Pith. sign in

REVIEW 4 cited by

Interpreting Attention Layer Outputs with Sparse Autoencoders

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2406.17759 v1 pith:ZMNEGPEZ submitted 2024-06-25 cs.LG

classification cs.LG
keywords saessparseattentionautoencodersfeaturesfindinterpretablemodel
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al.), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Legible-by-Construction: Attention and End-to-End Transformers

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Bounding attention values into fuzzy memberships or Boolean set operations yields 44–62% crisp selective detectors and an end-to-end legible transformer at baseline quality.

  2. Prototype Transformer: Towards Language Model Architectures Interpretable by Design

    cs.AI 2026-02 conditional novelty 6.0 of 10

    ProtoT is an autoregressive language model whose attention is replaced by learned prototype channels that are claimed to capture nameable concepts and allow targeted edits, at linear sequence cost but with slightly lo...

  3. Preserving Bilinear Weight Spectra with a Signed and Shrunk Quadratic Activation Function

    cs.LG 2025-09 reject novelty 5.0 of 10

    SQS is a new GLU activation that reportedly preserves bilinear weight spectra for interpretability while achieving competitive performance on MNIST, FMNIST, and Tiny Stories.

  4. Sparsification and Reconstruction from the Perspective of Representation Geometry

    cs.LG 2025-05 reject novelty 4.0 of 10

    Sparse encoding appears to stratify and compress feature representations, but the claimed causal link between cluster separation and reconstruction is not supported.

Pith tools