AI Breakthrough Cuts Training Costs by Two-Thirds While Enabling Million-Token Reasoning Capabilities

Researchers at Mila unveil 'Markovian Thinking' technique that slashes AI training costs by over two-thirds while enabling models to reason with up to 140,000 tokens and potentially scale to million-token capabilities through revolutionary linear compute scaling instead of traditional quadratic methods.

• Researchers at Mila develop 'Markovian Thinking' technique that allows AI models to perform complex reasoning with linear compute costs instead of quadratic scaling, cutting training costs by over two-thirds
• The Delethink environment breaks reasoning into fixed-size chunks of 8,000 tokens, forcing models to learn textual summaries between chunks while maintaining constant memory requirements
• Testing shows models trained with this method can reason effectively up to 140,000 tokens and continue improving beyond their training limits, opening possibilities for million-token AI reasoning capabilities