VibeThinker-3B records 94.3 AIME26 via curriculum SFT and GRPO on 3B parameters

The arXiv report details a post-training pipeline of curriculum supervised fine-tuning, multi-domain reinforcement learning via GRPO, and offline self-distillation applied to a 3B base. These steps produce verifiable reasoning scores that match or exceed models with 10-100x parameters on AIME26, LiveCodeBench, and recent LeetCode contests while preserving 93.4 on IFEval.

Benchmarks show 97.1 with claim-level test-time scaling and 96.1 percent acceptance on unseen contests. The Parametric Compression-Coverage Hypothesis in the report frames verifiable reasoning as compressible into small cores, distinct from broad knowledge coverage that demands larger parameter counts. This aligns with prior 1.5B results and contrasts with scaling curves observed in DeepSeek V3 and Gemini 3 series.

Operationally, the pipeline demonstrates that targeted reinforcement on verifiable tasks can compress frontier reasoning performance into deployment-efficient models. Subsequent work will likely test whether GRPO variants transfer across additional 1-7B bases without loss of instruction adherence.

The pattern indicates small-model regimes can serve as primary research vehicles for reasoning algorithms rather than mere distillation targets.