ReFigBench (Direct Code Generation): leaderboard
ReFigBench (Fan et al., 2026) asks a coding agent to rebuild a scientific overview figure from an arXiv paper as one editable PowerPoint slide; in this workflow the agent writes a program with a general PPTX library. A GPT-5.4 judge scores fidelity and native editability on a 100-point rubric (text, semantic structure, layout, editability and visual detail), averaged over 1,000 figures; higher is better, and each row is one model run inside one agent harness.
Metric: Overall rubric score (0-100) under the direct workflow: the agent writes a program that builds the slide with a general PPTX library (python-pptx); GPT-5.4 judge (gpt-5.4-2026-03-05, low effort, first of three runs) scores the rendered slide and its native object tree against the source figure on text 20, semantic structure 30, layout 15, editability 25 and visual detail 10 points; unrenderable artifacts score 0 and a full-slide raster paste is capped at 50; mean over 1,000 arXiv overview figures. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 75.6 | #48 (GPT-5.5) |
| 2 | Claude Opus 4.6 | 69.4 | #100 |
| 3 | MiniMax-M3 | 66.1 | #265 |
| 4 | MiMo-V2.5 | 63.4 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.