Research2h ago
Controlling Reasoning Effort in LLMs
New research demonstrates how LLMs learn distinct low, medium, and high-effort reasoning modes. By adjusting these levels, researchers can optimize the trade-off between compute cost and answer accuracy. This allows developers to trigger deep thinking only for complex queries. It reduces wasted inference cycles on trivial tasks while maintaining high-quality outputs for difficult problems.