Share your thoughts, 1 month free Claude Pro on usSee more

Correlation Metrics on MT-Bench (LLM-as-a-Judge)

0.689Pearson's r

Qwen3-32B REAL (ours)

Updated 1mo ago

Evaluation Results

Method
Qwen3-32B REAL (ours) 2026.03	0.689	0.691	0.552
TRACT 2025.03	0.672	0.639	-
Baseline B.4 (RAFT) 2025.03	0.618	0.614	-
Qwen3-8B REAL (ours) 2026.03	0.617	0.608	0.471
Qwen3-32B RAFT 2026.03	0.611	0.596	0.439
Mistral2-7B REAL (ours) 2026.03	0.593	0.569	0.422
Qwen3-8B TRACT 2026.03	0.558	0.586	0.436
TRACT 2025.03	0.555	0.529	-
Baseline B.3 (RAIL) 2025.03	0.547	0.583	-
Baseline B.1 (CE) 2025.03	0.541	0.556	-
Qwen3-8B RAFT 2026.03	0.541	0.517	0.385
Prometheus-2-7B 2026.03	0.538	0.511	0.389
Mistral2-7B Standard RL 2026.03	0.529	0.507	0.371
Mistral2-7B TRACT 2026.03	0.521	0.501	0.366
Prometheus-2-7B 2025.03	0.519	0.483	-
Prometheus-2-7B 2026.03	0.519	0.483	0.392
TRACT (Ablation: Objective CE) 2025.03	0.517	0.503	-
CLoud (Reward model) 2025.03	0.511	0.506	-
Qwen3-8B Standard RL 2026.03	0.495	0.535	0.397
Mistral-7B-Instruct (RAFT on GPT-4) 2025.03	0.483	0.469	-
Mistral-7B-Instruct (CE on GPT-4 CoT) 2025.03	0.48	0.482	-
Prometheus-1-13B 2026.03	0.467	0.455	0.345
Baseline B.2 (CE) 2025.03	0.466	0.494	-
Mistral-7B-Instruct (CE on Self-gen CoT) 2025.03	0.435	0.426	-
TRACT (Ablation: Stage 1 Init) 2025.03	0.432	0.421	-
Qwen3-32B Base 2026.03	0.425	0.468	0.353
GPT-3.5-Turbo-0613 2026.03	0.422	0.371	0.299
TRACT (Ablation: Data GPT-4) 2025.03	0.399	0.418	-
Mistral2-7B RAFT 2026.03	0.399	0.418	0.307
Prometheus-1-7B 2026.03	0.367	0.371	0.285
Qwen3-8B Base 2026.03	0.359	0.355	0.27
Qwen3-8B Base 2026.03	0.355	0.327	0.265
Mistral2-7B Base (w/ warmup) 2026.03	0.32	0.299	0.214
Mistral-7B-Instruct (Zero-shot RAIL) 2025.03	0.309	0.216	-
Mistral2-7B Base (w/ warmup) 2026.03	0.309	0.318	0.253
Mistral-7B-Instruct (CE on GPT-4 scores) 2025.03	0.279	0.268	-