Hasty Briefsbeta

Bilingual

Controlling Reasoning Effort in LLMs

2 days ago
  • Introduction of reasoning models with multiple effort modes similar to GPT-5.6.
  • Definition of reasoning models as generating intermediate reasoning traces.
  • Training methods: RLVR (Reinforcement Learning with Verifiable Rewards) and supervised fine-tuning.
  • Inference scaling: Adjusting compute and token usage to improve performance.
  • Implementation of reasoning effort levels using system prompts and length penalties.
  • Overview of effort modes in open-weight models like DeepSeek V4, Nemotron, Kimi, GLM-5, Qwen3, and Inkling.
  • Automatic effort selection as a future goal, though currently user-controlled via system prompts.