FILED: 2026-05-10 · TOPIC: Cornerstone · 8 MIN · BERLIN
Anthropic's Mechanistic Interpretability Work at Year Five — What Shipped
A working census of what the lab's mechanistic-interpretability program has actually published since the transformer-circuits thread opened in 2021. Sparse autoencoders, feature dictionaries, attribution graphs, and the open questions the published work has not yet closed.
The mechanistic-interpretability research program at Anthropic is, at the time of writing, the closest thing in the field to a sustained five-year arc — a continuous public publication record long enough that the criticism “we have no idea what is happening inside these models” can no longer be answered by linking the same three blog posts from 2022. The program has shipped enough that an honest survey is possible. This is an attempt at one. We have read the published material directly. We have not characterised internal positions the lab has not put on the record.
The thread, briefly
The program’s public spine is the Transformer Circuits Thread, a continuously updated set of short, technical write-ups that share a common methodology: identify a behavior inside a trained transformer, trace it to a specific computational circuit at the level of attention heads or MLP neurons, and publish the circuit in a form another researcher can replicate against an open-weight model. The 2021 framework paper, “A Mathematical Framework for Transformer Circuits,” is the methodological anchor. Most of what followed has been incremental: better techniques, larger models, more specific circuits, harder questions.
A reader new to the area can read the entire thread in a long weekend. We recommend it. The thread reads as the working notebook of a sustained research program rather than as a publication strategy, which is unusual.
What has actually shipped
Five publication lines are worth recording.
Induction heads and in-context learning. The earliest substantive circuit-level finding, published in “In-context Learning and Induction Heads” (2022), identified a specific two-head pattern that allows transformers to perform a primitive form of in-context generalization. The result is small in scope and load-bearing in implication: it suggests that a behavior previously characterized as emergent at scale is, in fact, implemented by an identifiable circuit that arises predictably during training. Subsequent work has located the same pattern across model families. The methodology — name a behavior, find the circuit, show it generalizes — is the template for most of what followed.
Superposition and the toy-models program. The 2022 paper “Toy Models of Superposition” laid out the central obstacle the program would spend the next three years working around: features inside a transformer are not stored one-per-neuron but in superposition, multiple features sharing the same neuron and decoded through learned weight patterns. The implication for interpretability is direct. If the unit of feature is not the neuron, then a methodology that reads neurons directly will read noise. The superposition paper was, in retrospect, the program’s “we know what the problem is” moment.
Sparse autoencoders and the dictionary-learning turn. The 2023–2024 sequence on sparse autoencoders is the program’s principal technical contribution to date. The technique trains a separate, much wider, sparse network to decompose a transformer’s activations into a dictionary of features, each of which has a more interpretable meaning than the raw activation. “Towards Monosemanticity” (2023) and “Scaling Monosemanticity” (2024) report the technique applied at successively larger model scales. The 2024 paper applied dictionary learning to Claude 3 Sonnet’s residual stream and identified a feature catalog in the millions, with named features ranging from concrete (the Golden Gate Bridge, immune-system metaphors) to abstract (deception, code that contains security vulnerabilities, sycophancy). The catalog is partial. The point is that it exists.
Circuit-level interventions. The “Scaling Monosemanticity” work moved from passive identification to active intervention: features identified in the dictionary can be clamped to higher or lower values during inference, and the model’s behavior changes in directions consistent with the feature’s identified meaning. The most-cited single demonstration was the “Golden Gate Claude” experiment, in which clamping the Golden Gate Bridge feature produced a model that returned the bridge in roughly every context. The result is small as a product demo. As an interpretability result it is large: the feature identified by the technique is causal, not merely correlational, with respect to the model’s behavior.
Attribution graphs and tracing. The 2024–2025 work on attribution graphs and circuit tracing takes the next step: given a specific model output, trace the feature-level computation that produced it. Working papers in 2024 introduced attribution methods that connect intermediate features to output features through the model’s MLP and attention components. Subsequent work has applied the technique to specific behaviors — sycophancy, refusal patterns, in-context arithmetic — and produced traces that name the contributing features at each layer. The traces are not yet complete in any strong sense. They are, however, audit artefacts of a kind the field did not have three years ago.
What the published work has not closed
The honest census requires naming the open questions.
The first is coverage. The feature dictionaries published to date cover a non-trivial fraction of the residual-stream variance in the studied models. They do not cover all of it. The fraction the dictionaries miss is the fraction interpretability cannot yet read. The lab has been explicit about this; the published papers carry the metric. The number is improving over publication generations. It is not yet “we read the model.”
The second is scale. The published interventions have been demonstrated at the scale of Claude 3 Sonnet. The publicly deployed frontier models, as of mid-2026, are larger. The dictionary-learning technique scales with effort; whether it scales gracefully to the next generation of frontier scale is a research question.
The third is agentic behaviour. The dictionaries have been built against single-turn behaviors. The interesting auditability questions for the agentic systems this publication covers — what does the orchestration layer do, what tool calls did it dispatch, what was the policy enforcement that conditioned the dispatch — are not directly answered by feature-level interpretability of a single forward pass. The bridge between mechanistic interpretability and orchestration-layer audit is not yet built. We have not seen a published proposal for what that bridge looks like.
The fourth is deception-detection. The feature dictionaries include features the lab has identified as related to deceptive behavior. The published work is careful about the claim: a feature whose activations correlate with deceptive outputs is not the same as a deception detector. The work on what the gap looks like, and what closing it would require, is ongoing.
What the program has bought the field
It is worth naming what the five-year arc has bought the field, separately from what is left to do.
The most useful purchase is the methodology. The transformer-circuits methodology — identify the behavior, find the circuit at the level of features, show the circuit generalises — is now the working template for a generation of academic and industrial interpretability research. The published material is the reference against which other groups’ results are compared. The convention is now: if you claim to interpret a model, you publish the dictionary, you publish the intervention, and you publish the metric that lets the next reader check the fraction of variance you have read.
The second purchase is the publication norm. The thread is open. The papers link the artefacts. The dictionaries are increasingly accompanied by browsers a reader can use to inspect the features directly. The model whose features are reported is named. The variance fraction is reported. This is the publication regime an audit-grade interpretability practice would require, and the program has been one of the principal contributors to making it standard.
The third purchase is the rhetorical floor. The field can no longer be argued at the level of “we have no idea what is happening inside these models.” A specific answer has been published. The answer is partial. The publication has been continuous enough that the partial answer cannot be dismissed as a vendor’s marketing claim.
What an auditability-oriented reader can do with the work
This publication’s beat is the auditability of AI systems in deployment. The mechanistic-interpretability work above is research-grade. The honest read on what an auditability-oriented reader can use the work for today is the following.
The work supplies a vocabulary an auditor can use to ask sharper questions of an opaque system’s operator. “What features in your published interpretability work activate when the system produces an output of class X” is a question a procurement-side auditor can now plausibly ask of a frontier-lab API vendor. The lab cannot fully answer it. The fact that the question is now well-formed is the contribution.
The work supplies a benchmark a buyer can use to evaluate competing vendors’ interpretability claims. A vendor that does not publish dictionaries, does not publish interventions, and does not publish coverage metrics is not at the published baseline. A vendor that does is at the published baseline. The buyer can ask which.
The work supplies a research direction the field can converge on. The dictionary-learning approach is not the only possible direction for interpretability research, but it is the direction with the longest continuous published record and the most replicable methodology. Other approaches — attention-based interpretability, probing-classifier methods, behavioral interpretability — continue to be published. The dictionary-learning line is currently the central one.
What we will be watching
The publication’s view, after five years of public publication, is that the program has earned its place in the auditability conversation. The next two years will determine whether the methodology scales to frontier deployment or stays a research-grade artefact.
We will be watching three publication lines specifically. The first is agentic-layer interpretability — whether the dictionary-learning technique extends from single-pass behaviors to multi-step orchestrated behaviors of the kind the agentic-stack vendors deploy. The second is deception-related features — whether the work on features correlated with deceptive output produces an intervention with high enough precision to be useful as an audit primitive. The third is buyer-facing tooling — whether the published browsers and inspection tools mature into something a procurement-side auditor can use without a dedicated research engineer translating each report.
The five-year arc is honest work. The next five will be the test of whether the work converts from a research program into the published infrastructure of audit-grade AI procurement. We will keep reading the thread.
Methodology. This piece is sourced from the public Transformer Circuits Thread, from Anthropic’s published research index, and from the lab’s referenced conference papers and working notes. Where we have characterized a result we have linked the underlying paper. Where the open questions named above are the lab’s own published self-criticism we have noted that. We have not invented any results.