New 'Thinking with Drafting' method reconstructs latent logic from visual data

Researchers have introduced a new method called Thinking with Drafting (TwD) to improve visual reasoning in multimodal large language models. TwD reconceptualizes processing visual inputs as optical decompression, reconstructing latent logical structures from visual tokens. This approach uses a minimalist Domain-Specific Language (DSL) as an intermediate representation, forcing models to draft their reasoning into executable code for self-verification. Experiments on a new visual algebra benchmark, VisAlg, show that TwD enhances cognitive scaffolding and visual generation acts as a logical verifier. AI

IMPACT Introduces a new framework for visual reasoning that could improve the accuracy and verifiability of multimodal AI systems.

RANK_REASON This is a research paper introducing a novel method for visual reasoning in multimodal models.

Read on arXiv cs.CL →

• arXiv
• Caijun Jia
• Domain-Specific Language
• Thinking with Drafting

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →