Graph Serializations for Reasoning LLMs: A Study Across Knowledge Extraction and Structural Reasoning

Recent research explores LLM capabilities regarding knowledge and reasoning. One study evaluates 15 graph representation formats across nine frontier LLM configurations, finding that serialization significantly impacts structural graph reasoning and ontology-based knowledge extraction performance. Simultaneously, another study investigates whether knowledge and reasoning benchmarks induce stable internal representation signatures in five models ranging from 7B to 31B parameters. Results indicate that while raw probes recover MMLU partitions, prompt text predicts these signatures nearly as effectively.