HORIZON - LLM Query Reformulation: leaderboard
HORIZON (2026) is a cross-domain, long-horizon user modeling benchmark built from Amazon Reviews; its LLM tasks run zero-shot on a held-out set of users whose histories end at a 2020 cut-off, and the text from the LLM is turned into catalog items by a fixed retrieval encoder. Here the LLM rewrites the history of the user into ten search queries; recall at 100 of the held-out next item, in percent; higher is better.
Metric: Recall@100 (%) of the held-out next item when the LLM rewrites a user's history into ten search queries, items retrieved from the Amazon Reviews catalog with the BLAIR item encoder and a FAISS index, zero-shot, on HORIZON's held-out set of one million out-of-distribution users after the 2020 cut-off; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 3 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 8B | 3.5 | #1873 |
| 2 | Llama 3.1 8B Instruct | 2.84 | #4600 |
| 3 | Gemma 2 9B (IT) | 2.66 | #2369 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=horizon-llm-query-reformulation · How It Works · Data refreshed daily, snapshot 2026-11-21.