World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments

# Diagnosing RL in FHIR Environments
Source: [https://arxiv.org/html/2607.01470](https://arxiv.org/html/2607.01470)
## World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments

###### Abstract

Clinical protocol\-execution tasks—checking a lab value, applying a threshold, placing a correctly structured FHIR order—are natural candidates for RL from world feedback: once clinical SMEs encode decision logic into a verifier, that verifier grades unlimited rollouts without per\-episode annotation\. But applying RL requires a sound feedback channel and sufficient base capability\. We audit MedAgentBench v1/v2, find a 41\.7% silent\-finish ceiling that makes inaction the RL dominant strategy, and constructMedAgentBench\-v3 \(MAB\-v3\)\(508 tasks, 8\.9% ceiling\)\. Training Qwen3\-8B exposes two structural barriers: a*capability ceiling*\(10/20 task types have 0% base performance, zero gradient\) and a*format\-knowledge barrier*\(3/20 types require exact clinical codes undiscoverable by exploration\)\. Pure RL reaches 18\.2% pass@1 vs\. 34\.1% for rule\-based SFT; the 15\.9 pp gap is attributable entirely to these barriers\. A decision/format\-knowledge/lookup taxonomy predicts RL learnability and prescribes the fix: SFT to inject codes, RL to learn conditionals\.

reinforcement learning, clinical AI, FHIR, world feedback, GRPO

## 1Introduction

A large class of clinical tasks involves*protocol execution*: given a known decision rule, the agent retrieves a lab value, applies the threshold, and if triggered places a correctly structured FHIR order\. These are administrative workflow tasks—not replacement of physician judgment, but execution of standing orders\(Jianget al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib8); Leeet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib14); Bediet al\.,[2026](https://arxiv.org/html/2607.01470#bib.bib13)\)\. MedAgentBench v1 and v2, defining the 20 task types studied here, were constructed with clinical teams who validated each workflow against real EHR practice\(Jianget al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib8); Chenet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib9)\)\.

#### Why RL from world feedback:

Protocol correctness is verifiable: once a clinical SME encodes the decision logic into a verifier, that verifier grades every subsequent rollout automatically\. This is qualitatively different from RLHF\(Christianoet al\.,[2017](https://arxiv.org/html/2607.01470#bib.bib2); Ouyanget al\.,[2022](https://arxiv.org/html/2607.01470#bib.bib19)\): SME effort is front\-loaded into environment design rather than spent labeling individual episodes\. The alternative—supervised SFT demos—requires manually coding every clinical rule, is biased toward action\-branch instances, and must be regenerated when protocols change\. RL from world feedback avoids this: the agent explores the environment and receives feedback from the verifier, with only a single update needed when a protocol changes\.

#### What our experiments reveal:

Applying RL here is non\-trivial\. Two structural barriers limit a naive approach\. First, the feedback channel must be clean: MedAgentBench v1/v2 had a 41\.7% silent\-finish ceiling \(41\.7% of tasks pass with no tool use\), making inaction the RL dominant strategy\. GRPO\(Shaoet al\.,[2024](https://arxiv.org/html/2607.01470#bib.bib4)\)on uncorrected MAB\-v2 converged to 0% action\-branch pass\. We constructMedAgentBench\-v3 \(MAB\-v3\)\(508 tasks, 8\.9% ceiling\) to fix this\. Second, even on a clean benchmark, 3 of 20 task types require exact clinical codes \(SNOMED, NDC\) undiscoverable by exploration—flat reward landscape—and 10 types have zero base capability, yielding zero gradient\. These two barriers explain the 15\.9 pp gap: SFT \(34\.1%\) injects codes and format; pure RL \(18\.2%\) cannot\. The approaches are complementary, and SFT\+RL is the prescription our results motivate\.

#### Contributions:

- •MAB\-v3 \+ environment\(Section[3](https://arxiv.org/html/2607.01470#S3),[4\.1](https://arxiv.org/html/2607.01470#S4.SS1)\): Corrected 508\-task benchmark \(silent\-finish ceiling 41\.7%→\\to8\.9%\) with a self\-contained world feedback environment: offline FHIR server, auditable rule\-based verifier, and deliberate reward shaping for conditional clinical behavior\.
- •Structured diagnosis of RL limitations\(Section[4](https://arxiv.org/html/2607.01470#S4)–[5](https://arxiv.org/html/2607.01470#S5)\): Base 16\.6%, SFT 34\.1%, pure RL 18\.2%\. The 15\.9 pp SFT/RL gap is attributable to format\-knowledge and capability\-ceiling failures, not to the RL algorithm or reward design\.
- •Task taxonomy\(Section[6](https://arxiv.org/html/2607.01470#S6)\): Decision / format\-knowledge / lookup framework, validated by per\-type results, that predicts RL learnability from first principles and prescribes SFT\+RL as the right combination for mixed\-structure clinical benchmarks\.

## 2Related Work

#### FHIR agent evaluation:

Jianget al\.\([2025](https://arxiv.org/html/2607.01470#bib.bib8)\)introduced MedAgentBench, grounding clinical agent evaluation in FHIR tool use; MAB v2\(Chenet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib9)\)extended it with 300 tasks\. Neither examines benchmark validity as a training signal\. FHIR\-AgentBench\(Leeet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib14)\)targets factual retrieval from MIMIC\-IV rather than clinical action\. HealthAdminBench\(Bediet al\.,[2026](https://arxiv.org/html/2607.01470#bib.bib13)\)documents a subtask/task reliability gap parallel to the action/aggregate divergence we quantify\.

#### RL from verifiable non\-human feedback:

Outcome\-based RL with deterministic verifiers has driven large gains in mathematics\(Shaoet al\.,[2024](https://arxiv.org/html/2607.01470#bib.bib4); Guoet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib5)\)\. Our contribution is applying this paradigm to clinical environments and showing that task structure determines whether the verifier provides learnable signal—a distinction absent from homogeneous domains like arithmetic\. Multi\-task gradient dominance\(Wuet al\.,[2026](https://arxiv.org/html/2607.01470#bib.bib17); Rameshet al\.,[2026](https://arxiv.org/html/2607.01470#bib.bib15)\)is a known failure mode we address via per\-task advantage normalization\.

## 3MedAgentBench\-v3: Restoring the World Feedback Signal

### 3\.1Why the Original Benchmark Breaks RL

Two properties combine to make inaction the dominant RL strategy on MAB\-v2\.

Branch imbalance:Four v2 task types have 70–97% of instances on the no\-action branch due to cohort composition\. RL rewards the most common outcome; on these types, abstaining is correct the vast majority of the time\.

Silent\-finish ceiling:Running a null agent \(immediatefinish\(\[\]\), no tool calls\) on all 600 MAB\-v2 tasks yields 41\.7% pass\. This is not a bug in any single task—many patients genuinely need no intervention—but it creates a dominant RL strategy: ignoring FHIR data and abstaining scores 41\.7% before any clinical action is learned\. On uncorrected MAB\-v2, GRPO converged to 0% action\-branch pass within 200 steps \(Figure[1](https://arxiv.org/html/2607.01470#S3.F1)\)\.

### 3\.2Additional Evaluation Failures

Undocumented format requirements:Graders for v1\-T5, v1\-T9, and v2\-T3 enforce format conventions absent from the task context: bare\-string route fields, tiered dose formulas, and two\-element return arrays\. Clinically correct but incorrectly formatted responses fail systematically\. We add the missing documentation to task context\.

Wall\-clock bug:The v2\-T1 grader callsdatetime\.now\(\)as the CT follow\-up reference date, misclassifying 4 of the 30 v2\-T1 task instances as action\-required when run in 2025–2026\. We freeze the timestamp to2023\-11\-13T10:15:00\+00:00\.

### 3\.3MAB\-v3 Construction

Four corrections: \(1\) context patches \(v1\-T5, v1\-T9, v2\-T3\); \(2\) fixed timestamp \(v2\-T1\); \(3\) silent\-finish labeling for all 600 tasks; \(4\) 1:1 branch balance cap per \(corpus, task type\)\.

Result\.MAB\-v3 draws from two source corpora—MAB v1 \(original 300 tasks\) and MAB v2\-new \(300 new tasks\)—spanning 20 task types across≈\\approx100 real anonymized patients \(30 task instances per type, drawn from the same patient pool\)\. After curation it contains508 tasks: 463 action\-required, 45 no\-action\. Silent\-finish ceiling:8\.9%, down from 41\.7%\. The v1/v2\-new split is preserved throughout evaluation to distinguish the two distinct sets of clinical workflows\.

![Refer to caption](https://arxiv.org/html/2607.01470v1/x1.png)Figure 1:Branch imbalance before \(hatched\) and after \(solid\) the 1:1 cap in MAB\-v3, and the resulting fall in the silent\-finish ceiling from 41\.7% to 8\.9%\. The original imbalance made inaction the RL dominant strategy\.

## 4Methods

### 4\.1Environment Design

Figure[2](https://arxiv.org/html/2607.01470#S4.F2)shows the complete world feedback loop\. A well\-designed RL environment should ensure that score variations reflect the capability being measured, not environment instability, verifier bugs, or reward shortcuts\. Our benchmark audit \(Section[3](https://arxiv.org/html/2607.01470#S3)\) reveals that MAB\-v2 fails this on multiple dimensions\. We document three design principles MAB\-v3 addresses and two failure modes we observed in practice\.

Reproducible, deterministic episodes:The FHIR environment runs against a fixed snapshot of an HAPI FHIR server covering≈\\approx100 real anonymized patients\. Every rollout is fully deterministic: the same query always returns the same response, there is no server state, and episodes complete in under one second\. This is a prerequisite for the thousands of rollouts RL requires and eliminates environment instability as a confound—a known failure mode where environment instability causes failures unattributable to model capability\.

Reward shaping for conditional behavior:The world feedback signal decomposes as:

r=rterminal\+raction\+rpenalty,r=r\_\{\\text\{terminal\}\}\+r\_\{\\text\{action\}\}\+r\_\{\\text\{penalty\}\},\(1\)whererterminal=1\.0r\_\{\\text\{terminal\}\}=1\.0if the grader passes;raction∈\{0\.10,0\.25\}r\_\{\\text\{action\}\}\\in\\\{0\.10,0\.25\\\}gives partial credit for correct resource type and POST structure;rspurious=−0\.15r\_\{\\text\{spurious\}\}=\-0\.15penalizes an off\-target POST on a no\-action task;rskip=−0\.20r\_\{\\text\{skip\}\}=\-0\.20penalizes finish with no tool use\. The partial credit andrspuriousr\_\{\\text\{spurious\}\}are*deliberate design choices*, not natural properties of the clinical world\. Without partial credit, the reward landscape is flat until the exact correct order is placed, making early exploration unrewarding\. Withoutrspuriousr\_\{\\text\{spurious\}\}, the model learns to always act since action always earns partial credit\. These two components together create a gradient for both the action and the conditional structure\.

Auditable grader:The rule\-based verifier \(1,340 lines implementing medical protocol specifications across all 20 task types\) is itself the feedback provider: it queries the same FHIR environment the agent uses, then checks POST payloads against clinical criteria\. Every failure is traceable to a specific criterion\. This auditability is what lets us identify the wall\-clock bug and undocumented format requirements as grader failures rather than model capability gaps\.

Format consistency:Training, evaluation, and SFT demonstration generation all use the<tool\_call\>named\-function interface—the standard tool\-use format that modern LLMs are pre\-trained on\. We deliberately chose this over the original MAB harness’s benchmark\-specific format \(free\-form HTTP strings:GET http://\.\.\.,POST http://\.\.\.,FINISH\(\[\.\.\.\]\)\), which is a quirky interface designed for the official harness but not the standard way LLMs call tools\. The practical consequence is that our frontier baseline numbers \(Table[1](https://arxiv.org/html/2607.01470#S5.T1)\) are evaluated on the official harness, while our trained\-model results use the<tool\_call\>format—a necessary methodological note\. When we attempted to generate SFT teacher demos by running GPT\-5\.5 through our<tool\_call\>interface, only∼\\sim16% of training tasks produced passing episodes, confirming that two harnesses implementing the same FHIR tools produce very different pass rates\. Our<tool\_call\>format is more demanding precisely because it tests the model’s native tool\-use capability, not its ability to emit HTTP\-style strings\.

Observed failure mode — reward hacking:Despite these precautions, we observed reward hacking in our initial RL runs\. On uncorrected MAB\-v2, GRPO discovered the 41\.7% silent\-finish shortcut within 200 steps, converging to 0% action\-branch pass\. The model found the cheapest reward path the environment made available—a classic eval design failure rather than model deficiency\. MAB\-v3’s 8\.9% ceiling closes this shortcut\.

Observed failure mode — capability ceiling:Qwen3\-8B is substantially smaller than any frontier model evaluated on MAB\-v3; its zero\-shot capabilities reflect a very different pre\-training profile\. For GRPO to provide a useful gradient signal on a given task, the model must occasionally succeed: all\-fail rollouts yield zero advantage and zero gradient\. We measure this directly viafrac\_reward\_zero\_std, the fraction of task groups per step where all rollouts share the same reward\. Across the 89\-step training run, the mean is 0\.195 \(max 0\.750\)—roughly one in five groups provides no gradient signal at all\. Our per\-type analysis confirms the cause: 10 of 19 evaluated task types show 0% pass@1 for the base model, all of which are v1 lookup or format\-knowledge tasks\. RL gets zero gradient from these types throughout training\. Per\-task advantage normalization \(\-\-per\-task\-norm\) prevents the high\-variance decision tasks from entirely drowning out the zero\-variance types, but cannot manufacture a gradient where none exists\.

![Refer to caption](https://arxiv.org/html/2607.01470v1/x2.png)Figure 2:World feedback loop for clinical FHIR agent training\. The policy model \(Qwen3\-8B \+ LoRA\) generates<tool\_call\>actions; the FHIR environment executes them against an offline server snapshot and returns deterministic responses\. At episode end, a rule\-based verifier grades the agent’s actions against clinical protocol specifications and returns rewardrr—no per\-episode human judgment involved \(SME effort is concentrated in the one\-time verifier design\)\. GRPO uses this world feedback signal to update model weights\. The multi\-turn episode box \(dashed\) shows the interact\-observe loop; the training loop closes through GRPO at bottom\.
### 4\.2Evaluation Protocol

We split MAB\-v3’s 508 tasks 80/20 \(401 train, 107 test\), stratified by corpus, task type, and action label\. Trained model results use the held\-out split with 4 rollouts per task \(temperature 0\.7\)\. We report pass@1 \(mean pass rate\), pass@4 \(unbiased Chen et al\. estimator\), any\_pass \(≥\\geq1 rollout correct\), and all\_pass \(consistency: all 4 correct\)\. Frontier results use all 508 tasks, 1 sample each\. Base model: Qwen3\-8B\(Qwen Team,[2025](https://arxiv.org/html/2607.01470#bib.bib7)\)\.

### 4\.3SFT: Programmatic Distillation

We construct rule\-based SFT demos for the 401 training tasks programmatically: a rule\-based agent applies the known clinical decision per task type, reads actual patient FHIR data, and generates the correct POST where action is required\. Every episode is validated by the verifier \(reward≥1\.0\\geq 1\.0\); 354 of 401 pass\. Each demo includes the full tool\-call sequence—GET with FHIR Bundle response, POST with acceptance message, finish—in the exact<tool\_call\>format used at eval time\.

Qwen3\-8B is fine\-tuned with LoRA \(rank 64,α\\alpha=128\)\(Huet al\.,[2022](https://arxiv.org/html/2607.01470#bib.bib6)\), assistant\-only loss,max\_seq\_length14 000, batch size 1, gradient accumulation 16, lr2×10−42\\\!\\times\\\!10^\{\-4\}, 3 epochs\. No per\-episode human labels; SME effort went into verifier design\.

### 4\.4RL\-GRPO: Training from World Feedback

We apply GRPO\(Shaoet al\.,[2024](https://arxiv.org/html/2607.01470#bib.bib4)\)to Qwen3\-8B \(8B parameters, substantially smaller than frontier models\) directly from the base model—no SFT warmstart— on the 401\-task training set\. The FHIR environment is the sole feedback source: 4 generations per prompt, rewards from the deterministic verifier, advantages normalized*per task type*to prevent high\-variance decision types from dominating low\-variance lookup and format\-knowledge types\(Rameshet al\.,[2026](https://arxiv.org/html/2607.01470#bib.bib15)\)\. Hyperparameters:β=0\.05\\beta=0\.05,εhigh=0\.28\\varepsilon\_\{\\text\{high\}\}=0\.28\(DAPO clipping\(Yuet al\.,[2025](https://arxiv.org/html/2607.01470#bib.bib3)\)clipping\), temperature 1\.8, max 8 episode steps, 1 epoch\. We trackfrac\_reward\_zero\_std\(fraction of task groups per step where all rollouts share the same reward\) as the primary diagnostic for capability\-ceiling dead zones\.

## 5Results

Table 1:Frontier models on MAB\-v3 \(508 tasks, 1 sample each, official MAB harness\)\.Act= action\-branch pass rate \(463 instances\);No\-act= no\-action\-branch pass rate \(45 instances\)\. Net = p@1−\-8\.9 pp silent\-finish baseline\. All values in %\.∗Format non\-compliance; 443/554 responses rejected by harness\.

Table 2:Trained Qwen3\-8B on 107\-task held\-out split \(4 rollouts/task\)\.p@1= mean pass rate;p@4= pass@4 unbiased estimator;any=≥\\geq1 rollout passes ;all= all 4 pass \(consistency\)\. All values in %\.#### Frontier models:

Most frontier models show higher no\-action than action pass \(GPT\-5\.5: 93\.3% vs\. 77\.3%; Gemini: 95\.6% vs\. 76\.5%\)—they are better calibrated for abstention than for correct clinical action, a difference invisible in aggregate scores and exposed only by the action/no\-action split that MAB\-v3’s 1:1 balance enables\. GPT\-4o is the exception \(74\.7% vs\. 68\.9%\), with the smallest caution bias among frontier models\.

#### Trained models:

The base Qwen3\-8B achieves 16\.6% p@1\. Two observations characterize its behavior\. First, pass@1 and any\_pass are nearly equal \(16\.6% vs\. 21\.4%\), meaning failures are genuine inability rather than near\-misses: the model has little latent capability that extra attempts would unlock\. Second, a sharp corpus gap: v2 tasks score 28\.0% while v1 tasks score only 4\.8%\. v2 clinical decision tasks \(CT follow\-up, potassium replacement, flu vaccine recall\) partially overlap with pre\-training knowledge; v1 administrative tasks \(exact MRN lookup from name\+DOB, age calculation, BP recording with exact FHIR category coding\) require format\-specific knowledge the base model lacks entirely\.

SFT reaches 34\.1% p@1 \(any 44\.9%, all 25\.5%\)\. The corpus gap persists but narrows: rule\-based SFT demos lift v1 from 4\.8% to 18\.0% \(\+13\.2 pp\) and v2 from 28\.0% to 49\.2% \(\+21\.2 pp\)\. Three pass@k metrics tell a complete story\. The p@1→\\top@4 gap \(34\.1%→\\to43\.4%\) reflects residual stochasticity: 4 attempts reach 9\.3 pp more tasks than 1\. The any\_pass \(44\.9%\) is the capability ceiling—nearly half of test tasks can be solved at least once\. The all\_pass \(25\.5%\) is the reliability floor—tasks the model solves consistently\. The 20 pp gap between any and all signals that a large fraction of capability is inconsistent, likely reflecting the limited demo count per task type \(24 demos covering all instances for most types, but as few as 1–2 for rare task types\)\.

RL \(GRPO, epoch 1\) achieves 18\.2% p@1 \(any 23\.5%, all 13\.3%\), a \+1\.6 pp improvement over base but 15\.9 pp*below*SFT\. Both v1 \(7\.7%\) and v2 \(32\.1%\) improve modestly over base \(4\.8%, 28\.0%\), but the corpus gap actually widens slightly \(24\.4 pp in RL vs\. 23\.2 pp in base\)\. The any\_pass≈\\approxp@1 pattern from base \(21\.4% vs\. 16\.6%\) is reproduced in RL \(23\.5% vs\. 18\.2%\)—confirming that RL, like the base model, has limited latent capability that extra attempts cannot unlock, rather than high capability with low consistency\. In contrast, SFT’s larger any/p@1 gap \(44\.9% vs\. 34\.1%\) reflects genuine stochasticity in a more capable model\. Section[6](https://arxiv.org/html/2607.01470#S6)interprets these patterns through the task taxonomy\.

## 6Analysis: When Does World Feedback Help?

### 6\.1Task Taxonomy and RL Learnability

We partition MAB\-v3’s 20 task types into three categories by reward landscape structure\. This taxonomy is derived*before*seeing RL results; it is a prediction, not a post\-hoc rationalization\.

Table 3:Task taxonomy\. RL learnability depends on whether reward variance exists across rollouts—i\.e\., whether the environment provides a gradient\.Table 4:Per\-type RL vs\. base pass@1 on held\-out test split\.nn= test instances \(4 rollouts each\)\. Format\-knowledge marked∗; dead\-zone types \(0% base*and*RL, zero gradient\) marked†\.Decision tasks\(11 types\) require reading a lab value and applying a clinical threshold\. The reward varies with the decision: ordering potassium for K=2\.8 mEq/L earnsr≥1\.0r\\geq 1\.0; ordering for K=4\.2 \(normal\) earnsr=−0\.15r=\-0\.15\. This gradient exists because the decision boundary is a learnable function of an observable FHIR value\. Both action and no\-action rollouts occur during exploration, and the spurious\-post penalty directly trains the conditional\.

Lookup tasks\(6 types\) require retrieving and returning an exact value: a patient MRN from name and DOB, a patient age, the most recent lab result\. The answer is read from the environment, not derived by reasoning\. RL can learn which tool to call and how to parse the response, but demonstration is at least as efficient and more reliable\.

Format\-knowledge tasks\(3 types\) require an exact clinical code not inferrable from context\. Naloxone coverage requires SNOMED 306181000000106; IV magnesium requires NDC 0338\-1715\-40\. Placing a clinically correct order with the wrong code earns−0\.15\-0\.15—the same reward as random wrong codes\. There is no gradient pointing toward the correct identifier: the reward landscape is flat\. RL cannot discover discrete codes by gradient ascent\.

Preliminary support from frontier results:The Appendix Table[5](https://arxiv.org/html/2607.01470#A1.T5)already supports this taxonomy: format\-knowledge types v2\-T5 and v2\-T8 score 0% for GPT\-5\.5 and Gemini despite their 78%\+ overall performance \(flat landscape, even frontier models struggle without in\-context code knowledge\), while decision type v2\-T9 \(flu vaccine recall\) is solved by all frontier models\. The base Qwen3\-8B scores 28\.0% on v2 \(predominantly decision tasks\) vs\. 4\.8% on v1 \(heavier in lookup/format\), consistent with the taxonomy\.

Empirical validation:Table[4](https://arxiv.org/html/2607.01470#S6.T4)reports per\-type RL vs\. base deltas\. The pattern is consistent with the taxonomy predictions:

- •*Dead zones*: 10 of 19 evaluated types are 0% for both base and RL \(all v1 types except task6/task7, plus v2/task3, task7\)\. These are predominantly v1 lookup and format\-knowledge tasks where the 8B base model has no zero\-shot capability, so all 4 RL rollouts fail identically and produce zero gradient\.
- •*Decision gains*: v2 types with clear threshold structure show positive RL deltas—v2/task1 \(\+4\.2 pp\), v2/task2 \(\+3\.6 pp\), v2/task5 \(\+12\.5 pp, though n=2\)\. The largest single gain is v1/task6 \(\+20\.8 pp\), a CBG\-average lookup that benefits from RL learning the correct FHIR code \(GLU, not CBG\) from partial reward\.
- •*Format\-knowledge harm*: v2/task8 \(naloxone, SNOMED 306181000000106\)*decreases*from 20\.8% to 16\.7% under RL\. The model had occasional lucky recall from pre\-training; RL training with wrong\-code rollouts overwrites this latent knowledge\.

The most striking finding is the*magnitude*of RL’s underperformance relative to SFT\. Pure RL from base reaches only 18\.2%—a 15\.9 pp gap behind supervised distillation—despite the same environment and reward signal that would, in principle, provide all necessary information\. This gap quantifies what world feedback alone cannot provide: format knowledge \(exact clinical codes, FHIR payload structure\) must be injected via supervised learning\. RL then refines the decision logic on top\. The SFT\+RL combination is the natural prescription, which we leave as immediate future work\.

### 6\.2Format\-Knowledge Tasks: The Hard Limit of World Feedback

The environment tells the model*that*it failed \(wrong code→\\tonegative reward\), but cannot tell it*what code to use*\. Trying 100 different SNOMED codes yields 100 identical failures\. SFT from rule\-based SFT demos solves this directly: the demo shows the exact code, and the model reproduces it for held\-out patients\. This is the one setting where world feedback is structurally insufficient and supervised injection is required\.

Partial evidence from pre\-training:Format\-knowledge type v1\-T8 \(orthopedic referral\) is solved by most frontier models at 87–100% \(Appendix[A](https://arxiv.org/html/2607.01470#A1)\), suggesting SNOMED 306181000000106 for ortho referral appears in frontier pre\-training data\. In contrast, v2\-T5 \(Mg IV, NDC 0338\-1715\-40\) and v2\-T8 \(naloxone, same SNOMED code in a different context\) score 0% for most models\. Format\-knowledge tasks are not uniformly hard—they depend on whether the exact code was seen in pre\-training\.

Inspection of RL rollouts on v2\-T8 \(naloxone, SNOMED 306181000000106\) confirms this pattern\. The RL model consistently attempts a ServiceRequest POST with the correct resource type but substitutes plausible\-sounding SNOMED codes \(e\.g\., codes for “medication administration” or “emergency treatment”\) rather than the exact required identifier\. Each attempt receivesr=−0\.15r=\-0\.15\(spurious\-post on a technically\-incorrect order\), but since all codes fail with the same reward, the model receives no gradient signal indicating which code is correct\. The base model occasionally produces the right SNOMED code from pre\-training knowledge \(any\_pass\>\>0 for v2\-T8\), but does so inconsistently \(all\_pass≈\\approx0\)\. SFT, having seen the exact code in rule\-based SFT demos, reproduces it reliably across all 4 rollouts for most patients\. This three\-way pattern—RL converges to a wrong attractor, base has occasional lucky recall, SFT has consistent reproduction—directly illustrates the flat\-landscape failure mode\.

### 6\.3RL and Conditional Reasoning

SFT trains exclusively on action\-branch demos \(the rule\-based SFT generator only produces no\-action traces for task types where the data\-driven decision yields no intervention\)\. RL, by contrast, sees both branches during exploration and receives explicit negative reward for spurious action, creating a direct gradient for the action/no\-action conditional\.

The any\_pass metric provides a proxy for this: models that correctly identify no\-action tasks will have lower any\_pass \(fewer tasks where any rollout passes\), while models that over\-act will have spuriously high any\_pass from action\-branch tasks\. RL’s any\_pass \(23\.5%\) is only marginally above base \(21\.4%\), suggesting the spurious\-post penalty did not substantially improve conditional calibration in one epoch from scratch\. SFT’s higher any\_pass \(44\.9%\) likely reflects better overall task comprehension rather than superior conditional reasoning, since the rule\-based SFT demos were predominantly action\-branch\. A full action/no\-action breakdown requires rerunning eval with the labeled test split; the any\_pass proxy is consistent with RL not yet learning the conditional from world feedback alone\.

## 7Discussion

#### What the results show vs\. what they imply:

The 15\.9 pp gap between SFT \(34\.1%\) and pure RL \(18\.2%\) should not be read as evidence that RL is the wrong approach for clinical agents\. It is evidence that*pure RL from a small base model without prior code knowledge*faces two structural barriers identified by our taxonomy\. This is a diagnosis, not a verdict\. The correct reading: rule\-based SFT has something RL lacks \(format knowledge and clinical codes\); RL has something rule\-based SFT lacks \(scalability through SME\-amortized environment design rather than per\-episode annotation, and natural coverage of no\-action branches\)\. The prescriptive implication is SFT\+RL—inject codes and format via supervised learning, then apply RL to refine conditional reasoning on the world feedback signal\. Our benchmark and environment are designed to test this combination, which we leave as immediate future work\(Gaoet al\.,[2023](https://arxiv.org/html/2607.01470#bib.bib11)\)\.

#### Clean world feedback requires a clean benchmark:

The analogy to reward model quality in RLHF\(Gaoet al\.,[2023](https://arxiv.org/html/2607.01470#bib.bib11)\)holds exactly: the uncorrected 41\.7% silent\-finish ceiling made inaction the RL dominant strategy, independent of algorithm choice\. MAB\-v3 restores the signal\. Any RL method applied to clinical protocol execution must first verify that the benchmark’s silent\-finish ceiling does not create a reward\-hacking shortcut\. Our four\-failure audit provides a checklist for this verification\.

#### Scope and limitations:

This work is scoped to*protocol\-execution tasks*: the agent executes known institutional decision rules; it does not make novel clinical judgments or replace physician oversight\. Results do not generalize to open\-ended clinical reasoning, differential diagnosis, or tasks where correctness is contested or context\-dependent\. MAB\-v3 uses≈\\approx100 real anonymized patients drawn from a HAPI FHIR server, with 30 task instances per type \(patients reused across types\); some types have as few as 1–2 action instances after the 1:1 cap, making per\-type estimates noisy\. Task instructions are more explicit than real clinical documentation, likely overstating readiness for deployment\. The 20 task types cover a specific set of administrative EHR workflows; extension to broader clinical protocols is future work\. The RL condition here is pure GRPO from base with 1 epoch—an SFT warmstart before RL is the direct next experiment\.

## References

- S\. Bedi, R\. Welch, E\. Steinberg, M\. Wornow, T\. M\. Kim, H\. Ahmed, P\. Sterling, B\. Purohit, Q\. Akram, A\. Acosta,et al\.\(2026\)HealthAdminBench: evaluating computer\-use agents on healthcare administration tasks\.arXiv preprint arXiv:2604\.09937\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.p1.1),[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px1.p1.1)\.
- E\. Chen, S\. Postelnik, K\. Black, Y\. Jiang, and J\. H\. Chen \(2025\)MedAgentBench v2: improving medical LLM agent design\.InBiocomputing 2026: Proceedings of the Pacific Symposium,pp\. 354–371\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.p1.1),[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px1.p1.1)\.
- P\. F\. Christiano, J\. Leike, T\. Brown, M\. Martic, S\. Legg, and D\. Amodei \(2017\)Deep reinforcement learning from human preferences\.InAdvances in Neural Information Processing Systems,Vol\.30\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.SS0.SSS0.Px1.p1.1)\.
- L\. Gao, J\. Schulman, and J\. Hilton \(2023\)Scaling laws for reward model overoptimization\.ICML\.Cited by:[§7](https://arxiv.org/html/2607.01470#S7.SS0.SSS0.Px1.p1.1),[§7](https://arxiv.org/html/2607.01470#S7.SS0.SSS0.Px2.p1.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi,et al\.\(2025\)Deepseek\-r1: incentivizing reasoning capability in llms via reinforcement learning\.arXiv preprint arXiv:2501\.12948\.Cited by:[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px2.p1.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.ICLR\.Cited by:[§4\.3](https://arxiv.org/html/2607.01470#S4.SS3.p2.2)\.
- Y\. Jiang, K\. C\. Black, G\. Geng, D\. Park, J\. Zou, A\. Y\. Ng, and J\. H\. Chen \(2025\)MedAgentBench: a virtual EHR environment to benchmark medical LLM agents\.NEJM AI2\(9\),pp\. AIdbp2500144\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.p1.1),[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px1.p1.1)\.
- G\. Lee, E\. Bach, E\. Yang, T\. Pollard, A\. Johnson, E\. Choi, J\. H\. Lee,et al\.\(2025\)FHIR\-AgentBench: benchmarking LLM agents for realistic interoperable EHR question answering\.arXiv preprint arXiv:2509\.19319\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.p1.1),[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px1.p1.1)\.
- L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.\(2022\)Training language models to follow instructions with human feedback\.Advances in Neural Information Processing Systems35,pp\. 27730–27744\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.SS0.SSS0.Px1.p1.1)\.
- Qwen Team \(2025\)Qwen3 technical report\.Note:[https://qwenlm\.github\.io/blog/qwen3/](https://qwenlm.github.io/blog/qwen3/)Cited by:[§4\.2](https://arxiv.org/html/2607.01470#S4.SS2.p1.1)\.
- S\. S\. Ramesh, X\. Ji, M\. Zimmer, S\. Yoon, Z\. Wang, H\. B\. Ammar, A\. Lucchi, and I\. Bogunovic \(2026\)Multi\-task grpo: reliable llm reasoning across tasks\.arXiv preprint arXiv:2602\.05547\.Cited by:[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px2.p1.1),[§4\.4](https://arxiv.org/html/2607.01470#S4.SS4.p1.2)\.
- Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.\(2024\)Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§1](https://arxiv.org/html/2607.01470#S1.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px2.p1.1),[§4\.4](https://arxiv.org/html/2607.01470#S4.SS4.p1.2)\.
- R\. Wu, A\. Samanta, A\. Jain, S\. Fujimoto, J\. Kwon, B\. Kretzu, Y\. Yu, K\. Hassani, B\. Vidolov, and Y\. Efroni \(2026\)Imbalanced gradients in rl post\-training of multi\-task llms\.InFindings of the Association for Computational Linguistics: EACL 2026,pp\. 3137–3150\.Cited by:[§2](https://arxiv.org/html/2607.01470#S2.SS0.SSS0.Px2.p1.1)\.
- Q\. Yu, Z\. Zhang, R\. Zhu, Y\. Yuan, X\. Luo, W\. Fan, J\. Liu,et al\.\(2025\)DAPO: an open\-source llm reinforcement learning system at scale\.arXiv preprint arXiv:2503\.14476\.Cited by:[§4\.4](https://arxiv.org/html/2607.01470#S4.SS4.p1.2)\.

## Appendix AMAB\-v3 Per\-Task Frontier Results

Table[5](https://arxiv.org/html/2607.01470#A1.T5)reports action\-branch pass rates per task type on MAB\-v3 for all six frontier models\.

Table 5:Action\-branch pass rate \(%\) by task type on MAB\-v3 \(508 tasks\)\.nn= action\-branch instances\. Format\-knowledge marked∗\.−\-=≤\\leq1 action instance after 1:1 cap\.G5\.5=GPT\-5\.5; Gem=Gemini 3\.1 Pro; G4o=GPT\-4o; L4M=Llama 4 Maverick; Mis=Mistral Large; Son=Claude Sonnet 4\.6\.

Three patterns support the taxonomy\. \(1\) Format\-knowledge types v2\-T5 \(nn=2\) and v2\-T8 \(nn=20\) illustrate the flat landscape at different scales: v2\-T8 has 20 action instances and still scores 0% for GPT\-5\.5/Gemini/Llama/Mistral, confirming the pattern is not a small\-sample artifact; v2\-T5 has only 2 instances so individual results should be read cautiously\. v1\-T8 \(ortho referral,nn=30\) is a format\-knowledge exception \(87–100%\), likely because SNOMED 306181000000106 appears in frontier pre\-training data\. \(2\) Decision types v2\-T9 \(flu vaccine\), v2\-T10 \(COVID booster\), v2\-T1 \(CT follow\-up\) are solved by all or most frontier models, consistent with learnable threshold structure\. \(3\) Lookup types show extreme variance: v1\-T1 \(MRN lookup from name\+DOB\) ranges from 7% \(GPT\-4o\) to 100% \(Gemini\), and v1\-T7 \(most recent CBG\) scores 17–40%, suggesting the exact value\-matching requirement interacts with output format in unpredictable ways\.

## Appendix BSFT: Programmatic Demo Construction

For each of the 401 training tasks, a rule\-based agent applied the known clinical decision per task type, queried the FHIR environment using grader\-compatible URL parameters \(without\_sort=\-date, which changes the query and returns fewer entries than the grader sees\), computed the data\-driven decision, and constructed the correct POST payload\. Each episode was validated by the FHIR verifier; 354/401 pass\. Key non\-obvious details: \(1\) potassium dose is computed as\(3\.5−K\)/0\.1×10\(3\.5\-K\)/0\.1\\times 10mEq from the actual observed value, not from the task label \(which occasionally mismatches the grader due to routing inconsistencies\); \(2\) Mg level decisions use data within the last 24h only, matching the grader’s window; \(3\) the “route” field in dosage instructions must be a plain string \(“IV”, “oral”\), not a FHIR coding object—the undocumented format requirement that caused systematic failures on v1\-T5 and v1\-T9 in the original benchmark\.