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

Researchers introduce Cog-DRIFT, a new framework that improves AI language model reasoning by transforming difficult problems into easier formats like multiple-choice questions, then gradually training models on increasingly complex versions. The method shows significant performance gains of 8-10% on previously unsolvable problems across multiple reasoning benchmarks.

• →Cog-DRIFT addresses a key limitation in reinforcement learning where models can't learn from problems too difficult to solve under current policies.
• →The framework reformulates hard reasoning problems into simpler formats like multiple-choice and fill-in-the-blank questions while preserving original answers.
• →Training uses adaptive curriculum learning, progressing from easier structured formats to harder open-ended problems.
• →Testing showed absolute improvements of +10.11% for Qwen and +8.64% for Llama models on originally unsolvable problems.
• →The method consistently outperformed standard training approaches across 2 models and 6 reasoning benchmarks with improved sample efficiency.