Circuits Updates – July 2025

A two-item Anthropic interpretability update. Chris Olah restates the 2021 Mathematical Framework's eigenvalue arguments for copy, previous-token and induction heads in terms of features and transformed features such as prev(X) and say(X); Sam Zimmerman reviews five early papers that applied sparse autoencoders to protein, DNA and cell models, as read in December 2024, with no new experiments in either item.

Chris Olah, Sam Zimmerman et al. (Anthropic) · 2025-07-25 · original

The July 2025 Circuits Update (dated 25 July 2025 on the Transformer Circuits index) has two items, neither of which reports a new experiment. The first, by Chris Olah, revisits A Mathematical Framework for Transformer Circuits (2021): it explains why that paper's eigenvalue tests for copying and induction worked, by restating the heads' behavior in terms of features. The second, by Sam Zimmerman, is a literature review of early work that applied sparse autoencoders (SAEs) to biological models, mainly protein language models, based on papers the team read in December 2024. The update appeared in a month when the team also posted Tracing Attention Computation Through Feature Interactions, Sparse mixtures of linear transforms and A Toy Model of Interference Weights; the update itself does not refer to them.

About the Circuits Updates format

Circuits Updates are informal posts on Anthropic's Transformer Circuits site, each a bundle of short, separately authored items. This one describes its contents as "developing ideas" and asks readers to treat them "like those of a colleague sharing some thoughts or preliminary experiments for a few minutes at a lab meeting, rather than a mature paper" (introduction). The web page is the version of record; no arXiv or anthropic.com version was found. The page has no numbered figures: the first item's argument is carried by six schematic matrix diagrams, and the second item's comparison of papers by a table. Locations below are item and subsection titles. The source runs to about 3,300 words, most of it the second item.

Terms used in this note

  • Feature. A direction in a model's activations that corresponds to an interpretable property, of the kind SAEs extract (see Towards Monosemanticity).
  • Superposition. A model representing more features than it has dimensions, so features share neurons (see Toy Models of Superposition).
  • OV and QK circuits. The two parts of an attention head: the OV circuit determines what the head reads from the attended position and writes to the output, and the QK circuit determines where it attends.
  • Copy head, previous-token head, induction head. A copy head raises the probability of tokens it attends to; a previous-token head moves information about the preceding token forward; an induction head attends to the token that followed an earlier occurrence of the current token and predicts it again (see In-context Learning and Induction Heads).
  • Transformed features. The item's notation for features related one-to-one to an original feature X: prev(X), "X was present at the previous token", and say(X), a feature that makes the model produce output that would trigger X.

Revisiting A Mathematical Framework in the language of features

Chris Olah (edited by Adam Jermyn) starts from a limitation of the 2021 paper: when it was written, the team "had no way to extract features from superposition", so ideas that are naturally about features were stated as eigenvalues of products of weight matrices ("Revisiting A Mathematical Framework with the Language of Features"). The item restates three cases as feature-to-feature matrices, drawn as grids with one red cell per strong entry. It does not compute these matrices for any model; the diagrams are idealized sketches.

The item adds a parenthetical hypothesis: in practice the directions for transformed features are likely to be linear transformations of the original features' directions, citing Wattenberg and Viégas (2024) on matrix binding and "echo" features.

Copy heads

A copy head's OV circuit should map each feature X to say(X) ("Copy Heads"). The Framework paper tested for this with positive eigenvalues of the product of the unembedding, the OV matrix and the embedding. The item's explanation of why that test was reasonable: in early layers the token embeddings act as a rough basis for simple per-token features, and in late layers the unembeddings act as a basis for the matching say(X) features, especially in small models. Multiplying through them isolates the block of the feature matrix that maps X to say(X), which should be roughly diagonal and therefore has positive eigenvalues. Positive eigenvalues in that block, the item says, mean that sets of features upweight the matching set of say features.

Previous-token heads

The item states as the team's belief that a previous-token head copies information about the preceding token forward as distinct features, mapping X to prev(X) ("Previous Token Heads"). If features for the previous token are already present, they may be copied in turn as "previous of previous" features; the item notes that a head attending two tokens back could also build those directly.

Induction heads

An induction head's OV circuit is a copying map, X to say(X), like a copy head's ("Induction Head"). The new content is the QK circuit, which the item paraphrases as: "If the query token is X, we want a key preceded by X. If the query token is preceded by Y, we want a key token which was Y two tokens back." The diagram shows a query feature X matching a key feature prev(X), and prev(Y) at the query matching prev(prev(Y)) at the key. The Framework paper had studied eigenvalues of the induction head's QK matrix multiplied by the previous-token head's OV matrix; in feature terms that product should again be roughly diagonal, which is why its eigenvalues came out positive.

What the item adds and does not

The item is an explanatory rewrite, not a new analysis: it gives no measurements, does not decompose any real head into features, and does not test the hypothesis that transformed features are linear transformations of the originals. Its stated aim is to "hopefully make the work clearer". The Framework paper's own caveat, that "there can be matrices with all positive eigenvalues that actually map some tokens to decreasing the logits of that same token" (a footnote in its section on detecting copying), is not revisited here. The later Tracing Attention Computation paper explains attention scores as sums of interactions between features at query and key positions; this note connects the two, and the update does not cite that paper.

Applications of interpretability to biology

Sam Zimmerman (edited by Harish Kamath and Adam Jermyn) summarizes work applying SAEs to biological sequence models ("Applications of Interpretability to Biology"). A footnote says the item covers only a few papers, "primarily ones reviewed by the Anthropic team during a lunch and learn in December 2024", and another that the review "is based on earlier versions of these papers available in December 2024", noting that InterProt and InterPLM had since been updated. The item reports no Anthropic experiments; every result below is the cited paper's, as the item describes it. This note has not checked those descriptions against the papers, except where stated.

The papers reviewed

The item's table compares five works ("Sparse Autoencoders in Biology"):

Work Model studied SAE type, as listed
InterPLM (Simon and Zou) ESM-2, 8M parameters L1, hidden dimension 10,420
InterProt (Adams et al.) ESM-2, 650M parameters TopK, hidden dimensions up to 16,384
Reticular (Parsan, Yang and Yang) ESM-2 3B, the base of ESMFold Matryoshka, dictionary size 10,240
Evo 2 (Brixi et al.) Evo 2, 7B parameters, a DNA model BatchTopK, dictionary size 32,768
Markov Bio (Green) a gene expression model not specified

The table also lists each work's validation method, for example Swiss-Prot annotations covering 433 concepts for InterPLM and linear probes on four tasks for InterProt. A footnote mentions SAE work on biomedical imaging models, excluded from the review.

Examples the item highlights

  • A missing annotation (InterPLM). Feature f/939 activates on the Nudix box motif, a conserved pattern in certain enzymes. One protein, B2GFH1, activated it strongly without carrying the Nudix box annotation in Swiss-Prot; the researchers found the motif present on checking the structure and other databases such as InterPro, and found several such cases in which apparent false positives were missing annotations ("An Example: Finding Missing Protein Annotations"). A later paragraph attributes the missing-annotation discovery to InterProt ("This parallels InterProt's discovery"), while every other mention credits InterPLM. This note searched the InterProt preprint's version 1 (February 2025), the version the item links, and found no missing-annotation result in it.
  • A prophage feature (Evo 2). Feature f/19746 activates across prophage regions in bacterial genomes and on CRISPR spacer sequences. Scrambling the spacers left its activation intact, while scrambling the CRISPR direct repeats removed it, which the item reads as the model having learned "the functional relationship between phages and bacterial immunity rather than superficial sequence similarity" ("Discovering Evolutionary Relationships Through Features"). The feature also activated on unannotated regions containing phage-associated genes such as integrases.
  • Structure from sequence (Evo 2). Features activating on α-helices (f/28741) and β-sheets (f/22326) were mapped onto AlphaFold 3 structure predictions, learned from genomic sequence alone ("A Spectrum of Discoveries").
  • Reticular. 8 to 32 active latents per position could reconstruct structural features, and higher layers reached 49% coverage of Swiss-Prot concepts. The table phrases the first result as "8–32 active latents total". The arXiv abstract of the Reticular paper confirms the scaling to ESM2-3B, the Matryoshka SAEs and a steering case study that increased solvent accessibility with the input sequence fixed.
  • InterProt and Markov Bio. InterProt found features for nuclear localization signals and thermostability, and features predictive of CHO cell expression with no clear biological interpretation. Markov Bio built feature-to-feature regulatory networks in a gene expression model to identify drivers of cell fate.

The item's synthesis and stated limits

The item lists cross-study patterns, each credited to specific papers: superposition more severe in biological models than in language models, more concept coverage at larger scale, features that align with biological units, and features from human genomes that annotated the woolly mammoth genome ("Cross Study Findings"). Under "Advanced Interpretability" it suggests these results point toward circuits, and it says biological models exhibit more hierarchical circuits than language models; the evidence given is Evo 2's features at levels from amino acids to whole proteins, with no comparison to a language model. Its limits section states that "Biological validation requires wet-lab experiments, not just computational metrics", that the link between SAE features and causal biological mechanisms needs experimental grounding, and that, as far as the author knows, multimodal biological models had not been studied this way ("Challenges and Measured Optimism"). It proposes validation protocols, multimodal SAE variants, cheaper training and circuit discovery as next steps, and says the team would be interested to see attribution graphs applied here.

Limitations and open questions

  • No new results. Neither item reports an experiment. The first is an explanatory sketch with idealized diagrams; the second summarizes others' preprints.
  • Dated sources. The review describes December 2024 versions of papers that had been revised by July 2025, and says so.
  • Secondhand claims. Results in the biology item, including the woolly mammoth annotation and the cross-study patterns, are the cited papers' claims as summarized; the item gives no numbers for most of them and no locations within the papers.
  • One apparent misattribution. The missing-annotation discovery is credited to InterProt in one sentence and to InterPLM everywhere else (see above).
  • Untested hypothesis. The linear-transformation account of transformed features is offered without a test in the item.

Research questions

  1. Do real copy and induction heads decompose into the item's feature matrices? The diagrams are idealized. Computing feature-to-feature OV and QK matrices for identified heads in a small model with SAE or transcoder features, and measuring how much weight lies on the predicted diagonal blocks, would test the account.
  2. Are prev(X) directions linear transforms of X directions? The item hypothesizes so. Fitting a single linear map from X features to candidate prev(X) features across many tokens, and measuring how well it generalizes to held-out tokens, would answer it.
  3. When does the eigenvalue test for copying mislead? The 2021 paper noted that positive eigenvalues do not guarantee every token raises its own logit. Comparing eigenvalue-based copying scores with feature-level diagonal strength across heads would show where the two disagree.
  4. How many SAE features in protein models correspond to missing annotations rather than errors? InterPLM reports cases. Taking a random sample of high-activation "false positives" across features and checking each against independent databases or experiments would give a rate.
  5. Is superposition more severe in biological models than in language models? The item states it. Measuring concepts per neuron, or SAE features per dimension at matched reconstruction error, for protein and language models of similar size would test it.
  6. Do attribution graphs work on protein or DNA models? The item suggests trying them. Building cross-layer transcoders for an open protein model and tracing a known structural prediction would show whether the method transfers.

How it connects

References

  • Olah, C., Zimmerman, S., Kamath, H., and Jermyn, A. (2025). Circuits Updates – July 2025. Transformer Circuits Thread, 25 July 2025 (date from the site index). https://transformer-circuits.pub/2025/july-update/index.html record
  • Elhage, N., Nanda, N., Olsson, C., et al. (2021). A Mathematical Framework for Transformer Circuits. Transformer Circuits Thread. https://transformer-circuits.pub/2021/framework/index.html record · 8 pages
  • Wattenberg, M., and Viégas, F. (2024). Relational Composition in Neural Networks: A Survey and Call to Action. arXiv:2407.14662. record
  • Simon, E., and Zou, J. (2024). InterPLM: Discovering interpretable features in protein language models via sparse autoencoders. bioRxiv 10.1101/2024.11.14.623630. record
  • Adams, E., Bai, L., Lee, M., Yu, Y., and AlQuraishi, M. (2025). From mechanistic interpretability to mechanistic biology: Training, evaluating, and interpreting sparse autoencoders on protein language models. bioRxiv 10.1101/2025.02.06.636901. record
  • Parsan, N., Yang, D. J., and Yang, J. J. (2025). Towards Interpretable Protein Structure Prediction with Sparse Autoencoders. arXiv:2503.08764. record
  • Brixi, G., Durrant, M. G., Ku, J., Poli, M., et al. (2025). Genome modeling and design across all domains of life with Evo 2. bioRxiv 10.1101/2025.02.18.638918. record
  • Green, A. (2024). Through a glass darkly: Mechanistic interpretability as the bridge to end-to-end biology. Markov Biosciences. https://www.markov.bio/research/mech-interp-path-to-e2e-biology record