Efficient Representations are Controllable Representations

TL;DR

This paper introduces a simple finetuning method that creates interpretable control features within LLMs by training specific residual stream dimensions as inert interpretability flags, enabling effective steering of model generation.

Contribution

The authors propose a straightforward auxiliary loss to embed controllable interpretability flags directly into LLMs, bypassing complex feature identification and intervention methods.

Findings

Inert flags become genuine internal features after finetuning.

Model relies on these flags during generation, enabling controllable steering.

Efficiency pressure encourages the model to organize around these interpretable features.

Abstract

What is the most brute-force way to install interpretable, controllable features into a model's activations? Controlling how LLMs internally represent concepts typically requires sophisticated methods to first identify, then intervene on the model's existing feature geometry. We bypass all of this. We finetune an LLM with a simple auxiliary loss, training 16 of its 3072 residual stream dimensions to be inert interpretability flags that simply indicate what concepts are required for generation. The model reorganizes around them anyway, learning to rely on these flags during actual generation tasks. As a result, these inert flags become genuine internal features: interpretable control switches that allow us to steer generation at inference time. Why does this work? When a feature is reliably supplied at a fixed location, gradient descent gradually eliminates redundant encodings elsewhere, and the model erodes its own alternative representations. A model's efficiency pressure is a lever - exploitable to induce interpretable, controllable representations.