ReFigBench (Direct Code Generation) - Human Preference Elo: leaderboard
Blinded human preferences on ReFigBench (Fan et al., 2026) for slides rebuilt with the direct code workflow: annotators compare two renderings of the same source figure, and a Bradley-Terry model over all ten evaluated configurations gives Elo strengths with mean 1,000. Higher is better, and each row is one model run inside one agent harness.
Metric: Human preference Elo (direct workflow: the agent writes a program that builds the slide with a general PPTX library (python-pptx); Bradley-Terry strength on the Elo scale (mean 1,000 over the ten configurations) from 4,270 blinded pairwise human comparisons of renderings of the same source figure, a tie counting half a win). Source: arxiv.org. Saturation forecast: Around June 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 1281.6 | #48 (GPT-5.5) |
| 2 | Claude Opus 4.6 | 885.5 | #100 |
| 3 | MiniMax-M3 | 816.7 | #265 |
| 4 | MiMo-V2.5 | 738 | #353 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=refigbench-direct-code-generation-human-preference-elo · How It Works · Data refreshed daily, snapshot 2026-11-12.