reasoning-degeneration-dev/

t1-distill-aime-397b-to-30b-thinking-heldout-enhanced

Part of the Strategy Distillation experiment exploring whether synthesized reasoning strategies from large models can boost smaller models on AIME math.

pass@1: 60% (12/20) 20 heldout AIME questions (unseen by trace model), with distilled strategy. Thinking enabled.

	Full Experiment Results




	

Condition pass@1

Instruct base (original, n=10) 50% (5/10)

Instruct + strategy (original, n=10) 30%… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/t1-distill-aime-397b-to-30b-thinking-heldout-enhanced.