AspectSim (Summarize-Then-Embed): leaderboard
Aspect-conditioned document similarity (AspectSim, Hossain et al., 2026) with the LLM writing an aspect-focused summary of each of two related documents before an embedding model compares them. The score is the Spearman correlation with GPT-4o similarity labels over about 26,000 aspect pairs from five corpora, averaged over nine embedding models, on a -1 to 1 scale; higher is better.
Metric: Spearman correlation (-1 to 1) between the embedding similarity of the evidence the LLM extracts from each document and the GPT-4o aspect-similarity labels, mean over nine embedding models; the LLM writes an aspect-conditioned summary. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B Instruct | 0.57 | #1093 |
| 2 | Phi-4 | 0.57 | #1920 |
| 3 | Qwen 2.5 14B Instruct | 0.56 | #1745 |
| 4 | Qwen 2.5 32B Instruct | 0.56 | #1216 |
| 5 | Qwen 3 8B | 0.56 | #1838 |
| 6 | DeepSeek R1 Distill Qwen 32B | 0.55 | #1780 |
| 7 | Llama 3.3 70B Instruct | 0.54 | #1353 |
| 8 | Gemma 3 12B (IT) | 0.53 | #1849 |
| 9 | Qwen 3 4B | 0.52 | #2297 |
| 10 | Gemma 2 27B (IT) | 0.52 | #1815 |
| 11 | DeepSeek R1 Distill Qwen 14B | 0.48 | #2594 |
| 12 | Mistral 7B Instruct | 0.47 | #7693 |
| 13 | Llama 3.2 3B Instruct | 0.4 | #7185 |
| 14 | Phi-4 Mini Instruct | 0.36 | #6452 |
| 15 | Gemma 3 1B (IT) | 0.06 | #8752 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aspectsim-summarize-then-embed · How It Works · Data refreshed daily, snapshot 2026-11-12.