Controlling Reasoning Effort in LLMs
Summary
Sebastian Raschka's Deep-dive analyzes how reasoning effort can be controlled in large language models, detailing training-time and inference-time scaling, different reasoning modes, and how various open-weight models implement on/off switches and token budgets. The piece covers practical mechanisms (RLVR, SFT, prompts, and budgets), examples from multiple models, and the potential for automatic effort selection in future systems.