Learning to Plan and Generate Text with Citations
This work addresses the need for verifiable systems in information-seeking scenarios, offering an incremental improvement in attribution capabilities for domain-specific applications.
The paper tackles the problem of improving attribution quality in LLM-generated responses by proposing plan-based models that use question sequences as blueprints, showing that planning consistently improves attribution and yields more accurate citations compared to non-planning LLM pipelines.
The increasing demand for the deployment of LLMs in information-seeking scenarios has spurred efforts in creating verifiable systems, which generate responses to queries along with supporting evidence. In this paper, we explore the attribution capabilities of plan-based models which have been recently shown to improve the faithfulness, grounding, and controllability of generated text. We conceptualize plans as a sequence of questions which serve as blueprints of the generated content and its organization. We propose two attribution models that utilize different variants of blueprints, an abstractive model where questions are generated from scratch, and an extractive model where questions are copied from the input. Experiments on long-form question-answering show that planning consistently improves attribution quality. Moreover, the citations generated by blueprint models are more accurate compared to those obtained from LLM-based pipelines lacking a planning component.