Small language models struggle with graph property inference, study finds

A new study investigates the capabilities of small language models in inferring formal properties of relational structures presented textually. Researchers found that these models generally struggle to reliably estimate graph properties, with errors often exceeding intrinsic property dispersion and weak rank correlations across various configurations. However, the study identified that using adjacency-list encodings and multi-branch reasoning strategies can lead to measurable improvements in accuracy and consistency, even with limited model capacity. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Highlights limitations in small LLMs for structured reasoning, suggesting representational and reasoning strategy improvements are key.

RANK_REASON Academic paper on the limitations and potential improvements for small language models in structured reasoning tasks.

Read on arXiv cs.LG →

• arXiv
• Michal Podstawski

• paper
• other