Diverge - Length-controlled Win Rate: leaderboard
Diverge v1 held-out1,740 instances from2026 papers,16domains; three idea cards generated as one set except DARLING three single-card draws. Qwen custom variants separate, proposed checkpoint600; generic Nemotron120B revision unresolved, never inferred as current Nemotron3Super. Impact eval GPT5.4mini both candidate/gold orderings; distance eval Gemini3.5Flash,temp0,max1024 (training uses different judges); LC win rate length-controlled pooled logistic adjustment. CU greedy-order patience0.9 quality-times-distance sum divided by1+0.9+0.9². Grounding/distance are diagnostics, not monotonically better capability. Dataset/code release currently coming soon. Metric: Length-controlled impact win rate (0–1). Higher is better. Source: https://arxiv.org/html/2610.22324v1
Metric: Length-controlled impact win rate (0–1). Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 0.21 | #36 |
| 2 | GPT-6 Astra | 0.15 | #5 |
No result here: #3 Claude Opus 5.5, #7 Gemini 4 Argon, #8 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=diverge-length-controlled-win-rate · How It Works · Data refreshed daily, snapshot 2026-10-17.