Why the control mechanisms used in LLMs don't translate to smaller models

Transferring control from one LLM to another appears to work only once a certain size threshold is reached. In the study, this threshold was found to be approximately 1.7 billion parameters—below this threshold, the similarity between models loses its practical significance.

The authors compared five open-weight models across two architectural approaches and three scales. For each, they trained a Sparse Autoencoder, identified directions for 15 semantic concepts, and tested all 20 directed model pairs.

For models with 1.7 billion parameters or more, between 47% and 49% of feature pairs showed noticeable alignment. But what’s more interesting is this: in the behavior control test, vectors transferred from another model achieved a 71% success rate, compared to 68% for native vectors from the same model.

This does not mean that there is a single universal lever for all LLMs. A single common vector worked in four out of five models and achieved a 67.3% success rate. However, the fact itself is significant: for some models, behavior can be controlled without fine-tuning precisely because their internal spaces turned out to be functionally compatible.

The practical implication for developers of interpretability tools is unpleasant: a result obtained on a 7B-level model cannot be transferred to a smaller model without verification. The study covers only five models, and for one of them, generation was unstable; transfer performance also deteriorated significantly for models with fewer than 1.7 billion parameters. Therefore, the identified threshold should currently be viewed as a strong indicator rather than a universal rule applicable to all architectures and tasks.

If you’re building LLM management tools, which is more important to you: a single transferable mechanism for different models, or individually testing each model before use?

Source: cs.CL updates on arXiv.org