Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models

Patrick Leask, Neel Nanda, Noura Al Moubayed
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:32803-32829, 2025.

Abstract

Sparse Autoencoders (SAEs) are a popular method for decomposing Large Language Model (LLM) activations into interpretable latents, however they have a substantial training cost and SAEs learned on different models are not directly comparable. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activation models (ITDAs). ITDAs are constructed by greedily sampling activations into a dictionary based on an error threshold on their matching pursuit reconstruction. ITDAs can be trained in 1% of the time of SAEs, allowing us to cheaply train them on Llama-3.1 70B and 405B. ITDA dictionaries also enable cross-model comparisons, and outperform existing methods like CKA, SVCCA, and a relative representation method on a benchmark of representation similarity. Code available at https://github.com/pleask/itda.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-leask25a, title = {Inference-Time Decomposition of Activations ({ITDA}): A Scalable Approach to Interpreting Large Language Models}, author = {Leask, Patrick and Nanda, Neel and Al Moubayed, Noura}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {32803--32829}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/leask25a/leask25a.pdf}, url = {https://proceedings.mlr.press/v267/leask25a.html}, abstract = {Sparse Autoencoders (SAEs) are a popular method for decomposing Large Language Model (LLM) activations into interpretable latents, however they have a substantial training cost and SAEs learned on different models are not directly comparable. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activation models (ITDAs). ITDAs are constructed by greedily sampling activations into a dictionary based on an error threshold on their matching pursuit reconstruction. ITDAs can be trained in 1% of the time of SAEs, allowing us to cheaply train them on Llama-3.1 70B and 405B. ITDA dictionaries also enable cross-model comparisons, and outperform existing methods like CKA, SVCCA, and a relative representation method on a benchmark of representation similarity. Code available at https://github.com/pleask/itda.} }
Endnote
%0 Conference Paper %T Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models %A Patrick Leask %A Neel Nanda %A Noura Al Moubayed %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-leask25a %I PMLR %P 32803--32829 %U https://proceedings.mlr.press/v267/leask25a.html %V 267 %X Sparse Autoencoders (SAEs) are a popular method for decomposing Large Language Model (LLM) activations into interpretable latents, however they have a substantial training cost and SAEs learned on different models are not directly comparable. Motivated by relative representation similarity measures, we introduce Inference-Time Decomposition of Activation models (ITDAs). ITDAs are constructed by greedily sampling activations into a dictionary based on an error threshold on their matching pursuit reconstruction. ITDAs can be trained in 1% of the time of SAEs, allowing us to cheaply train them on Llama-3.1 70B and 405B. ITDA dictionaries also enable cross-model comparisons, and outperform existing methods like CKA, SVCCA, and a relative representation method on a benchmark of representation similarity. Code available at https://github.com/pleask/itda.
APA
Leask, P., Nanda, N. & Al Moubayed, N.. (2025). Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:32803-32829 Available from https://proceedings.mlr.press/v267/leask25a.html.

Related Material