Cangjie-bench - Code Generation (ILA-agent): leaderboard
Cangjie-bench (Shen et al., 2026) tests coding in Cangjie, a statically typed language Huawei released in 2024 with little public code. This board: 155 HumanEval problems solved in Cangjie under the ILA-agent protocol, where the model searches and reads the official Cangjie documentation and compiles and tests its code through tools for up to 15 turns; accuracy is the share of solutions passing all private unit tests, in percent (higher is better).
Metric: Accuracy (%; share of solutions passing all private unit tests; 155 HumanEval problems written in the Cangjie language; ILA-agent: the model explores the official Cangjie documentation and verifies code with the compiler through tools, at most 15 turns). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 81.94 | #244 |
| 2 | Qwen 3 Max | 73.55 | #401 |
| 3 | DeepSeek V3.2 | 63.23 | #403 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cangjie-bench-code-generation-ila-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.