RoundTripCodeEval - Run-Length Encoding: leaderboard
RoundTripCodeEval (RTCE; Maveli, Vergari and Cohen, 2026) tests whether models can execute and invert code, here a run-length compressor written in Python, on 250 synthetic inputs: predict the encoder output and the decoder output, directly and by inverting the function they are given. The board is the paper average of exact match, edit similarity and pass@5 over the four tasks, in percent (higher is better); zero-shot with one worked example.
Metric: Mean of exact match, edit similarity and pass@5 (%) over the four round-trip tasks for run-length encoding (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 | Overall rank |
|---|---|---|---|
| 1 | QwQ-32B | 57.23 | #909 |
| 2 | Qwen 2.5 Coder 32B Instruct | 41.51 | #1553 |
| 3 | DeepSeek R1 Distill Qwen 32B | 36.37 | #1780 |
| 4 | Codestral-22B-v0.1 | 30.68 | #2680 |
| 5 | DeepSeek R1 Distill Qwen 14B | 26.97 | #2594 |
| 6 | Qwen 2.5 7B Instruct | 17.39 | #2583 |
| 7 | Phi-3.5-mini-instruct | 13.67 | #5757 |
| 8 | Phi-3 Mini 128K Instruct | 12.01 | #6952 |
| 9 | Yi-Coder-9B-Chat | 11.85 | #6839 |
| 10 | Mistral 7B Instruct (v0.3) | 11.52 | #6698 |
| 11 | Llama 3.1 8B Instruct | 9.36 | #4277 |
| 12 | DeepSeek R1 Distill Qwen 1.5B | 4.3 | #9257 |
| 13 | Llama 3.2 1B Instruct | 0.15 | #9109 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=roundtripcodeeval-run-length-encoding · How It Works · Data refreshed daily, snapshot 2026-11-12.