QEncodeBench: leaderboard
QEncodeBench (Che, Wu, Yuan and Yu, 2026) asks a language model to encode a classical constraint problem as a Grover-style phase oracle in Qiskit, within a qubit budget and without oracle-lookup APIs. A verifier decides by exhaustive simulation whether the circuit marks exactly the solution set up to a global phase and restores its ancillas; the board is semantic pass@1 in percent over the 480-instance core set, higher is better.
Metric: Semantic pass@1 (%; L3 gate on the 480-instance core set of six constraint-problem families at three graded tiers; the model writes a Qiskit build_oracle function in one shot and the verifier decides full solution-set equivalence up to a global phase by exhaustive simulation, with ancillas restored and a measurement-free unitary circuit; pass@1, one attempt per instance (five samples at temperature 0.7 for the non-reasoning DeepSeek-V4-Flash row)). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra | 83.1 | #60 |
| 2 | GPT-OSS-120B (High) | 81.2 | #699 (GPT-OSS-120B) |
| 3 | GPT-OSS-20B (High) | 68.8 | #1150 (GPT-OSS-20B) |
| 4 | Claude Opus 4.8 | 65.2 | #56 |
| 5 | Claude Haiku 4.5 | 49 | #568 |
| 6 | DeepSeek V4 Flash (Non-reasoning) | 10.2 | #254 (DeepSeek V4 Flash) |
| 7 | DeepSeek R1 0528 Qwen3 8B | 1 | #2101 |
| 8 | Qwen 2.5 Coder 7B Instruct | 0 | #3884 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=qencodebench · How It Works · Data refreshed daily, snapshot 2026-11-12.