TRACE the Evidence: Constructing Knowledge-Grounded Reasoning Chains for Retrieval-Augmented Generation
A paper proposing TRACE, a method for constructing knowledge-grounded reasoning chains to enhance multi-hop question answering.
The authors propose TRACE, a method that constructs knowledge-grounded reasoning chains to improve multi-hop question answering. TRACE uses a KG Generator and Autoregressive Reasoning Chain Constructor to build reasoning chains from retrieved documents. Experimental results show an average performance improvement of up to 14.03% compared to using all retrieved documents.
Based on: TRACE the Evidence: Constructing Knowledge-Grounded Reasoning Chains for Retrieval-Augmented Generation