🤖 AI Summary
This work addresses the challenge of efficiently embedding interpretable and controllable internal representations within large language models without resorting to complex feature identification and intervention procedures. The authors propose introducing a lightweight auxiliary loss during fine-tuning to designate 16 fixed dimensions in the residual stream as inert yet semantically meaningful concept indicators. By leveraging the model’s inherent pressure for computational efficiency, this approach encourages the model to actively rely on these pre-defined markers during generation, thereby automatically suppressing redundant encoding. Experiments demonstrate that this method enables precise and controllable generation interventions within a 3072-dimensional residual stream, validating the feasibility of constructing interpretable and manipulable internal representations through efficiency-driven training dynamics.
📝 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.