Spatial Priming Outperforms Semantic Prompting: A Grid-Based Approach to Improving LLM Accuracy on Chart Data Extraction

Authors: Andrei Lazarev, Dmitrii Sedov, Alexander Galkin

arXiv · PDF

Summary

The authors tested whether multimodal LLMs extract data from scientific charts more accurately when given high-level semantic guidance (like metadata or chain-of-thought) versus low-level spatial cues. Semantic methods (two-stage metadata-first, chain-of-thought) produced no significant improvement, but a simple spatial intervention — overlaying a coordinate grid on the chart image — significantly reduced extraction error (SMAPE dropped from 25.5% to 19.5%, p < 0.05) on a synthetic dataset. The takeaway is that current multimodal models benefit more from explicit spatial scaffolding than from abstract reasoning prompts for this task.

Main takeaways:

• High-level semantic prompting (metadata-first, chain-of-thought) failed to improve chart data extraction accuracy.
• A coordinate grid overlay — simple spatial priming — significantly reduced error (SMAPE 25.5% → 19.5%, p < 0.05).
• For current multimodal models, explicit spatial context is more effective than semantic guidance on structured visual tasks.
• The result suggests that what helps vision-language models is concrete perceptual scaffolding, not higher-level reasoning prompts.

Relevance

Tangential — only relevant if I ever explore how different kinds of context (spatial scaffolding vs. semantic instructions) affect persona behavior or marker implantation, though the core question (what kind of input structure actually changes model behavior?) loosely parallels my installation-path work.

Abstract

arXiv:2605.08220v1 Announce Type: new Abstract: The automated extraction of data from scientific charts is a critical task for large-scale literature analysis. While multimodal Large Language Models (LLMs) show promise, their accuracy on non-standardized charts remains a challenge. This raises a key research question: what is the most effective strategy to improve model performance (high-level semantic priming) or low-level spatial priming? This paper presents a comparative investigation into these two distinct strategies. We describe our exploratory experiments with semantic methods, such as a two-stage metadata-first framework and Chain-of-Thought, which failed to produce a statistically significant improvement. In contrast, we present a simple but highly effective spatial priming method: overlaying a coordinate grid onto the chart image before analysis. Our quantitative experiment on a synthetic dataset demonstrates that this grid-based approach provides a statistically significant reduction in data extraction error (SMAPE reduced from 25.5% to 19.5%, p < 0.05) compared to a baseline. We conclude that for the current generation of multimodal models, providing explicit spatial context is a more effective and reliable strategy than high-level semantic guidance for this class of tasks.