RAISE: Enhancing Scientific Reasoning in LLMs via Step-by-Step Retrieval

2025-06-10

AI Summary

To address challenges in scientific reasoning—including difficulty in multi-step logical derivation, weak comprehension of domain-specific terminology, and lagging adaptation to dynamically evolving knowledge—this paper proposes a logic-relevance-driven, stepwise retrieval-augmented framework. First, complex questions are decomposed into logically coherent subtasks; second, structured logical queries are generated; and third, precise literature retrieval is performed along the reasoning path. Departing from conventional semantic-similarity-based retrieval, the framework achieves dual alignment between domain knowledge and logical structure. Technically, it integrates task decomposition, logic-aware retrieval, and collaborative reasoning with large language models. Evaluated on multiple scientific reasoning benchmarks, it significantly outperforms existing methods, yielding retrieved documents that exhibit both domain specificity and logical coherence. This work establishes a novel, interpretable, and traceable reasoning-augmentation paradigm for scientific AI.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Search and Optimization: Metareasoning and Metaheuristics

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Querying, indexing, and retrieval in Web-related graphs

Abstract

Scientific reasoning requires not only long-chain reasoning processes, but also knowledge of domain-specific terminologies and adaptation to updated findings. To deal with these challenges for scientific reasoning, we introduce RAISE, a step-by-step retrieval-augmented framework which retrieves logically relevant documents from in-the-wild corpus. RAISE is divided into three steps: problem decomposition, logical query generation, and logical retrieval. We observe that RAISE consistently outperforms other baselines on scientific reasoning benchmarks. We analyze that unlike other baselines, RAISE retrieves documents that are not only similar in terms of the domain knowledge, but also documents logically more relevant.

Problem

• Enhancing long-chain reasoning in scientific LLMs
• Retrieving domain-specific and updated scientific knowledge
• Improving logical relevance in document retrieval

Innovation

• Step-by-step retrieval-augmented framework RAISE
• Problem decomposition and logical query generation
• Retrieves logically relevant documents effectively