up benchmarks reporting whether that hybrid approach closes the 15.9 percentage-point gap the authors measured. Also watch for broader adoption of MAB-v3 (508 tasks) as a diagnostic standard for clinical RL in FHIR environments.

Authors, identifier and date: the paper is titled "World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments," arXiv:2607.01470, submitted 1 Jul 2026, by Ananya Mantravadi, Harshit Rajgarhia, Prasanna Desikan and Abhishek Mukherji.