Learning to Reduce: Optimal Representations of Structured Data in Prompting Large Language Models

TL;DR

This paper introduces a framework called Learning to Reduce that fine-tunes language models to generate condensed versions of structured data, enhancing reasoning and performance on downstream tasks involving long contexts.

Contribution

The paper presents a novel reinforcement learning-based method to train models for effective context reduction, improving evidence selection and reasoning in large language models.

Findings

Achieves comparable accuracy in evidence selection

Enhances LLM performance on downstream tasks with long contexts

Demonstrates good generalizability across datasets

Abstract

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.