Rank-1 LoRAs Encode Interpretable Reasoning Signals

TL;DR

This paper demonstrates that small, interpretable rank-1 parameter modifications to language models can recover most reasoning capabilities, providing insights into the underlying mechanisms of reasoning in large language models.

Contribution

The work introduces a minimal rank-1 LoRA adapter that captures most reasoning performance and reveals interpretable, reasoning-specific activations, advancing understanding of model behavior.

Findings

Rank-1 LoRA recovers 73-90% of reasoning performance.

LoRA activations are as interpretable as MLP neurons.

Autoencoder analysis uncovers fine-grained, monosemantic features.

Abstract

Reasoning models leverage inference-time compute to significantly enhance the performance of language models on difficult logical tasks, and have become a dominating paradigm in frontier LLMs. Despite their wide adoption, the mechanisms underpinning the enhanced performance of these reasoning models are not well understood. In this work, we show that the majority of new capabilities in reasoning models can be elicited by small, single-rank changes to base model parameters, with many of these changes being interpretable. Specifically, we use a rank-1 LoRA to create a minimal parameter adapter for Qwen-2.5-32B-Instruct which recovers 73-90% of reasoning-benchmark performance compared to a full parameter finetune. We find that the activations of this LoRA are as interpretable as MLP neurons, and fire for reasoning-specific behaviors. Finally, we train a sparse autoencoder on the entire activation state of this LoRA and identify fine-grained and monosemantic features. Our findings highlight that reasoning performance can arise largely from minimal changes to base model parameters, and explore what these changes affect. More broadly, our work shows that parameter-efficient training methods can be used as a targeted lens for uncovering fundamental insights about language model behavior and dynamics.