Benchmark · Open Source · 2026

CareBench: When Correct Diagnoses
Come from Unsafe Reasoning

Elias Hossain1, Tasfia Nuzhat Ornee1, Shubhashis Roy Dipta2, Andrea Rivera-Sepulveda3, Pegah Khosravi1, Niloofar Yousefi1

1University of Central Florida  ·  2University of Maryland, Baltimore County  ·  3Nemours Children’s Health

Medical-agent leaderboards score one thing: did the agent reach the right diagnosis. That hides a measurable failure mode — an agent can be right by way of a process no clinician would endorse. CareBench measures the process, not just the answer.

Code 🤗 Dataset
1,000
Trajectories
5
Frontier Models
50
Clinical Cases
0.788
Pooled B-LCR

The Problem

Lucky correctness

Across 1,000 trajectories from five frontier models, nearly four in five correct diagnoses were reached without any invoked safety check (pooled B-LCR = 0.788), replicating an earlier preliminary study at 0.795.

An agent that commits to a diagnosis before gathering the necessary evidence, or that skips a red-flag check, can still land on the right answer. On an outcome-only leaderboard that trajectory is indistinguishable from careful clinical reasoning. We call the gap lucky correctness, and it is invisible to any metric that reads only the final answer.

Outcome hides process

Diagnostic accuracy is one bit per case. It cannot distinguish an agent that ordered the right workup from one that guessed early and got lucky, even though only one of those is safe to deploy.

Rankings invert

The model that ranks first on diagnostic accuracy in our evaluation ranks near the bottom on process safety. Choosing a model by outcome alone can select the least safe reasoner in the pool.

Policy stays hidden

Each case ships a JSON clinical policy — necessary evidence, red flags, accepted diagnoses and treatments, closure threshold — visible only to the scoring layer, never to the agent or the patient simulator.


What It Measures

Outcome and process, measured separately

Every trajectory is scored on both what the agent concluded and how it got there. CareBench ships as a third clinical domain of the tau-bench tool-agent-user framework, alongside retail and airline, reusing its agent loop, tool-calling interface, user simulator, and pass^k pipeline.

MetricWhat it captures
DxAccDiagnostic accuracy, via a conservative layered diagnosis matcher
ENREvidence Necessity Recall: fraction of case-required evidence actually gathered
PEPathway Efficiency: fraction of gathered evidence that was necessary
PCRPremature Closure Rate: committed before sufficient evidence
RFCRRed-Flag Capture Rate: clinically relevant red flags acknowledged
NOSAFETYReference-free action-log signal: detects commitment without a required safety check
LCRLucky Correctness Rate: correct diagnosis reached through an unsafe process
B-LCRBehavioral variant, policy-independent, read only from the action log

The two primary trajectory-level signals are ENR (evidence coverage) and NOSAFETY (the reference-free action-log signal); the remaining pathway metrics are reported as supporting diagnostics. B-LCR is the correctness-conditioned summary of NOSAFETY, and it is the headline number precisely because it is policy-independent: computed from what the agent did, it cannot be gamed by disagreement over what a case policy should have required.

The outcome-process partition

Safe · correct
Right diagnosis, defensible pathway. The only quadrant an outcome-only metric and a clinician would both accept.
Lucky · correct
Right diagnosis, unsafe pathway. Scored as a win by every outcome-only leaderboard.
Safe · wrong
Wrong diagnosis, sound process. Penalised identically to reckless failure by outcome-only scoring.
Unsafe · wrong
Wrong on both counts.

A composite pass^k requires both a correct outcome and a safe process, so a model cannot buy the score with lucky correctness.


Results

Outcome-only and process-aware selection disagree

Five frontier models, 200 trajectories each. Read the first two numeric columns together: the ordering by diagnostic accuracy is close to the reverse of the ordering by process safety.

ModelDxAcc ↑B-LCR ↓PE ↑ENR ↑Safe-correctLucky-correct
gpt-4o-mini0.1201.0000.7250.225024
llama-70b0.0750.7330.8120.238114
deepseek-chat0.0650.3850.8500.263310
gpt-4o0.0550.8180.8550.26929
claude-sonnet0.0151.0000.4550.14103

Outcome-only vs process-aware model selection. n = 200 trajectories per model. gpt-4o-mini ranks first on diagnostic accuracy while scoring B-LCR = 1.000 — every one of its 24 correct diagnoses came from an unsafe pathway, and it produced zero safe-correct trajectories. deepseek-chat ranks third on accuracy but first on process safety, with the most safe-correct trajectories in the pool. Pick by accuracy and you ship the least safe reasoner.

Replicated

The pooled B-LCR of 0.788 reproduces an earlier preliminary study at 0.795, on independently generated trajectories. The effect is not an artifact of one run or one scoring pass.

Evidence is thin everywhere

ENR sits between 0.141 and 0.269 across all five models: even the best of them gathers barely a quarter of the evidence its cases require before concluding.

Regenerable offline

Every table and figure rebuilds from saved trajectories with no further model calls, so the analysis is reproducible without paying for inference twice.


Quick Start

Run the benchmark

The case pool is 50 clinical cases: 19 translated from MedChain and 31 from PMC-Patients. The harness routes through LiteLLM, so any provider it supports can be evaluated.

install.shBash
# Clone and install
git clone https://github.com/eliashossain001/carebench.git
cd carebench
pip install -e .

# Pull the case pool and trajectories
huggingface-cli download EliasHossain/carebench \
  --repo-type dataset \
  --local-dir data/carebench/
run.shBash
# Evaluate an agent on the clinical domain
python run.py \
  --env carebench \
  --model gpt-4o \
  --model-provider openai \
  --user-model gpt-4o-mini \
  --user-model-provider openai \
  --agent-strategy tool-calling \
  --num-trials 4 \
  --max-concurrency 8

--num-trials sets k for pass^k; --agent-strategy is one of tool-calling, act, react, or few-shot.

Research use only. CareBench is an evaluation benchmark, not a clinical tool. Its cases and policies exist to measure agent behaviour, and nothing here is validated for patient care. Cases derive from MedChain and PMC-Patients; please cite those datasets alongside CareBench when using the corresponding cases.

Citation

Cite CareBench

If CareBench or the lucky-correctness metrics are useful in your work, please consider citing it.

@misc{hossain2026carebench, title = {CareBench: When Correct Diagnoses Come from Unsafe Reasoning}, author = {Hossain, Elias and Ornee, Tasfia Nuzhat and Dipta, Shubhashis Roy and Rivera-Sepulveda, Andrea and Khosravi, Pegah and Yousefi, Niloofar}, year = {2026}, note = {Process-aware evaluation for clinical LLM agents. Code: \url{https://github.com/eliashossain001/carebench} Data: \url{https://huggingface.co/datasets/EliasHossain/carebench}}, }