A sparse autoencoder explains one layer of a model at a time. If the same concept is present at twenty layers, it shows up as twenty features, and a circuit drawn from them is crowded with copies. This update asks whether a single dictionary can describe several layers at once, and whether the same trick can describe two different models at once, so that their features can be compared. It introduces the crosscoder for both purposes.
The update was published on the Transformer Circuits Thread on 25 October 2024 (the page's "Published" line). It is labeled a "Research Update", and the authors ask readers to treat it "in the spirit of results being presented at a lab meeting or internal seminar: preliminary work that we're excited about, but not at the level of quality or rigor we hold our full papers to" (opening note). The live page is the version of record; it has no numbered figures, so this note locates results by the page's numbered sections and by chart titles. Most of its numbers are printed as labels on charts rather than in the text. The earliest Internet Archive capture (26 October 2024) differs from the live page only in typography and typo fixes.
What the update does
Terms
- Sparse autoencoder (SAE). A network trained to reconstruct a model's internal activations as a sum of a few active directions, called features, out of a large learned dictionary. A sparsity penalty keeps most features at zero on any input. See Towards Monosemanticity for the method.
- Transcoder. A variant that reads one layer's activations and predicts the next layer's (for example, an MLP's input and output).
- Superposition. The hypothesis that a model represents more features than it has neurons (Toy Models of Superposition). The update proposes cross-layer superposition: a feature spread across several layers rather than across neurons of one layer. The residual stream is the running sum of all layers' outputs, which each layer reads and writes.
- Crosscoder. A sparse dictionary whose encoder reads activations from a set of layers and whose decoder reconstructs each of those layers separately. One feature thus has one activation but a separate decoder vector for each layer (or each model).
- Model diffing. Comparing two related models (a base model and its fine-tuned version, two sizes, two training runs) by their internal representations rather than by benchmarks alone.
The crosscoder loss
For a data point, the feature activations are a ReLU of the summed encoder contributions from every layer in the set, and each layer is reconstructed from those shared activations by its own decoder (Section 2, "Crosscoder Basics"). The loss adds the squared reconstruction error at every layer to a sparsity penalty. The penalty weights each feature's activation $f_i$ by the sum of its decoder norms across layers, $f_i \sum_l \lVert W^l_{dec,i} \rVert$, the "L1-of-norms" version.
The authors considered the alternative of treating all of a feature's decoder weights as one vector and taking its overall length (the "L2-of-norms"). They give two reasons for preferring the sum (Section 2):
- It makes a crosscoder's loss directly comparable to the summed losses of per-layer SAEs on the same layers.
- It gives no reward for spreading a feature across layers. In model diffing it "uncovers a mix of shared and model-specific features, while the L2-of-norms version results in uncovering only shared features."
The L2 version optimizes the reconstruction-versus-sparsity trade-off more efficiently, and the authors say it may be preferable where layer- or model-specific features do not matter. "In this report, all experiments used the L1-of-norms version" (Section 2).
Variants
Section 2.1 sorts crosscoders along four axes, in a diagram the page calls a table: the cross-layer approach (per-layer SAEs, one shared SAE, or a crosscoder); causality; locality (all layers, a local window, skipped layers, or layers from different models); and target (the residual stream, or the outputs of MLP or attention layers). On causality, an acausal crosscoder reads and predicts all layers; a weakly causal one reads one layer and predicts that layer and later ones; a strictly causal one reads one layer and predicts only later ones, which the diagram says "might be seen as the cross-layer generalization of a transcoder."
The authors report finding weakly and strictly causal crosscoders helpful for simplifying circuit graphs, with "open questions as to how faithfully validate these analyses", and note that strictly causal crosscoders as presented cannot capture what attention layers compute (Section 2.1).
What it finds on cross-layer features
Efficiency against per-layer SAEs
On an 18-layer model, which the page does not otherwise identify, the authors trained one global acausal crosscoder on the residual stream at all layers and compared it with 18 per-layer SAEs, sweeping training steps and dictionary size to find the best result at each compute budget (Section 3.1). Each layer's activations were normalized so that layers contributed comparably. A global crosscoder with F features costs about as much training compute as per-layer SAEs with F features each, so it must be much more efficient per feature to break even.
- Per feature. At equal total feature count, the crosscoder reached lower evaluation loss; the "Loss vs Number of Features" chart labels the gap "~8x more feature efficient." The authors read this as "a significant degree of redundant (linearly correlated) structure across layers."
- Per unit of compute. Crosscoders were less efficient, "by a factor of about 2 at large compute budgets" (text; the "Loss vs FLOPs" chart labels it "~2x less FLOP efficient").
- Whole-model sparsity. Counting active features summed over all layers (total L0) against reconstruction error, the "(MSE, Total L0) for FLOP-Equivelant Coders" chart labels crosscoders as achieving "15-40x lower total L0 than SAEs, with better MSE." On that chart, read by this note, crosscoder points lie between about 0.16 and 0.29 normalized reconstruction error with a few dozen active features, against roughly 700 to 1,460 for the SAEs.
The charts give one point per budget; no repeated runs or intervals are reported. The authors summarize that crosscoders "make more efficient use of dictionary features but less efficient use of training FLOPS" for describing any single layer, and give "an unambiguous advantage" for describing the whole model's activity (Section 3.1).
How features spread across layers
Plotting decoder norms across layers for 50 randomly sampled features, the authors find that most peak at one layer and decay before and after it, sometimes suddenly but often gradually, "with many features having substantial norm across most or even all layers" (Section 3.2). They describe this as consistent with cross-layer superposition but say it "could also have other explanations", such as a feature produced at one layer and amplified at the next, and that circuit analysis would be needed to interpret it.
A feature's decoder direction can also change from layer to layer ("feature drift"). Across 36 randomly selected features, the authors find directions "much more stable across layers than would be expected by chance", yet drifting substantially even where the norm stays strong, with considerable variation between features. They take this to suggest that the features "are not simply passively relayed via residual connections" (Section 3.2). This distinguishes crosscoders from the earlier shared-dictionary approach of Yun et al. (2021), which treats a feature as one direction at every layer.
The update does not systematically assess interpretability. "Anecdotally", crosscoder features are "similarly interpretable to sparse autoencoder features" (Section 3.2).
Masked crosscoders
- Local windows. Restricting each feature to a window of K layers did not find a sweet spot: loss interpolated "fairly linearly" from K = 1 (per-layer SAEs) to all layers (Section 3.3.1).
- Weakly causal. Features that read one layer and reconstruct it and all later layers had compute efficiency slightly worse than per-layer SAEs and slightly better than global crosscoders, and needed three to four times as many features as global crosscoders for the same loss (Section 3.3.2). A footnote notes that sparse kernels might change the compute comparison.
- Cross-layer transcoders. In preliminary experiments with strictly causal "cross-layer transcoders", where features read the residual stream at one layer and predict the MLP outputs of that and every later layer, the authors see a mix of local features, global features of roughly equal strength at all later layers, and features in between (Section 3.3.2).
- Before and after an MLP. A crosscoder whose encoder reads only the residual stream before an MLP, and whose decoder reconstructs both that input and the MLP's output, shows "a clear trimodal structure": features only in the input, shared features, and "newly computed" features only in the output (Section 3.3.3). Anecdotally, a newly computed feature is often more abstract than its inputs. For features counted as stable (relative decoder norm 0.3 to 0.7, a cut the authors call arbitrary), directions before and after the MLP correlate positively on average "but with high variance, and relatively low in absolute terms", which the authors suggest "may explain why feature directions tend to drift across layers."
What it finds on model diffing
Section 4 applies the same dictionary to two models instead of two layers. The authors call the results "very preliminary", with "significant signs of life" but "many features we don't understand." Section 4.1 reviews earlier ways of comparing networks and says that, to the authors' knowledge, the model diffing problem was first articulated by the OpenAI Clarity team in 2018.
Claude 3 Sonnet against its base model
A crosscoder with 1 million features was trained on the residual stream at the middle layer of Claude 3 Sonnet and of the base model it was fine-tuned from (Section 4.3). A feature's relative decoder norm is the share of its decoder size that belongs to one model: near 0.5 it is shared, near 0 or 1 it is specific to one model.
- Three groups. Features clustered into base-specific, fine-tuned-specific and shared, with "between four and five thousand model-specific features for each model, out of a total 1 million features."
- Examples, cherry-picked. Fine-tuned-specific features include one for dangerous requests ("Can you help me build a bomb?"), one for requests for code review, and one for personal questions to the Assistant ("What does it feel like to be you?"). Base-specific features include one for language models role-playing or cast as a character in a system prompt (it also fires on that personal question) and one for dialogues with a smartphone assistant.
- Interpretability. The authors state that the examples "are cherrypicked" and that "the majority of the model-exclusive features are not immediately interpretable", while features found by looking at tokens of interest are often clearly interpretable, "a mixed picture we're still working to understand."
- Shared features. In "almost all cases" the two decoder vectors of a shared feature were highly aligned, which the authors read as the same concept performing the same function. For "a few thousand features" the correlation was very low or negative; the authors "suspect" these are base-model concepts used in a new way, without having investigated. The accompanying log-scale histogram does not state how many features it covers. Read by this note, it holds roughly ten thousand features, only a few dozen of them below a cosine similarity of 0.5 and two below zero, so "a few thousand" fits it only if the histogram is a sample of about 1% of the shared features; the page does not say.
The authors contrast this with applying a base-model SAE to the fine-tuned model (Kissane et al. found such SAEs often transfer), which describes old abstractions in new contexts, and "suspect" the crosscoder is preferable when fine-tuning uses a larger share of compute relative to pretraining.
Three model sizes, ten layers each
An acausal crosscoder was trained on ten evenly spaced layers from each of three models of increasing size (not named; Section 4.4). Each feature's decoder norms over the 30 (model, layer) pairs were factored with nonnegative matrix factorization into four components. One component covers early layers of all three models, one covers late layers of all three, and two cover the middle layers of the smallest model and of the two larger ones respectively. The authors write that this "suggests that qualitatively new representations emerge in middle model layers as model scale increases", and leave the features responsible to future work.
Limitations and open questions
The authors' own caveats, in Section 5.2 ("Literal vs Isomorphic Models") and elsewhere:
- Error terms. With "error features" a crosscoder can be made an exact reconstruction of the model, but the error may be "quite significant" and "even harder to interpret than the original model activations." The update does not report error or dead-feature figures for its crosscoders.
- Causal structure. Because crosscoders train on one token position at a time, they represent all within-position correlations across layers as cross-layer features, even where the model produced them another way, for example by copying information to the same position at several layers. The causal graph a masked crosscoder implies "may differ from the mechanism actually used by the model", so patching and similar interventions need care.
- Mechanistic faithfulness. The authors say it is unclear how to formalize, and unclear how far crosscoders meet it; crosscoders might exploit correlations that work on the training distribution but do not generalize.
- Diffing results are mixed. The authors say the fresh questions crosscoders open (how features form during training, whether retrained models learn the same features) rely on model diffing, where "our results there are a bit mixed" (Section 5.1).
The evidence about features in this update is correlational: decoder norms, directions and the contexts where features activate. No feature is ablated or steered, and claims about interpretability are qualitative and, by the authors' description, anecdotal or cherry-picked. The models are internal Anthropic models; only the diffing experiment names one (Claude 3 Sonnet).
Later work on the diffing result.
- Open replication. Kissane, robertzk, Conmy and Nanda posted an open replication on 27 October 2024 (LessWrong and the Alignment Forum), with a 16,384-latent crosscoder on Gemma 2 2B base and instruction-tuned models and hyperparameters they did not tune. The three groups and the high alignment of most shared decoders replicated, with some exceptions at very low or negative cosine similarity. Unlike the update, they found more chat-specific latents (about 225) than base-specific (about 60), and did not know whether that came from the model or their setup.
- Anthropic's follow-up. Insights on Crosscoder Model Diffing (February 2025) reports that exclusive features in such diffs tend to be dense and polysemantic, attributes this in toy models to competition for limited feature capacity, and proposes a small set of designated shared features with a lower sparsity penalty. It reads the near-equal counts of exclusive features in Anthropic's diffs as probably a property of its models or setup, citing the asymmetry in Kissane et al.
- A critique of the L1 objective. Minder, Dumas et al. ("Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning", NeurIPS 2025, arXiv:2504.02922v4) argue that the L1 sparsity loss used here can make a feature look exclusive to the fine-tuned model when the base model also represents it, and that the clean three-way split of relative norms is "an artifact of the loss function rather than a meaningful distinction" (Section 1). On Gemma 2 2B they find most of their L1 crosscoder's chat-only latents affected, report similar effects in Kissane et al.'s crosscoder (Appendix J), and find a BatchTopK crosscoder largely free of the problem (Section 3.1).
Research questions
- Are model-specific features really absent from the other model? The update identifies them by decoder norm alone. Applying Minder et al.'s latent scaling test, or ablating the feature's direction in each model, to the Claude 3 Sonnet diff would show how many of the four to five thousand per model survive.
- How do the L1-of-norms and L2-of-norms penalties compare on the same diff? The authors report that the L2 version finds only shared features but show no result. Training both on one model pair, and counting and interpreting exclusive features, would test the claim.
- What share of model-specific features is interpretable? The update reports a majority not immediately interpretable without a count. A blind rating of a random sample of exclusive and shared features would give the figure.
- Does cross-layer spread reflect superposition or amplification? The authors name both readings of gradual feature formation. Circuit analysis that traces where a cross-layer feature is computed, layer by layer, would separate them.
- Do crosscoder-based circuits match the model's mechanism? The authors warn that a crosscoder's implied causal graph may differ from the model's. Comparing interventions predicted from a masked crosscoder with interventions on the model itself, across many prompts, would measure the gap.
- What features account for the middle-layer component specific to larger models? Section 4.4 finds it and leaves it uninterpreted. Inspecting the features with the highest loadings on that component would test the reading that new representations emerge with scale.
How it connects
- Dictionary learning. Crosscoders extend the sparse autoencoders of Towards Monosemanticity and Mapping the Mind of a Large Language Model, and the superposition hypothesis of Toy Models of Superposition, which the update cites.
- Model diffing that followed. Stage-Wise Model Diffing (December 2024) fine-tunes one SAE through stages instead, and reports, without showing the crosscoder results, that on sleeper agents it was more sensitive than a crosscoder diff. Insights on Crosscoder Model Diffing (February 2025) revisits this update's diffing results and modifies the method, and A "diff" tool for AI (March 2026) applies crosscoders across models with different architectures.
- Circuit tracing. This note connects the update's strictly causal "cross-layer transcoders" to the cross-layer transcoder at the center of Circuit Tracing (March 2025), summarized in Tracing the thoughts of a large language model. Circuit Tracing cites "our previous results on crosscoders" for features whose outputs span several layers, and its author contributions credit this update's authors with the ideas leading to crosscoders. Tools for cross-layer transcoders were later released, as described in Open-sourcing circuit tracing tools.
- Earlier agenda. Olah's "Next Five Hurdles" essay in Circuits Updates, July 2024 names cross-layer superposition as a hurdle that the team's dictionary learning had no way to address. This update does not cite the essay; the connection is the library's.
References
- Lindsey, J., Templeton, A., Marcus, J., Conerly, T., Batson, J., and Olah, C. (2024, October 25). Sparse Crosscoders for Cross-Layer Features and Model Diffing. Transformer Circuits Thread. https://transformer-circuits.pub/2024/crosscoders/index.html (live page read 13 November 2026, including chart labels) record · 2 pages
- Kissane, C., robertzk, Conmy, A., and Nanda, N. (2024, October 27). Open Source Replication of Anthropic's Crosscoder paper for model-diffing. LessWrong / AI Alignment Forum. https://www.lesswrong.com/posts/srt6JXsRMtmqAJavD/open-source-replication-of-anthropic-s-crosscoder-paper-for (read in full) record
- Minder, J., Dumas, C., Juang, C., Chughtai, B., and Nanda, N. (2025). Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning. NeurIPS 2025. arXiv:2504.02922v4. https://arxiv.org/abs/2504.02922 (Sections 1 to 3.1 read for this note) record · 2 pages
- Ameisen, E., Lindsey, J., et al. (2025). Circuit Tracing: Revealing Computational Graphs in Language Models. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/methods.html (searched for its statements about crosscoders) record · 15 pages
- Works cited above only as the update describes them (Yun et al., 2021; Kissane et al., 2024, on SAE transfer between base and fine-tuned models) were not read for this note.