IdeaTaste - Code generation - frontier - Match-Cite: leaderboard
IdeaTaste Table 2:300 proposed ideas per area/prompt setting, matched to held-out 2026 research papers with GPT-5.6 Luna using published citation and idea-overlap rubric. Four prompt-information conditions separate. Six actual agent models listed in AppendixD; native percent overlap retained, no claim that citation imitation is novelty. Human comparison and diversity concentration diagnostics are not model observations. Reasoning controls not reported. Metric: Reference overlap (%).
Metric: Reference overlap (%). Source: arxiv.org. Saturation forecast: Around 2030. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B | 5.05 | #123 |
| 2 | DeepSeek V4 Flash | 4.58 | #155 |
| 3 | K2 Horizon 375B A23B | 4.24 | #165 |
| 4 | Gemini 3.5 Flash Lite | 4.23 | #174 |
| 5 | Claude Haiku 4.5 | 3.9 | #287 |
| 6 | Llama 4 Scout | 3.14 | #648 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ideataste-code-generation-frontier-match-cite · How It Works · Data refreshed daily, snapshot 2026-11-05.