AspectSim (Sentence-Level Retrieval): leaderboard
Aspect-conditioned document similarity (AspectSim, Hossain et al., 2026): given two related documents and a free-form aspect, the LLM retrieves the single most relevant sentence from each, an embedding model scores the two sentences, and the score is the Spearman correlation with GPT-4o similarity labels (validated against humans) over about 26,000 aspect pairs from news, medical, political, peer-review and hotel-review 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 retrieves the single most relevant sentence. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 0.59 | #1353 |
| 2 | Qwen 2.5 32B Instruct | 0.59 | #1216 |
| 3 | Qwen 2.5 72B Instruct | 0.59 | #1093 |
| 4 | Qwen 3 8B | 0.58 | #1838 |
| 5 | Phi-4 | 0.58 | #1920 |
| 6 | Qwen 2.5 14B Instruct | 0.58 | #1745 |
| 7 | Gemma 2 27B (IT) | 0.57 | #1815 |
| 8 | Qwen 3 4B | 0.57 | #2297 |
| 9 | DeepSeek R1 Distill Qwen 32B | 0.57 | #1780 |
| 10 | DeepSeek R1 Distill Qwen 14B | 0.57 | #2594 |
| 11 | Gemma 3 12B (IT) | 0.55 | #1849 |
| 12 | Mistral 7B Instruct | 0.52 | #7693 |
| 13 | Phi-4 Mini Instruct | 0.48 | #6452 |
| 14 | Llama 3.2 3B Instruct | 0.46 | #7185 |
| 15 | Llama 3.2 1B Instruct | 0.11 | #9109 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aspectsim-sentence-level-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-12.