SCDBench - Format Completeness: leaderboard
SCDBench (Qin, Song and Gervais, 2026) tests LLM decompilation of 600 verified Ethereum smart contracts in three difficulty tiers: from EVM assembly and function-selector hints the model must write Solidity that compiles and behaves like the original. Percentage of the 600 responses that contain parseable compiler settings, self-contained code and a target contract name, zero-shot; higher is better.
Metric: Complete responses (%; parseable compiler settings, self-contained code and a target contract name, of 600 contracts, zero-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.3 Codex (High) | 94 | #114 (GPT-5.3 Codex) |
| 2 | GLM-5 (Non-reasoning) | 90 | #256 (GLM-5) |
| 3 | GLM-5 (Thinking) | 86.67 | #256 (GLM-5) |
| 4 | Claude Opus 4.7 (High) | 72.5 | #75 (Claude Opus 4.7) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scdbench-format-completeness · How It Works · Data refreshed daily, snapshot 2026-11-12.