reasoning-degeneration-dev/

t1-strategy-arena-frozenlake-backward_chaining-together_ai-qwen-qwen3-next-80b-a3b-thin-fea1b895

Strategy compliance baseline — FrozenLake arena evaluation. No strategy instruction was given. This is the baseline to observe the model's natural reasoning patterns across game turns. Per-turn reasoning is extracted from transcripts for manual review.

	Results




	

Metric Value

Win rate 9/10 (90.0%)

Avg turns 30.7

Errors 0

	Details




	

Parameter… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/t1-strategy-arena-frozenlake-backward_chaining-together_ai-qwen-qwen3-next-80b-a3b-thin-fea1b895.