HPR-Bench - Tag Inference: leaderboard
User-profile inference from long behavioural histories (HPR-Bench, UserGPT Technical Report, Xuan et al., 2026): given one user's multi-source e-commerce and service history, the model answers single- and multiple-choice questions about 18 profile attributes in five dimensions. The score is Avg@10, the accuracy averaged over ten sampled answers and over the 18 attributes, in percent; higher is better.
Metric: Avg@10 Tag Accuracy (%; mean of 10 samples over 18 profile tags). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 Plus (Thinking) | 73.29 | #217 (Qwen 3.6 Plus) |
| 2 | GLM-5 (Thinking) | 71.86 | #256 (GLM-5) |
| 3 | Qwen 3.5 397B A17B (Thinking) | 71.65 | #242 (Qwen 3.5 397B A17B) |
| 4 | Qwen 3.5 Plus (Thinking) | 71.54 | #227 (Qwen 3.5 Plus) |
| 5 | Kimi K2.5 (Thinking) | 70.64 | #252 (Kimi K2.5) |
| 6 | GLM-4.7 (Thinking) | 69.38 | #365 (GLM-4.7) |
| 7 | DeepSeek V3.2 (Thinking) | 66.84 | #403 (DeepSeek V3.2) |
| 8 | DeepSeek R1 0528 | 66.13 | #425 |
| 9 | Kimi K2 (Thinking) | 65.56 | #435 (Kimi K2) |
| 10 | Qwen 3 235B A22B 2507 (Thinking) | 64.34 | #460 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 52.44 | #619 |
| 12 | Qwen 3 8B (Thinking) | 50.35 | #1838 (Qwen 3 8B) |
| 13 | Qwen 3 8B (Non-reasoning) | 46.12 | #1838 (Qwen 3 8B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hpr-bench-tag-inference · How It Works · Data refreshed daily, snapshot 2026-11-12.