RoundTripCodeEval - LZW: leaderboard
Metric: Mean of exact match, edit similarity and pass@5 (%) over the four round-trip tasks for LZW (output prediction and input prediction, each directly and through the inverted function), 250 inputs, zero-shot with one worked example, five completions. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | QwQ-32B | 24.14 |
| 2 | DeepSeek R1 Distill Qwen 32B | 23.81 |
| 3 | Qwen 2.5 Coder 32B Instruct | 21.06 |
| 4 | DeepSeek R1 Distill Qwen 14B | 14.03 |
| 5 | Codestral-22B-v0.1 | 7.77 |
| 6 | Phi-3.5-mini-instruct | 4.81 |
| 7 | Qwen 2.5 7B Instruct | 4.46 |
| 8 | Yi-Coder-9B-Chat | 4.21 |
| 9 | Phi-3 Mini 128K Instruct | 3.65 |
| 10 | Mistral 7B Instruct (v0.3) | 2.25 |
| 11 | Llama 3.1 8B Instruct | 0.82 |
| 12 | DeepSeek R1 Distill Qwen 1.5B | 0.32 |
| 13 | Llama 3.2 1B Instruct | 0.05 |
Interactive version: theaggregate.ai/benchmark?slug=roundtripcodeeval-lzw · How It Works · Data refreshed daily, snapshot 2026-09-26.