Ideation Arena - Specificity: leaderboard
Ideation Arena (arXiv 2608.29696v1, Table 2): research proposals from 14 LLMs and five research-agent workflows on GPT-4o and DeepSeek V3.2, all given the same closed citation context, compared in 6,441 double-blind pairwise votes by 105 computer-science researchers. Bradley-Terry Elo for concrete technical detail; higher is better.
Metric: Elo (Bradley-Terry over expert pairwise votes). Source: arxiv.org. Saturation forecast: Not forecast. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 1275.3 | #145 |
| 2 | Claude Opus 4.5 | 1240.3 | #87 |
| 3 | Claude Sonnet 4.5 | 1214.7 | #148 |
| 4 | Kimi K2 (Thinking) | 1118.6 | #252 (Kimi K2) |
| 5 | Mistral Large 3 | 1106.3 | #410 |
| 6 | DeepSeek V3.2 | 1071.3 | #210 |
| 7 | GLM-4.7 | 1059.9 | #202 |
| 8 | O3 | 1059.6 | #132 |
| 9 | Gemini 3 Flash (Preview) | 1055.3 | #90 |
| 10 | Grok 4 Fast | 1049.8 | #266 |
| 11 | Gemini 3 Pro (Preview) | 1031.5 | #71 |
| 12 | Grok 4 | 962.8 | #189 |
| 13 | Qwen 3 Max | 950.2 | #213 |
| 14 | GPT-4o | 836.2 | #357 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ideation-arena-specificity · How It Works · Data refreshed daily, snapshot 2026-11-05.