Wenwei Zhang

Mixed SFT outperforms next-chunk reasoning RL while consuming over 60 times less compute, reshaping our understanding of effective training strategies with no-CoT data.

Achieving similar performance to larger models with significantly less data and faster inference speeds could redefine efficiency benchmarks in foundation models.

Intern-S2-Preview-397B not only excels in multimodal scientific reasoning but also enhances biological instruction performance without altering its foundational architecture.

Task-specific optimization in multimodal models can significantly boost performance by dynamically reallocating shared computation resources between vision and language components.

GEPO achieves balanced cross-task improvements in RL for LLMs by dynamically adjusting advantages based on group entropy, outperforming traditional methods.

AdvancedMathBench reveals that even state-of-the-art models struggle with advanced mathematical reasoning, achieving only 75.8% accuracy in proof generation.

Visual Pretraining outperforms text-only methods, revealing that rich visual cues can enhance language model performance in ways previously underestimated.