The Answer Lies Within: Self-Derived Rewards Enable Explainable Relation Extraction

A new framework that mimics human text-processing and uses self-derived rewards substantially improves both the accuracy and explainability of one-shot relation extraction in large language models, according to research posted on arXiv [1]. Large language models, or LLMs, are machine learning models with many parameters trained on vast amounts of text for natural language processing tasks [10]. Despite their capabilities, these models still struggle with one-shot relation extraction when predefined relation labels are absent [1]. Researchers Zhengliang Shi and colleagues identify two specific pitfalls: models are often misled by irrelevant tokens instead of relation-conveying semantics, and they frequently fail to align with the abstraction level that human annotators expect [1]. To address these gaps, the team introduces a framework with two components [1]. The first, COGRE, is a cognitively-inspired reasoning framework that structures relation extraction into a series of processes designed to mimic human text-processing [1]. Human cognitive behavior involves mental processes of learning, memory, and decision-making, which the COGRE framework attempts to replicate in a structured way [3]. The second component, HIT@DICT, is a reinforcement learning intermediate reward strategy that encourages reasoning to align with relational labels by rewarding relation-relevant phrases [1]. The reward is derived from a credit dictionary automatically extracted from correct predictions [1]. Experiments demonstrate significant gains. COGRE paired with the Qwen2.5-14B-Instruct model — part of the Qwen family of LLMs developed by Alibaba Cloud [11] — achieved a 24.65% F1 score on the One-shot NYT29 dataset, surpassing prior reasoning-based designs [1]. Optimizing this approach with reinforcement learning using HIT@DICT further improved performance by +23.46 percentage points [1]. A human evaluation also found that the best model generated relational phrases closely aligned with gold labels, increasing human explanation quality ratings by 54% relative to baselines [1]. The work was submitted to arXiv on October 7, 2025, and last revised on June 13, 2026 [1]. The paper is indexed on the Hugging Face Hub, where users can find related models, datasets, and demos linked to the arXiv identifier [6].