Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
By Samuel Marks, C. Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, Aaron Mueller. International Conference on Learning Representations, 2024
394 citations recorded by the source corpus.
We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.
Subfields: Artificial Intelligence
View Paper: https://www.semanticscholar.org/paper/b8f280d8bf685f8da7c83068e73f000528072d6b
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Efstratios Gavves
- Leonard Bereska
- Vikrant Varma
- Neel Nanda
- Wes Gurnee
- Arthur Conmy
- Aaquib Syed
- Nina Rimsky
Citations in the corpus
- Mechanistic Interpretability for AI Safety - A Review (2024, background): “Recent work by Marks et al. (2024) and Braun et al. (2024) demonstrates the potential of integrating sparse autoen-17 Under review as submission to TMLR coders with automated circuits discovery (Conmy et al., 2023; Syed et al., 2023), combining feature-level analysis with circuit-level…”
- Refusal in Language Models Is Mediated by a Single Direction (2024, positive): “Inspired by the rapid progress of mechanistic interpretability (Bricken et al., 2023; Marks et al., 2024; Nanda et al., 2023; Templeton et al., 2024) and activation steering (Panickssery et al., 2023; Turner et al., 2023; Zou et al., 2023a), this work leverages the internal representations of chat…”
- Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 (2024, positive): “How can we practically extend the SAE feature circuit finding algorithm in Marks et al. (2024) to larger models?”
- Scaling and evaluating sparse autoencoders (2024, neutral): “Marks et al. [2024] showed that the resulting features from sparse autoencoders can find sparse circuits in language models.”