Figma to code, expert preferencen = 33 tasksEloReported

1GPT-5.2Codex · effort not reported1054
1Claude Opus 4.6Claude Code · effort not reported1039
3Gemini 3.1 ProGemini CLI · effort not reported907

Bradley-Terry Elo from 135 pairwise votes by a domain expert on the built pages, as the paper reports it (Table 10). 95% intervals.

Programmatic content, expert preferencen = 41 tasksEloReported

1Claude Opus 4.6Claude Code · effort not reported1173
2GPT-5.2Codex · effort not reported915
2Gemini 3.1 ProGemini CLI · effort not reported912

3 agentsSource

Bradley-Terry Elo fit to the 279 released pairwise expert votes on rendered chapters, a tie counted as half a win, centred on 1000; 95% bootstrap intervals over votes. The paper reports these votes as win rates.

Figma to coden = 33 tasksOutput scoreReported

1GPT-5.2 ProCodex · effort not reported4.27
2Claude Opus 4.6Claude Code · effort not reported4.19
3GPT-5.2Codex · effort not reported3.94
4Claude Opus 4.5Claude Code · effort not reported3.88
5Gemini 3.1 ProGemini CLI · effort not reported3.73
6Claude Sonnet 4.5Claude Code · effort not reported3.66
7Gemini 3 ProGemini CLI · effort not reported3.59

7 agentsSource

A visual judge scores the built page against the design, 1 to 5, on eight weighted rubrics. 95% bootstrap intervals over tasks where the run count allows.

Programmatic contentn = 41 tasksArtifact scoreReported

1Claude Opus 4.6Claude Code · effort not reported0.612
1Gemini 3.1 ProGemini CLI · effort not reported0.526
2GPT-5.2Codex · effort not reported0.478

3 agentsSource

A visual judge scores each rendered chapter against rubrics written by educators, normalised to 0 to 1. Course-level means over 41 courses and 183 chapters, 95% bootstrap intervals.

01The benchmark

Subjective enterprise work, scored three ways: expert rubrics on the process, curated artifacts on the output, pairwise human preference on the ranking.

Figma to code, 33 tasks. Real Figma designs built into React and Tailwind, scored by a visual judge against the design and by expert preference.
Programmatic content, 41 tasks. Courses from source documents rendered as video, scored on citation grounding, educator rubrics and expert preference.
Financial documents, 37 tasks. Business bank statements read into labelled transactions, scored on extraction and categorisation against risk analysts' labels.
Figma design, 10 tasks. Design work inside a headless Figma on one design system, checked for completeness and consistency.

02Data

Figma to code tasks (open). The Figma designs the agents build from, each with its file key and description.
Figma to code preferences (open). Pairwise expert votes on the built pages, each with its reasoning and a quality rating.
Programmatic content tasks (open). Courses with their source documents, line-level citation ground truth and educator rubrics.
Programmatic content preferences (open). Pairwise expert votes on the rendered chapters, each with its reasoning.
Financial documents preview (open). A preview of the bank statement cases with the transaction taxonomy and graded labels.
Financial documents (on request). The redacted bank statement cases with extraction and categorisation ground truth.

Cite

@article{chandwani2026lhbench,
  title   = {LH-Bench: Skill-Grounded Evaluation of Long-Horizon Agents on Subjective Enterprise Tasks},
  author  = {Chandwani, Abhishek and Gupta, Ishan},
  journal = {arXiv preprint arXiv:2603.22744},
  year    = {2026}
}

Paper: LH-Bench: Skill-Grounded Evaluation of Long-Horizon Agents on Subjective Enterprise Tasks