GISTBench (Amazon Music): leaderboard
Metric: Median per-user interest-groundedness F1 (%): the harmonic mean of IG precision (share of the model's predicted interest categories whose cited engagement evidence, after a Llama-3.3-70B-Instruct relevance filter, meets the positive and negative signal thresholds) and IG recall (share of the categories verified for that user by any of the 11 evaluated models, a pooled oracle), with interests mapped to a 325-category taxonomy, on Amazon Music review users (1-5 star ratings, sparse histories); the model reads the user's interaction history and returns interests with cited evidence; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 66.7 | #91 |
| 2 | GPT-OSS-120B | 66.7 | #330 |
| 3 | Gemini 3.5 Flash | 66.7 | #55 |
| 4 | DeepSeek R1 | 57.1 | #245 |
| 5 | Qwen 3 32B | 50 | #424 |
| 6 | Qwen 3 VL 235B A22B Instruct | 50 | #264 |
| 7 | Qwen 2.5 72B Instruct | 43.3 | #436 |
| 8 | Llama 3.1 8B Instruct | 37.3 | #1018 |
| 9 | DeepSeek R1 Distill Llama 70B | 33.3 | #582 |
| 10 | Qwen 2 7B Instruct | 28.6 | #976 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=gistbench-amazon-music · How It Works · Data refreshed daily, snapshot 2026-10-11.