llm-post-training on Hacker News: 1 stories with screenshots

Screenshot of arxiv.org: VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO

VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO (arxiv.org)

357 points by timhigins 4 months ago | 185 comments

arXiv technical report introducing VibeThinker-3B, a 3B-parameter dense model trained with curriculum SFT, multi-domain RL, and self-distillation for verifiable reasoning.

small-language-models reasoning reinforcement-learning vibethinker

#llm-post-training on Hacker News

VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO(arxiv.org)

VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO (arxiv.org)