Summary * We present a method for performing circuit analysis on language models using "transcoders," an occasionally-discussed variant of SAEs that provide an interpretable approximation to MLP sublayers' computations. Transcoders are exciting because they allow us not only to interpret the output of MLP sublayers but also to decompose the...
Epistemic status: preliminary/exploratory. Work performed as a part of Neel Nanda's MATS 5.0 (Winter 2023-2024) Research Sprint. TL;DR: We develop a method for understanding how sparse autoencoder features in transformer models are computed from earlier components, by taking a local linear approximation to MLP sublayers. We study both how the...
This sounds like it could work. I can think of a few reasons why this approach could be challenging, however:
1. We don't really know how transcoders (or SAEs, to the best of my knowledge) behave when they're being trained to imitate a model component that's still updating
2. Substituting multiple transcoders at once is possible, but degrades model performance a lot compared to single-transcoder substitutions. Substituting one transcoder at a time would require restarting the forward pass at each layer.
3. If the transcoders are used to predict next tokens, they may lose interpretability and return to superposition.
Under a "transcoder-aware" training regime, these would be the first things I would check for.
Also, you may be interested in Jacob's comment here for some details on when we tried to co-train SAEs and transcoders to have sparse connections to one another. This is a very different question, of course, but it provides some preliminary evidence that the fidelity-interpretability tradeoff persists across more elaborate training settings.
Epistemic status: preliminary/exploratory.
Work performed as a part of Neel Nanda's MATS 5.0 (Winter 2023-2024) Research Sprint.
TL;DR: We develop a method for understanding how sparse autoencoder features in transformer models are computed from earlier components, by taking a local linear approximation to MLP sublayers. We study both how the feature is activated on specific inputs, and take steps towards finding input-independent explanations via examining model weights. We demonstrate this method with several deep-dive case studies to interpret the mechanisms used by simple transformers (GELU-1L and GELU-2L) to compute some specific features, and validate that it agrees with the results of causal methods.
A core aim of mechanistic interpretability is tackling the curse of dimensionality,... (read 12541 more words →)
I see. I was in fact misunderstanding this detail in your training setup. In this case, only engineering considerations really remain: these boil down to incorporating multiple transcoders simultaneously and modeling shifting MLP behavior with transcoders. These seem like tractable, although probably nontrivial and, because of the LLM pretraining objective, quite computationally expensive. If transcoders catch on, I hope to see someone with the compute budget for it run this experiment!