HPR-Bench - Profile Summarization - Tag Coverage: leaderboard
Profile summarization in HPR-Bench (UserGPT Technical Report, Xuan et al., 2026): the model condenses a user's long behavioural history into a free-text persona summary, and Qwen3-32B extracts the atomic profile tags the summary states. The score is the coverage ratio of those extracted tags over the reference profile, in percent; higher is better.
Metric: Extracted-Tag Coverage (%; COVEx, share of profile tags recovered from the summary). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 (Thinking) | 96.38 | #460 |
| 2 | Qwen 3 235B A22B 2507 Instruct | 94.71 | #619 |
| 3 | GLM-5 (Thinking) | 94.22 | #256 (GLM-5) |
| 4 | DeepSeek V3.2 (Thinking) | 93.41 | #403 (DeepSeek V3.2) |
| 5 | Kimi K2.5 (Thinking) | 93.32 | #252 (Kimi K2.5) |
| 6 | DeepSeek R1 0528 | 92.21 | #425 |
| 7 | Qwen 3.6 Plus (Thinking) | 90.46 | #217 (Qwen 3.6 Plus) |
| 8 | Qwen 3 8B (Non-reasoning) | 88.21 | #1838 (Qwen 3 8B) |
| 9 | Qwen 3 8B (Thinking) | 88.01 | #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-profile-summarization-tag-coverage · How It Works · Data refreshed daily, snapshot 2026-11-12.