Sumit Saraswat
feat: final submission with 70B dashboard and SOTA ReAct loop
5afe05e
metadata
title: ClinicalBench
emoji: 🔬
colorFrom: blue
colorTo: green
sdk: docker
app_port: 8000
pinned: false
tags:
  - openenv

🔬 ClinicalBench

A Benchmark for Evaluating Agentic Reasoning in Safety-Critical Clinical Workflows

OpenEnv HF Space Docker License 70B Score 405B Score 720 Patients Multi-Hop

🎯 Llama 3.3 70B beats the 405B frontier model (0.66 vs 0.50). ClinicalBench is an OpenEnv benchmark where LLMs audit 720 oncology patient records against procedurally generated protocols. By utilizing multi-hop comorbidity traps, Simpson's Paradox confounders, and a brutal -0.30 false-positive penalty, ClinicalBench proves that agentic tool-calling efficiency (3.3 70B) outperforms raw parameter size (3.1 405B) in safety-critical workflows.

Live Demo · Architecture · Results · Quick Start · Leaderboard


🖥️ The Enterprise Audit Dashboard (Live Demo)

Because safety-critical AI requires transparency, ClinicalBench includes a production-ready enterprise dashboard to visualize the agent's ReAct loop in real-time.

Launch the Hugging Face Space to see the 70B reasoning agent actively triage patients, compute bias distributions, and flag protocol violations while safely navigating the 8K token context limit.


The Problem

Clinical data auditing is one of medicine's most consequential workflows. A single undetected protocol violation can invalidate years of trial data, delay drug approvals, and — in worst cases — put patients at risk. Today's AI systems fail at this task in three specific ways:

Failure Mode What Happens Why It Matters
Overflagging LLMs flag valid edge cases (e.g., Stage IV patients with extended treatment windows) as violations False alarms waste reviewer time and erode trust in AI-assisted auditing
Temporal Confusion Models miss impossible date orderings (death before treatment) while fixating on superficial anomalies Critical safety signals go undetected
Bias Misinterpretation Models detect demographic skew in raw statistics but cannot distinguish genuine selection bias from confounded high-risk cohorts Naive bias detection causes incorrect escalations or dangerous dismissals

ClinicalBench is designed to evaluate and train agents that can overcome all three failure modes simultaneously.


Why ClinicalBench Exists

Existing RL benchmarks for agents fall into two categories: game-like environments (code golf, math puzzles) where memorization helps, and static dataset tasks (classification, extraction) where the answer is fixed. Neither captures the reality of clinical auditing, where:

  • Rules change every episode — eligibility criteria, timing windows, and bias thresholds are protocol-specific
  • Edge cases are not errors — Stage IV patients legitimately have longer treatment windows
  • Statistics lie without context — a minority group's higher mortality rate may reflect disease severity, not unfair sampling
  • The step budget is limited — agents must prioritize which patients and which patterns to investigate

ClinicalBench fills this gap by generating a new procedural dataset and protocol for every reset(), forcing agents to read and reason rather than memorize.


Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    ClinicalBench Architecture                   │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  reset(seed, task_id)                                           │
│        │                                                        │
│        ▼                                                        │
│  ┌──────────────────────┐    ┌─────────────────────────────┐    │
│  │  Procedural Dataset  │───▶│  Episode-Specific Protocol  │    │
│  │  Generator           │    │  Excerpt                    │    │
│  │  • 300-720 patients  │    │  • Dynamic age range        │    │
│  │  • Seeded RNG        │    │  • Variable timing windows  │    │
│  │  • Adversarial traps │    │  • Stage IV exceptions      │    │
│  │  • Hidden confounders│    │  • Bias thresholds          │    │
│  └──────────────────────┘    └─────────────────────────────┘    │
│        │                              │                         │
│        ▼                              ▼                         │
│  ┌─────────────────────────────────────────────────────────┐    │
│  │              Agent Interaction Loop                     │    │
│  │  Thought → Tool → Observation → Flag → Report           │    │
│  ├─────────────────────────────────────────────────────────┤    │
│  │  investigate_pattern(var)   → distribution summary      │    │
│  │  compute_distribution(var) → cohort breakdown           │    │
│  │  flag_error(patient, type) → correct/false positive     │    │
│  │  submit_report(text)       → quality score              │    │
│  └─────────────────────────────────────────────────────────┘    │
│        │                                                        │
│        ▼                                                        │
│  ┌─────────────────────────────────────────────────────────┐    │
│  │              Multi-Dimensional Grading                  │    │
│  │  Recall (70%) + Precision (15%) + Workflow (5%)         │    │
│  │  + Efficiency (5%) + Report Quality (5%)                │    │
│  │  Dense step rewards + episode benchmark score           │    │
│  └─────────────────────────────────────────────────────────┘    │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Key Design Decisions

  1. Procedural Generation — Each reset() samples a new protocol with different age ranges, timing windows, and bias thresholds using seeded stochastic processes. No two environments are identical, preventing memorization.

  2. Adversarial Traps — Valid edge cases (boundary ages, near-window delays, valid Stage IV exceptions) are deliberately injected to punish agents that use naive threshold-based heuristics.

  3. Confounder-Aware Bias — Hard episodes may contain either genuine selection bias OR a confounded high-risk cohort. The confounder (high-risk outreach site with more late-stage patients) creates an overall mortality gap that disappears after stage-stratified analysis. Agents must perform this adjustment before flagging.

  4. Phase-Gated Workflow — Agents must investigate variables before flagging errors, and compute distributions before claiming bias. Skipping phases is penalized, encouraging structured reasoning over guessing.


Task Suite

Task 1: task_easy — Dynamic Eligibility Screening

Property Value
Dataset ~300 patients
Error types invalid_age
Difficulty source Age bounds are episode-specific (e.g., 35-75, 45-85), not fixed at 18-120
Traps Valid boundary ages at exact protocol limits
Step budget 25

Task 2: task_medium — Protocol Timeline Audit

Property Value
Dataset ~480 patients
Error types invalid_age, temporal_inconsistency, protocol_window_violation
Difficulty source Treatment-start window is protocol-specific; Stage IV has a longer valid window
Traps Near-boundary delays, valid Stage IV exceptions, near-immediate valid deaths
Step budget 50

Task 3: task_hard — Equity + Protocol Audit

Property Value
Dataset ~720 patients with 25+ fields (including 11 clinical noise columns)
Error types invalid_age, temporal_inconsistency, protocol_window_violation, selection_bias
Difficulty source Multi-hop comorbidity exception, Simpson's Paradox bias, context dilution from EHR noise
Traps Comorbidity-negated Stage IV exceptions, confounder cohorts, treatment-arm skew, near-boundary windows
Step budget 75 (tight for 29 batches + investigations + flags)

Why ClinicalBench Is Hard

This benchmark is designed to expose fundamental limitations in current AI systems:

Challenge Why It Breaks Naive Agents
Dynamic protocols Rules embedded in natural language change every episode — hardcoded thresholds fail
Multi-hop comorbidity override Stage IV exception is revoked when comorbidity_index > threshold — requires 3-step cross-referencing (stage → comorbidity → window) that LLMs almost always miss
Clinical noise columns 11 realistic EHR fields (BMI, LDH, medications, etc.) dilute LLM attention across 720 × 25+ field records
Simpson's Paradox High-risk sites inflate mortality for minorities, but the cause is disease severity, not sampling bias — overall stats look fine
Tight step budget 75 steps for 40+ errors in 720 patients — agents must triage across 29 batches and cannot check everything
Phased workflow Flagging before investigating is blocked and penalized — forces structured reasoning
Overconfidence penalty High-confidence wrong flags are penalized 1.8× — discourages guessing

Benchmark Results

All scores are from genuine LLM inference — the model reads raw patient data, decides what to flag, and gets scored by the environment. No Python detectors, no hardcoded logic. The LLM is the brain; Python is just the hands.

Reproducible benchmark scores (seed=20260402):

Agent Easy Medium Hard Average Precision Description
🔴 Naive LLM 0.19 0.16 0.02 0.12 10% Single prompt, tiny sample, zero feedback
🟡 Heuristic 0.98 0.79 0.73 0.83 67% Deterministic Python rules (honestly labeled, no LLM)
🟠 ReAct (3.1 405B) 0.77 0.38 0.34 0.50 26% Massive parameters lead to false-positive hallucinations
🟢 ReAct (3.3 70B) 0.98 0.60 0.40 0.66 45% Specialized tool-calling efficiently avoids logic traps

🧠 The Generational Leap: Why 3.3 70B beats 3.1 405B

When forced to play the game fairly, the 405-billion parameter frontier model scored just 0.50, while the newer, smaller Llama 3.3 70B scored 0.66. ClinicalBench successfully exposed the exact architectural difference between the two generations:

  1. The Overthinking Trap (405B's Flaw): Because 3.1 405B is a massive generalist, it looks at the EHR noise in our Hard task and hallucinates complex, non-existent clinical reasons to flag a patient. Our brutal -0.30 penalty for false positives caused the 405B to destroy its own score.
  2. Agentic Tool Mastery (70B's Advantage): Llama 3.3 was heavily fine-tuned for ReAct logic. It doesn't hallucinate ghosts; it calls the [INV] tool, reads the JSON, flags the exact patients, and stops. It navigates the environment better because it is a better "driver."

What This Proves:

  • Language understanding ≠ clinical reasoning.
  • Bigger is not always better in auditing. Raw parameter size leads to overconfidence and false-positive hallucinations.
  • Meta's 3.3 architecture works. ClinicalBench independently verifies that 3.3's agentic fine-tuning directly translates to safer, more accurate clinical compliance.

🏆 Frontier Model Leaderboard

We challenge all frontier models to beat the benchmark. Submit your scores via PR.

Rank Model Easy Medium Hard Avg Score
1 Meta-Llama-3.3-70B-Instruct 0.98 0.60 0.40 0.66
2 Meta-Llama-3.1-405B-Instruct 0.77 0.38 0.34 0.50
Your model here

Challenge: Can any model beat 0.66 average on genuine ReAct evaluation? The 2-hop comorbidity trap, overconfidence penalty, and Simpson's Paradox remain a stress test for every model we evaluate.

🏗️ ReAct Agent Architecture

┌────────────────────────────────────────────────────────────┐
│                    INFERENCE ENGINE                        │
│  ┌──────────┐  ┌──────────────┐  ┌──────────────────────┐  │
│  │ Phase 1  │  │ Phase 2      │  │ Phase 3              │  │
│  │ INVEST.  │→ │ BATCHED SCAN │→ │ REPORT               │  │
│  │ 1 LLM call│ │ 25 pts/batch │  │ 1 LLM call           │  │
│  │ ~500 tok │  │ ~2K tok each │  │ ~500 tok             │  │
│  │          │  │ MEMORY WIPE ↻│  │                      │  │
│  └──────────┘  └──────────────┘  └──────────────────────┘  │
│                                                            │
│  Token Budget: ~2K per call (fits 8K context window)       │
│  Memory Policy: FRESH context each batch (no snowball)     │
│  Error Budget: -0.30 per false positive, 1.8x overconf     │
└────────────────────────────────────────────────────────────┘
          ↕ JSON actions (investigate/flag/report)
┌────────────────────────────────────────────────────────────┐
│            OPENENV ENVIRONMENT (Grading)                   │
│  Procedural Generation → Phase Gate → Scoring → Feedback   │
└────────────────────────────────────────────────────────────┘

Action Space

class AuditAction(Action):
    action_type: str           # investigate_pattern | compute_distribution |
                                # flag_error | propose_fix | submit_report
    variable: Optional[str]     # Field to investigate or compute
    patient_id: Optional[str]   # Patient to flag
    error_type: Optional[str]   # invalid_age | temporal_inconsistency |
                                # protocol_window_violation | selection_bias
    reason: Optional[str]       # Justification text
    proposed_value: Optional[str]
    report: Optional[str]       # Final audit report
    confidence: Optional[float] # 0.0-1.0 confidence in the flag

Observation Space

class AuditObservation(Observation):
    done: bool                          # Episode finished?
    reward: float                       # Dense step reward
    task_id: str                        # task_easy | task_medium | task_hard
    task_type: str                      # Audit category
    task_description: str               # Task instructions
    protocol_title: str                 # Episode protocol ID
    trial_protocol_excerpt: str         # Natural language protocol rules
    dataset: list[dict]                 # Full patient records
    errors_found: list[str]             # Correctly flagged patients
    patterns_investigated: list[str]    # Variables investigated
    distributions_computed: list[str]   # Distributions computed
    feedback: str                       # Step-by-step feedback
    score_so_far: float                 # Current benchmark score [0, 1]
    dense_reward_total: float           # Cumulative dense reward
    score_breakdown: dict[str, float]   # {recall, precision, workflow, efficiency, report}
    attempts_remaining: int             # Steps left in budget
    phase: str                          # investigation | flagging

Reward Design

ClinicalBench uses two scoring layers to separate RL training signal from benchmark evaluation:

Dense Step Reward (for RL training)

  • Correct flag: +0.16
  • False positive: −0.26 (asymmetric to penalize guessing)
  • Duplicate flag: −0.08
  • New investigation: +0.04
  • Overconfident wrong flag: reward × −1.8
  • Per-step cost: −0.004 × step_count (increasing pressure)

Episode Benchmark Score (for evaluation)

Component Weight Signal
Recall 70% What fraction of real errors were caught?
Precision 15% How many flags were correct?
Workflow Discipline 5% Did the agent investigate before flagging?
Efficiency 5% Ratio of useful actions to total actions
Report Quality 5% Does the report cite protocol, root cause, risk, corrective action, fairness?

This separation keeps the RL signal dense (partial progress on every step) while preventing early score saturation from hiding later mistakes.


Procedural Generation

Each episode generates a unique dataset with new protocol constraints:

python3 server/dataset_generator.py

Guarantees:

  • Same seed → identical dataset, protocol, and ground truth
  • Different seeds → different protocols with different rules
  • Deterministic grading: reproducible scores across machines
  • Hard mode alternates between true_bias and confounded_no_bias

Example validated profile (seed=42):

  • Easy: 300 patients, 8 errors, 13 traps
  • Medium: 480 patients, 23 errors, 25 traps
  • Hard: 720 patients, 43 errors, 40 traps (incl. 10 comorbidity override traps)

Quick Start

1. Start the Server

cd server
PYTHONPATH=.. python3 -m uvicorn app:app --host 0.0.0.0 --port 8000

2. Open the Dashboard

Navigate to http://localhost:8000 to see the enterprise audit command center. Select an agent and task, then click Start Audit to watch the reasoning loop in real time.

3. Health Check

curl -s http://localhost:8000/health

4. Run Baseline Inference

# Full comparison (all 3 agents × all 3 tasks)
ENV_BASE_URL=inprocess python3 inference.py --mode all --seed 20260402

# Single agent mode
python3 inference.py --mode full

5. OpenEnv Validation

openenv validate .

Docker

docker build -t clinical-bench:latest .
docker run -p 8000:8000 clinical-bench:latest

The container exposes:

  • /health for health checks
  • / for the enterprise dashboard
  • WebSocket endpoints for OpenEnv reset() / step() / state()

Real-World Relevance

ClinicalBench models tasks that clinical data managers perform daily:

Real-World Task ClinicalBench Equivalent
ICH-E6(R2) protocol compliance review Age eligibility + treatment window verification
FDA 21 CFR Part 11 data integrity audit Temporal consistency checking
DSMB safety signal assessment Stage-adjusted outcome disparity analysis
IRB equity review Confounder-aware selection bias detection

This benchmark is immediately useful for evaluating whether an LLM-based agent can be safely deployed in a clinical data management workflow — one of healthcare AI's highest-value, highest-risk applications.


OpenEnv Compliance

  • Typed Action, Observation, State models (Pydantic)
  • reset(seed, task_id) → Observation
  • step(action) → Observation
  • state → current state
  • openenv.yaml with metadata and 3 tasks
  • openenv validate . passes
  • 3 tasks with deterministic graders, scores in [0.0, 1.0]
  • Dense reward shaping + benchmark rubric
  • Reproducible inference.py at repo root
  • Dockerized with health check
  • Inference runtime < 3 minutes
  • Runs on 2 vCPU / 8GB memory

Project Structure

clinical_trial_auditor/
├── openenv.yaml              # OpenEnv manifest with 3 tasks
├── inference.py              # Baseline inference (naive/heuristic/full)
├── client.py                 # EnvClient implementation
├── models.py                 # Typed Action/Observation/State
├── README.md
├── Dockerfile
├── requirements.txt
├── pyproject.toml
├── docs/
│   └── architecture.md       # Detailed system architecture
└── server/
    ├── app.py                # FastAPI + dashboard API
    ├── clinical_trial_auditor_environment.py
    ├── dataset_generator.py  # Procedural adversarial data engine
    ├── models.py
    ├── requirements.txt
    └── static/
        └── index.html        # Enterprise audit dashboard

Built for the Meta × Scaler School of Technology OpenEnv Hackathon

🧬 Developer Note & Lineage

ClinicalBench is deeply informed by my ongoing research and architecture development on a SEER (Surveillance, Epidemiology, and End Results) based oncology project, active since 2024. The complexities modeled in this benchmark—specifically the Simpson's Paradox confounders, Stage IV comorbidity overrides, and the immense noise of real-world Electronic Health Records—are direct reflections of the challenges encountered when processing live clinical oncology data.

Because the hardest thing about AI in healthcare isn't the model — it's knowing when to trust it.
Sumit Saraswat | GLA University