Cog-DRIFT: Exploration on Adaptively Reformulated Instances Enables Learning from Hard Reasoning Problems

Cog-DRIFT constructs reformulated variants of tasks and arranges them into an adaptive curriculum by difficulty.

Training proceeds from easier formats to harder formats so models learn progressively.

Task reformulation transforms complex problems into simpler formats while preserving the original answer.

Cog-DRIFT makes hard problems solvable that were previously unsolvable under standard RL post-training and improves pass@k at test time.

Across 2 models and 6 reasoning benchmarks it outperforms standard GRPO and strong guided-exploration baselines—absolute gains: +10.11% (Qwen), +8.64% (Llama); average gains over second-best: +4.72% (Qwen), +3.23% (Llama).