When Dimensionality Hurts: The Role of LLM Embedding Compression for Noisy Regression Tasks

TL;DR

This paper investigates how embedding compression in large language models can improve performance on noisy regression tasks by reducing overfitting, while noting it may hinder tasks with strong causal input-output dependencies.

Contribution

It demonstrates that embedding compression via autoencoders can enhance LLM-based regression performance in noisy settings, revealing a regularizing effect and limitations based on task dependency.

Findings

Compression improves performance in noisy financial prediction tasks.

Compression mitigates overfitting in regression tasks.

Performance decreases in tasks with high causal input-output relationships.

Abstract

Large language models (LLMs) have shown remarkable success in language modelling due to scaling laws found in model size and the hidden dimension of the model's text representation. Yet, we demonstrate that compressed representations of text can yield better performance in LLM-based regression tasks. In this paper, we compare the relative performance of embedding compression in three different signal-to-noise contexts: financial return prediction, writing quality assessment and review scoring. Our results show that compressing embeddings, in a minimally supervised manner using an autoencoder's hidden representation, can mitigate overfitting and improve performance on noisy tasks, such as financial return prediction; but that compression reduces performance on tasks that have high causal dependencies between the input and target data. Our results suggest that the success of interpretable compressed representations such as sentiment may be due to a regularising effect.