RRescue: Ranking LLM Responses to Enhance Reasoning Over Context
Keywords: reasoning, long context, large language models, pairwise preference, ranking of LLM responses
TL;DR: We present a novel approach to optimize LLMs by ranking a set of contextually-grounded candidate responses, emphasizing partial ordering of responses for enhanced robustness.
Abstract: Effectively using a given context is paramount for large language models (LLMs). A context window can include task specifications, retrieved documents, previous conversations, and even model self-reflections, functioning similarly to episodic memory. While efforts are being made to expand the context window, studies indicate that LLMs do not use their context optimally for response generation. In this paper, we present a novel approach to optimize LLMs using ranking metrics, which teaches LLMs to rank a collection of contextually-grounded candidate responses. Rather than a traditional full ordering, we advocate for a partial ordering. This is because achieving consensus on the perfect order for system responses can be challenging. Our partial ordering is more robust, less sensitive to noise, and can be acquired through human labelers, heuristic functions, or model distillation. We test our system's improved contextual understanding using the latest benchmarks, including a new multi-document question answering dataset. We conduct ablation studies to understand crucial factors, such as how to gather candidate responses, determine their most suitable order, and balance supervised fine-tuning with ranking metrics. Our approach, named RRescue, suggests a promising avenue for enhancing LLMs' contextual understanding via response ranking.
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 888
Loading