HPR-Bench - Profile Summarization - Tag Accuracy: leaderboard
Metric: Extracted-Tag Accuracy (%; AccEx, tags extracted from the summary by Qwen3-32B). Source: arxiv.org. Saturation forecast: Around December 2027. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3 235B A22B 2507 (Thinking) | 70.14 |
| 2 | GLM-5 (Thinking) | 67.75 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 64.23 |
| 4 | Kimi K2.5 (Thinking) | 63.56 |
| 5 | DeepSeek V3.2 (Thinking) | 63.04 |
| 6 | DeepSeek R1 0528 | 62.55 |
| 7 | Qwen 3.6 Plus (Thinking) | 57.13 |
| 8 | Qwen 3 8B (Thinking) | 51.92 |
| 9 | Qwen 3 8B (Non-reasoning) | 51.6 |
Interactive version: theaggregate.ai/benchmark?slug=hpr-bench-profile-summarization-tag-accuracy · How It Works · Data refreshed daily, snapshot 2026-09-25.