LLM-FDC: LLM-Driven Fuzzy Rule Generation and Confidence-Based Decision-Making

Decision-making under uncertainty is a fundamental challenge for autonomous systems and intelligent robots operating in interactive environments, where task rules are described in natural language but their applicability must be validated through continuous interaction. Existing decision-making approaches under uncertainty either rely on deterministic rule formalisms or use large language models to directly generate fixed symbolic rules. Both approaches have notable drawbacks: the former cannot capture the uncertainty of rule applicability in specific task instances, while the latter lacks explicit and interpretable mechanisms for evolving and weighting uncertain knowledge during decision-making. To address these limitations, we propose LLM-FDC (LLM-Driven Fuzzy Rule Generation and Confidence-based Decision-making), a framework that integrates fuzzy rule induction with confidence-guided planning. At a high level, LLM-FDC generates complementary fuzzy rules from natural-language task descriptions, continuously updates their confidence values through interaction, and evaluates candidate actions via a confidence-weighted decision engine. This design preserves instance-level uncertainty, enables transparent belief evolution, and supports robust decision-making in interactive environments. Experiments conducted in the ScienceWorld environment demonstrate that LLM-FDC achieves higher success rates and faster reward accumulation compared with baseline methods. These results confirm that natural-language fuzzy rules, combined with explicit confidence evolution, provide a lightweight and interpretable mechanism for autonomous decision-making under uncertainty.