List your company

Compositional Reasoning on D4 V2 (test)

89.9Stability (%)Qwen2.5-7B-Inst
78.35681.35384.3587.347May 26, 2026

Evaluation results

MethodLinks
Qwen2.5-7B-InstPost-training recipe=RLHF, Inference temperature=greedy (T=0), Response structure=XML-structured, Consistency matcher=Press et al.2026.0589.91269.878.981.1
DeepHermes-3Post-training recipe=SFT distill, Inference temperature=greedy (T=0), Response structure=XML-structured, Consistency matcher=Press et al.2026.058847.461.973.1100
Qwen3-8BPost-training recipe=RLVR, Inference temperature=greedy (T=0), Response structure=XML-structured, Consistency matcher=Press et al.2026.0587.7055.378.9100
Mistral-7B-InstPost-training recipe=RLHF, Inference temperature=greedy (T=0), Response structure=XML-structured, Consistency matcher=Press et al.2026.0578.82555.666.7—