Research & Papers

Flow of Spans: Generalizing Language Models to Dynamic Span-Vocabulary via GFlowNets

This new approach could fundamentally change how language models generate text...

Deep Dive

Researchers introduced Flow of SpanS (FoSS), a new AI framework that generates text using flexible text spans instead of fixed tokens. By employing Generative Flow Networks (GFlowNets) to navigate a directed acyclic graph structure, FoSS achieves up to 12.5% higher MAUVE scores than standard Transformers in text generation and shows 3.5% gains on knowledge tasks. The model benefits from scaling with larger models and more data, outperforming state-of-the-art methods.

Why It Matters

This breakthrough could lead to more flexible, creative, and knowledge-rich AI writing assistants and content generators.