Research2h ago
Controlling Reasoning Effort in LLMs
New research identifies how LLMs toggle between low, medium, and high-effort reasoning modes. The study demonstrates that models can be steered to allocate more compute to complex problems while bypassing overkill for simple tasks. This efficiency gain allows developers to optimize inference costs without sacrificing accuracy on difficult logic puzzles.