Proceedings of the Annual Meeting of the Cognitive Science Society

Complex relational understanding tasks such as Document-level Relation Extraction require resolving semantic ambiguity amid a quadratic explosion of entity pairs, causing severe class imbalance and false negatives. Although large language models show strong reasoning ability, directly applying them to DocRE is inefficient and prone to hallucination. Inspired by dual-process theories of human cognition, we propose DocRE-Thinker, a hybrid framework integrating fast intuition with controlled deliberation. A fine-tuned backbone serves as System 1, efficiently generating candidate relations while estimating uncertainty. A frozen large language model acts as System 2 and is invoked only for high-uncertainty cases, performing uncertainty arbitration and attribute-based rule induction to justify relations. These symbolic rules are converted into supervision through a text-gradient feedback mechanism, enabling System 1 to internalize analytical reasoning in a rule-aware manner. Experiments on DocRED and Re-DocRED show state-of-the-art performance with significant recall gains across benchmarks, validating robust and efficient cognitively grounded relational reasoning.