QC-Stark Benchmark

QC-Stark is a multi-task benchmark for evaluating large language models on quantum computing (QC) tasks. It covers the full practitioner workflow—circuit construction, debugging, compilation, verification, simulation, and error correction—with systematic difficulty scaling and contamination resistance via procedural generation.

Key facts

Property	Value
Tasks	11
Difficulty levels per task	5 (Textbook → Homework → Exam → Research → Open)
Seeds per (task, level)	5
Total instances	275
Auto-verification	Yes (Qiskit execution)
Contamination resistance	Seed-based procedural generation

Task list

Task ID	Name	Workflow Stage	Mean accuracy (10 models)
T1	State Preparation	Construct	0.440
T2	Trotterization	Construct	0.252
T3	Oracle Synthesis	Construct	0.412
T4	Debugging	Understand	0.020
T5	Noise Discrimination	Understand	0.272
T6	Reverse Engineering	Understand	0.448
T7	Equivalence Checking	Verify	0.824
T8	Hardware Routing	Compile	0.184
T9	Noise Fidelity	Simulate	0.516
T10	VQE	Simulate	0.592
T11	QEC Decoding	Error Correction	0.440

Schema

Each instance has the following fields:

• task_id (string): T1–T11
• task_code (string): internal code (e.g. B1_debugging)
• task_name (string): human-readable task name
• workflow_stage (string): QC workflow stage
• level (int): difficulty level 1–5
• difficulty (string): Textbook / Homework / Exam / Research / Open
• seed (int): procedural generation seed 1–5
• prompt (string): the full problem statement given to the model

Evaluation

Responses are verified by executing model-generated Qiskit code against auto-generated test cases. No human grading. Verification metrics vary by task (state fidelity, functional equivalence, syndrome correctness, etc.).

Citation

@misc{gupta2026qcstark,
  title={QC-Stark: A Multi-Task Benchmark Revealing Capability Dissociations
         in LLMs for Quantum Computing Tasks},
  author={Gupta, Pranav},
  year={2026},
  institution={Cisco Collaboration AI}
}