HarnessOpt-Bench - OfficeQA: leaderboard
A task of HarnessOpt-Bench (Scale AI, 2026): an optimizer LLM working through a coding harness edits the deliberately untuned seed harness of an OfficeQA Pro document question-answering agent under fixed evaluation budgets, never seeing the held-out test cases, and nominates a final harness. Normalized gain is the fraction of the headroom above the held-out score of the seed agent that the nominated harness captures, negative when it is worse than the seed; higher is better.
Metric: Normalized gain ((held-out score minus seed score) / (1 minus seed score); negative when the nominated harness is worse than the seed; mean over the contestant's optimization rounds, each held-out case scored three times; OfficeQA Pro document question answering with a deepseek-v4-flash target agent; seed held-out score 0.341, 99 test cases; the optimizer edits the seed harness under fixed development and validation evaluation budgets and never sees the test partition). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 0.63 | #22 |
| 2 | Claude Sonnet 5 | 0.51 | #111 |
| 3 | Kimi K3 | 0.41 | #54 |
| 4 | GPT-5.6 Sol | 0.29 | #30 |
| 5 | GPT-5.6 Terra | 0.14 | #62 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harnessopt-bench-officeqa · How It Works · Data refreshed daily, snapshot 2026-11-13.