Planning vs Reasoning: Ablations to Test Capabilities of LoRA layers

TL;DR

This paper explores the effectiveness of LoRA layers in enhancing reasoning and planning abilities in language models, introducing a new dataset and analyzing the low-rank properties of LoRA matrices.

Contribution

It introduces HashChain Reasoning, a new dataset for testing reasoning, and systematically studies how LoRA layers improve reasoning and planning in GPT-2.

Findings

LoRA layers enhance reasoning capabilities in low-rank spaces.

Reasoning tasks require 2-3x lower rank in LoRA matrices than planning tasks.

Reasoning prefers low-parameter spaces for better generalization.

Abstract

Low-Rank Adaptation (LoRA) layers have emerged as a promising approach for efficient model fine-tuning, but their capabilities and limitations have not been fully explored. This paper: 1) Investigates the fundamental question of whether LoRA layers are effective at increasing reasoning + planning abilities 2) We introduce HashChain Reasoning, a novel evaluation dataset that deterministically tests reasoning capabilities. Through systematic ablation studies on GPT-2, we demonstrate that reasoning capabilities appear to exist primarily in low-rank spaces and can be effectively enhanced using LoRA layers. The effective rank analysis of trained LoRA matrices reveals a 2-3x lower rank requirement for reasoning tasks compared to planning tasks, giving context on where LoRA layers would be effective. This also provides evidence for reasoning fundamentally preferring low-parameter spaces for generalization.