KAISA: Knowledge And Internal States Augmented Agent for Chemistry Knowledge Question Answering

Large Language Models (LLMs) often struggle with rigorous scientific tasks, suffering from hallucinations and calculation errors due to the lack of precise domain knowledge. While Retrieval-Augmented Generation (RAG) can mitigate these issues, standard approaches frequently encounter knowledge conflicts between parametric memory and retrieved context, as well as sensitivity to noise. To address these challenges, we introduce KAISA-Agent (Knowledge And Internal States Augmented Agent), a novel framework that synergizes external knowledge with internal inference states. Our approach employs an iterative tool invocation mechanism for precise computation and a dual-path hybrid sampling strategy that dynamically arbitrates between internal reasoning and external evidence using a sequence-level confidence metric. Experiments demonstrate that KAISA-Agent significantly enhances a lightweight 8B model, achieving a 10.34% accuracy improvement.