# Tiny Model, Big Logic: Diversity-Driven Optimization Elicits Large-Model Reasoning Ability in VibeThinker-1.5B

**Authors:** Sen Xu, Yi Zhou, Wei Wang, Jixin Min, Zhibin Yin, Yingwei Dai, Shixi Liu, Lianyu Pang, Yirong Chen, Junlin Zhang  
**arXiv:** [2511.06221](https://arxiv.org/abs/2511.06221)  
**Date:** 2025-11-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2511.06221v1/

## Key points

- VibeThinker-1.5B is a groundbreaking 1.5B-parameter model challenging the belief that small models have limited reasoning capabilities
- Developed through the Spectrum-to-Signal Principle (SSP) showcasing superior reasoning performance compared to larger models like DeepSeek R1 and Kimi k2
- Utilizes Two-Stage Diversity-Exploring Distillation (SFT) followed by MaxEnt-Guided Policy Optimization (RL) for exceptional reasoning abilities on math benchmarks such as AIME24, AIME25, and HMMT25
- Surpasses DeepSeek R1 in tasks while reducing training costs to $7,800
- Outperforms Magistral Medium and its own base model on LiveCodeBench V6, highlighting prowess in specialized domains and complex reasoning tasks
- Remarkable performance in diverse scenarios when compared against state-of-the-art models with Long-CoT capabilities and non-reasoning models
- Evaluation settings using vLLM for accurate assessments of performance metrics
- "Tiny Model, Big Logic" arXiv query emphasizes large-model reasoning ability elicited through diversity-driven optimization in VibeThinker-1.5B
- Challenges existing Scaling Law assumptions by showing small models can achieve remarkable reasoning capabilities comparable to larger counterparts while reducing training and inference costs

## Summary

Introducing VibeThinker-1.5B: A Revolutionary Model with Exceptional Reasoning Capabilities

VibeThinker-1.5B is a groundbreaking 1.5B-parameter model that challenges the conventional belief that small models are limited in their reasoning capabilities. Developed through our innovative Spectrum-to-Signal Principle (SSP), this compact yet powerful model showcases superior reasoning performance compared to larger models like DeepSeek R1 and Kimi k2. Utilizing a Two-Stage Diversity-Exploring Distillation (SFT) followed by MaxEnt-Guided Policy Optimization (RL), VibeThinker-1.5B demonstrates exceptional reasoning abilities on challenging math benchmarks such as AIME24, AIME25, and HMMT25. In fact, it surpasses even the much larger DeepSeek R1 in these tasks while also significantly reducing training costs to only $7,800. Furthermore, on LiveCodeBench V6, VibeThinker-1.5B outperforms both Magistral Medium and its own base model by a significant margin. This highlights its prowess in specialized domains and complex reasoning tasks. By comparing VibeThinker-1.5B against a wide range of state-of-the-art models across different scales and categories - including advanced reasoning models with Long-CoT capabilities and top-tier non-reasoning models - we establish its remarkable performance in diverse scenarios. Our evaluation settings using vLLM for inference backend ensure accurate assessments of the model's performance metrics. Additionally, a recent arXiv query titled "Tiny Model, Big Logic: Diversity-Driven Optimization Elicits Large-Model Reasoning Ability in VibeThinker-1.5B" further emphasizes the significance of our approach in eliciting large-model reasoning ability through diversity-driven optimization. Overall, our findings challenge existing Scaling Law assumptions by showcasing that small models like VibeThinker-1.5B can achieve remarkable reasoning capabilities comparable to larger counterparts while significantly reducing training and inference costs. This not only democratizes advanced AI research but also prompts a necessary re-evaluation of traditional scaling paradigms in the field of artificial intelligence.