Title: CECOR: Correction-oriented synthetic data construction for factual error correction

Abstract:

Factual Error Correction (FEC) focuses on updating inaccurate text so that it aligns with external evidence. While current approaches excel at single-hop corrections, they frequently handle claims as isolated units, leading to poor performance in multi-hop scenarios that demand compositional reasoning across various sources of evidence. This limitation is exacerbated by the scarcity of paired training data and the complexity of identifying semantic errors within intricate reasoning chains. To address these issues, we introduce CECoR (Compositional Error Correction via Reasoning-aware Synthesis), a framework designed for reasoning-aware synthesis. CECoR employs a Decomposition and Injection paradigm to facilitate compositional error correction. It breaks down multi-hop assertions into transparent reasoning steps and applies controlled perturbations to generate high-quality training pairs. By integrating supervised fine-tuning with reinforcement learning in a two-stage learning process, the model enhances both factual precision and robustness. Extensive evaluations indicate that CECoR delivers superior results on multi-hop benchmarks, surpassing distantly supervised techniques and few-shot LLM baselines. Furthermore, the method generalizes well to single-hop tasks and maintains stability even when faced with noisy evidence, highlighting its practical utility for real-world factual correction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC