Stepwise Reasoning Enhancement for LLMs via External Subgraph Generation
Researchers have developed two new frameworks to improve the accuracy and interpretability of large language models (LLMs). The frameworks, SGR and SCA, aim to address limitations in complex multi-step reasoning.
SGR, proposed in a paper submitted on June 3, 2026[1], integrates LLMs with external knowledge graphs through query-relevant subgraph generation. It extracts key entities, relations, and constraints to construct a structured schema, then retrieves compact subgraphs from a knowledge graph. SGR combines direct Cypher-based reasoning with collaborative reasoning integration, improving reasoning accuracy and Hits@1 performance over standard prompting and several knowledge-enhanced baselines. Another framework, SCA, was introduced in a separate paper[2]. SCA assigns step-level confidence based on generated reasoning traces and applies the Information Bottleneck principle to identify potentially erroneous steps. Two complementary methods, NIBS and GIBS, are proposed. SCA reliably identifies low-confidence steps strongly correlated with reasoning errors. Using step-level confidence to guide self-correction improves the correction success rate by up to 13.5%[2].
research-papersafety-researchtool-releasebenchmark
Background sources we checked (4)
- arxiv.org ↗ Large language models have shown strong performance in natural language generation and downstream reasoning tasks, but they still struggle with logical consistency, factual grounding, and interpretability in complex multi-step reasoning. To address these limitations, this paper p…
- arxiv.org ↗ Stepwise Reasoning Enhancement for LLMs via External Subgraph Generation [...] # Stepwise Reasoning Enhancement for LLMs via External Subgraph Generation [...] Large language models have shown strong performance in natural language generation and downstream reasoning tasks, but t…
- arxiv.org ↗ Large Language Models (LLMs) have achieved strong performance across a wide range of natural language processing tasks in recent years, including machine translation, text generation, and question answering. As their applications extend to increasingly complex scenarios, however,…
- arxiv.org ↗ SGR: A Stepwise Reasoning Framework for LLMs with External Subgraph Generation [...] # SGR: A Stepwise Reasoning Framework for LLMs with External Subgraph Generation [...] Large Language Models (LLMs) have demonstrated strong capabilities across diverse NLP applications, such as …