Saved in:
Bibliographic Details
Main Authors: Lee, Younghun, Kim, Sungchul, Yu, Tong, Rossi, Ryan A., Chen, Xiang
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2402.14195
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914688671940608
author Lee, Younghun
Kim, Sungchul
Yu, Tong
Rossi, Ryan A.
Chen, Xiang
author_facet Lee, Younghun
Kim, Sungchul
Yu, Tong
Rossi, Ryan A.
Chen, Xiang
contents Large Language Models (LLMs) have been widely used as general-purpose AI agents showing comparable performance on many downstream tasks. However, existing work shows that it is challenging for LLMs to integrate structured data (e.g. KG, tables, DBs) into their prompts; LLMs need to either understand long text data or select the most relevant evidence prior to inference, and both approaches are not trivial. In this paper, we propose a framework, Learning to Reduce, that fine-tunes a language model to generate a reduced version of an input context, given a task description and context input. The model learns to reduce the input context using On-Policy Reinforcement Learning and aims to improve the reasoning performance of a fixed LLM. Experimental results illustrate that our model not only achieves comparable accuracies in selecting the relevant evidence from an input context, but also shows generalizability on different datasets. We further show that our model helps improve the LLM's performance on downstream tasks especially when the context is long.
format Preprint
id arxiv_https___arxiv_org_abs_2402_14195
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Learning to Reduce: Optimal Representations of Structured Data in Prompting Large Language Models
Lee, Younghun
Kim, Sungchul
Yu, Tong
Rossi, Ryan A.
Chen, Xiang
Computation and Language
Large Language Models (LLMs) have been widely used as general-purpose AI agents showing comparable performance on many downstream tasks. However, existing work shows that it is challenging for LLMs to integrate structured data (e.g. KG, tables, DBs) into their prompts; LLMs need to either understand long text data or select the most relevant evidence prior to inference, and both approaches are not trivial. In this paper, we propose a framework, Learning to Reduce, that fine-tunes a language model to generate a reduced version of an input context, given a task description and context input. The model learns to reduce the input context using On-Policy Reinforcement Learning and aims to improve the reasoning performance of a fixed LLM. Experimental results illustrate that our model not only achieves comparable accuracies in selecting the relevant evidence from an input context, but also shows generalizability on different datasets. We further show that our model helps improve the LLM's performance on downstream tasks especially when the context is long.
title Learning to Reduce: Optimal Representations of Structured Data in Prompting Large Language Models
topic Computation and Language
url https://arxiv.org/abs/2402.14195