T2J-Bench - DPO Codebases: leaderboard
The 15 direct preference optimization trainers of T2J-Bench (Song et al., 2026), converted from PyTorch to JAX by an agent in the authors' shared coding scaffold; the fixed verifier also compares policy and reference log-probabilities and the preference loss. Percent of codebases whose first conversion passes all three verifier stages; higher is better.
Metric: Conversion Pass Rate (%; first attempt, 15 DPO codebases). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 40 | #75 |
| 2 | Gemini 3.1 Pro (Preview) | 26.7 | #91 |
| 3 | Claude Sonnet 4.6 | 26.7 | #158 |
| 4 | Claude Opus 4.6 | 13.3 | #100 |
| 5 | GLM-5 | 6.7 | #256 |
| 6 | Qwen 3 Coder 480B A35B Instruct | 0 | #641 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=t2j-bench-dpo-codebases · How It Works · Data refreshed daily, snapshot 2026-11-12.