DOMINO: A Dual-System for Multi-step Visual Language Reasoning

TL;DR

This paper introduces DOMINO, a dual-system approach for multi-step visual language reasoning that separates visual information extraction from deliberate reasoning, improving accuracy over existing methods.

Contribution

The work presents a novel dual-system framework that decomposes visual language reasoning into two steps, enhancing interpretability and performance with minimal training data.

Findings

Outperforms prior methods on chart and plot datasets.

Fine-tuning System-2 improves accuracy by over 5%.

Achieves state-of-the-art results with limited data.

Abstract

Visual language reasoning requires a system to extract text or numbers from information-dense images like charts or plots and perform logical or arithmetic reasoning to arrive at an answer. To tackle this task, existing work relies on either (1) an end-to-end vision-language model trained on a large amount of data, or (2) a two-stage pipeline where a captioning model converts the image into text that is further read by another large language model to deduce the answer. However, the former approach forces the model to answer a complex question with one single step, and the latter approach is prone to inaccurate or distracting information in the converted text that can confuse the language model. In this work, we propose a dual-system for multi-step multimodal reasoning, which consists of a "System-1" step for visual information extraction and a "System-2" step for deliberate reasoning. Given an input, System-2 breaks down the question into atomic sub-steps, each guiding System-1 to extract the information required for reasoning from the image. Experiments on chart and plot datasets show that our method with a pre-trained System-2 module performs competitively compared to prior work on in- and out-of-distribution data. By fine-tuning the System-2 module (LLaMA-2 70B) on only a small amount of data on multi-step reasoning, the accuracy of our method is further improved and surpasses the best fully-supervised end-to-end approach by 5.7% and a pipeline approach with FlanPaLM (540B) by 7.5% on a challenging dataset with human-authored questions.