Vibe-Bench - BM25 Retrieval - Response BLEU-4: leaderboard
Vibe-Bench tests personalization from user histories: selecting an emotion-regulation strategy, generating a personalized response, and predicting occupational preferences. The paper separates no-history, full-history, profile, retrieval and personalized fine-tuning protocols. P-SFT checkpoints retain distinct identities from their untrained backbones. This series reports Vibe-Bench - BM25 Retrieval - Response BLEU-4 from 2609.00921v1, A11.T10: Score (0-100). Higher is better.
Metric: Score (0-100). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 Instruct | 4.56 | #643 |
| 2 | Gemma 3 27B (IT) | 3.67 | #1375 |
| 3 | Mistral Small 3.2 | 2.98 | #1673 |
| 4 | Qwen 3 14B | 2.77 | #1309 |
| 5 | Llama 3.1 70B Instruct | 2.55 | #1629 |
| 6 | Llama 3.1 8B Instruct | 2.34 | #4318 |
No result here: #4 Claude Opus 5.5, #5 GPT-6 Astra, #9 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vibe-bench-bm25-retrieval-response-bleu-4 · How It Works · Data refreshed daily, snapshot 2026-11-08.