CAT — MT-Bench — Honest persona + prefill: leaderboard
Appendix A.6 Table 4, method=Honest persona + prefill. Base means unmodified Llama-3.3-70B-Instruct or Qwen3.6-27B with reasoning OFF, not pretrained base. MMLU: all 57 subjects, n=14,042, zero-shot. MT-Bench: 80 two-turn questions, Gemma-4-31B-it judge, 1–10. Prompt/steering conditions remain separate boards; SFT/CAPO/CATCH learned adapters are source-qualified distinct model identities and never merged into the original checkpoint. Their public weight release is unverified. Metric: Judge score (1–10). Higher is better. Source: https://arxiv.org/html/2610.22292v1
Metric: Judge score (1–10). Source: arxiv.org. Saturation forecast: Around January 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 27B (Non-reasoning) | 8.43 | #167 (Qwen 3.6 27B) |
| 2 | Llama 3.3 70B Instruct | 6.95 | #532 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=cat-mt-bench-honest-persona-plus-prefill · How It Works · Data refreshed daily, snapshot 2026-10-17.