AnthropicP422026-05-15full textmechanistic interpretabilityattention headstransformer circuitspolysemanticitysparse features

HeadVis

Two titles, both real. The heading above is how the lab announced this work. The document actually behind it is titled “HeadVis: An Interactive Tool For Investigating Attention Heads” — everything below is read from that document.

Anthropic built and open-sourced HeadVis, a tool for visualizing attention heads across a full data distribution, revealing that heads' behavior often differs sharply from narrow-task expectations.

It gives interpretability researchers a concrete, open-source workflow and vocabulary for telling whether an attention head does one thing or several, which is still a live methodological gap.

Read the paper ↗All Anthropic papers →Compare its abstract →

R. Luger · Harish Kamath · Doug Finkbeiner · Purvi Goel · Adam Jermyn · Sam Zimmerman · Joshua Batson · Tom Conerly — meet the researchers →

How much of this do you want?

Orient me keeps four things: the abstract, the method, the claim↔evidence panel, and where it leads next. Everything adds constructs, the model table, every reported statistic, the discussion framing and the style moves. Switching hides nothing permanently and never changes what a section says — it only changes how many are on screen.

Abstract

Two readings, equal authority

How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.

This source carries no verbatim abstract.

Constructs

What this paper defines

Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.

Induction score

“The induction score is the average attention pattern of a head on (query, key) token pairs that correspond to the literal copying case of induction.”Fuzzy Induction

In plain terms: A number that measures how much a head does plain copy-the-earlier-token induction.

QK attribution

“QK attribution writes the attention score as a sum of (query feature, key feature) interactions.”Fuzzy Induction

In plain terms: A breakdown of why a head attends where it does, expressed as pairs of interacting features rather than raw numbers.

OV attribution

“OV attribution writes the head's output as a sum of (value feature, output feature) interactions: value features at the attended token, output features at the query token.”Fuzzy Induction

In plain terms: A breakdown of what a head actually copies or writes once it has decided where to attend, expressed as interacting features.

Polysemantic head

“The clean separation across Q, K, and O — together with the token-level patterns — makes us fairly confident this head is polysemantic: a single head implementing several unrelated behaviors.”A Polysemantic Attention Head

In plain terms: One attention head that is really doing several unrelated jobs at once.

K-composition

“This is K-composition : one head writes to a position, a later head reads from it.”Same-Set Suppression

In plain terms: One head leaves information at a token position that a later head then reads and uses.

Same-set suppression geometry

“A country's query vector has strongly negative cosine similarity (mean −0.47) with the key vectors of its own cities and near-zero (mean +0.01) with cities elsewhere.”Same-Set Suppression

In plain terms: The mechanism by which a head avoids attending from an item to other members of its own group, built directly into the geometry of its query and key vectors.

Attentional feature

“The goal is spiritually the same: learn monosemantic, sparse "attentional features" that faithfully reconstruct an attention layer.”Discussion

In plain terms: A hoped-for basic, single-purpose unit of attention computation, analogous to a sparse feature in an MLP layer.

Method

What they actually did

Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.

Metric-based head selection

A scatter plot of per-head metrics helps find heads sitting at interesting extremes.

→

Attention-pattern browsing

Browse the selected head's attention patterns across many dataset examples.

→

PCA clustering check

Project Q, K, and O activations at high-attention token pairs to check for distinct behavior clusters, as a hypothesis-generation step.

→

QK / OV feature attribution

Decompose the attention score and the head's output into interacting query/key and value/output features.

→

Targeted confirmation

Confirm the hypothesis with custom sequences, controlled sweeps, or by tracing composition with other heads.

The typical HeadVis workflow: find an interesting head by an extreme metric value, browse how it behaves across the dataset, check for clustered sub-behaviors with PCA, then use QK/OV feature attributions and targeted tests to pin down the mechanism.

Click any box to open it.

Built HeadVis to combine attention-pattern visualizations, quantitative per-head metrics, low-rank projections, and feature-level attributions into one tool for investigating attention heads.
Trace this step to the paper
“We built HeadVis to easily generate and test hypotheses about what an attention head does, using visualizations of attention patterns and head outputs, quantitative distribution metrics, low-rank component projections, and SAE feature attributions through QK and OV circuits.”Introduction
Used a configurable scatter plot of per-head metrics as the entry point for finding heads worth investigating, on the idea that heads at metric extremes tend to be the interpretable ones.
Trace this step to the paper
“The first component of the HeadVis UI is a scatter plot of heads, with both axes and point color configurable to any per-head quantity, like its layer, head index, or one of several computed metrics.”Fuzzy Induction
Identified a fuzzy induction head using the induction score metric, then ran QK and OV attributions on a top-activating example to decompose its behavior into feature-level interactions.
Trace this step to the paper
“To understand what makes this induction fuzzy, we run QK and OV attributions from the second token of gleam back to the newline after dream .”Fuzzy Induction
For a previously studied line-width head, collected Q, K, and O activations at its highest-attention (query, key) token pairs and ran PCA on each as a hypothesis-generation check for distinct sub-behaviors.
Trace this step to the paper
“We take the (query, key) token pairs with highest attention output norm, and collect the head's Q, K, and O activations at those pairs ... and run PCA on each set. If the top components separate into clusters, that's a hint the head has distinct behaviors worth investigating separately.”A Polysemantic Attention Head
Pinned down the exact boundary of one of the line-width head's behaviors by sweeping the year value in a fixed prompt rather than relying on attention patterns or attributions alone.
Trace this step to the paper
“sweeping the year in a fixed prompt shows the head attends strongly only for 1000–1986, a constraint neither the attention patterns nor the attributions surfaced.”A Polysemantic Attention Head
For the answer-selection head, ran QK and OV attributions across examples spanning the full data distribution, not just multiple-choice questions, to see whether a shared feature template underlies apparently different behaviors.
Trace this step to the paper
“Running QK and OV attributions on examples from the full distribution, a pattern emerges. The OV circuit always does the same thing: it copies content stored at the attended token, much like induction.”Answer Selection
Projected a known feature through the head's key and query weights to find which other features land nearby, as a virtual-weights test of whether disjoint-looking behaviors share an underlying mechanism.
Trace this step to the paper
“We project the "correct answer" feature through the head's key weights, and ask which other features land nearby in that projected space.”Answer Selection
Ran a systematic controlled test for the same-set suppression head by generating a batch of templated sentences, each with one item that does not belong to the group, and measuring how often the head attends to the outlier.
Trace this step to the paper
“To test this systematically we generated 185 sentences of the form "{five cities} are all cities in {country}", where four cities belong to that country and one is foreign.”Same-Set Suppression
Computed mean-subtracted cosine similarities between country query vectors and city key vectors to locate the suppression mechanism directly in the QK circuit's geometry.
Trace this step to the paper
“For each country, we compute the head's query vector on "The country {X}", subtract the mean across countries, and do the same for city key vectors on "The city {X} is well known".”Same-Set Suppression
Traced a K-composition chain by ranking all heads by their attention pattern on a specific (query, key) pair, to find which earlier head wrote the feature that the target head's key side reads.
Trace this step to the paper
“we looked for a head that attended between a and chikumbuso , and checked its OV circuit. That head writes exactly the Zambian features our head's key side reads!”Same-Set Suppression
Open-sourced the HeadVis frontend together with a specification of its interface to the two backend components, so others can connect it to their own models.
Trace this step to the paper
“We open-source the frontend along with a specification of the interface with the two backends.”Open Source HeadVis

The models under study

Exactly what was run, and how

Model	Developer	Temp	Effort / reasoning	Deployment	Other settings
Claude Haiku 3.5	Anthropic	not reported	not reported	unstated	A hosted demo covers only a subset of Haiku 3.5's heads, using cached QK/OV attributions for a few paper examples rather than a live backend.
Gemma 3 1B	Google	not reported	not reported	unstated	—

Source for Claude Haiku 3.5 settings

“We present a few case studies using HeadVis to interpret attention heads from Claude Haiku 3.5.”Introduction

Source for Gemma 3 1B settings

“HeadVis on all heads in Gemma 3 1B is here .”Open Source HeadVis

What they reported — and what they left out

The paper names Claude Haiku 3.5 and Gemma 3 1B as the models its demos and case studies cover, and notes the Haiku demo is restricted to a subset of heads with cached attributions, but it never reports temperature, sampling settings, context length, or deployment method for either model.

Results

The numbers they report

The line-width head's 'years' behavior has a sharp cutoff that isn't visible from attention patterns or attributions alone.

attends strongly only for years 1000–1986

See it in the paper

“sweeping the year in a fixed prompt shows the head attends strongly only for 1000–1986, a constraint neither the attention patterns nor the attributions surfaced.”A Polysemantic Attention Head

A country token reliably attends more to a foreign city than to its own cities in controlled test sentences.

98% of 185 templated sentences

See it in the paper

“The country token attends more to the foreign city than to any of its own 98% of the time.”Same-Set Suppression

A country's query vector is strongly anti-aligned with its own cities' key vectors, and roughly neutral toward other cities.

mean cosine similarity −0.47 (own cities) vs. +0.01 (other cities)

See it in the paper

“A country's query vector has strongly negative cosine similarity (mean −0.47) with the key vectors of its own cities and near-zero (mean +0.01) with cities elsewhere.”Same-Set Suppression

The suppression is a flip introduced by the head's own weights: a country and its own cities actually have positive similarity in the residual stream before the head's projections act on them.

mean cosine similarity +0.17 in the residual stream

See it in the paper

“in the residual stream a country and its own cities have positive cosine similarity (mean +0.17), so the head flips the sign.”Same-Set Suppression

Swapping the out-of-place country for the in-set country in a hand-built example collapses the suppressed attention edge.

attention to 'a' dropped from 0.87 to 0.08 (Zambia); 0.10 (Lusaka)

See it in the paper

“we replaced Ivory Coast with Zambia in the sequence, and attention to a dropped from 0.87 to 0.08; the same happened with Lusaka , the Zambian capital (0.10).”Same-Set Suppression

Swapping in any other African country (rather than the in-set one) leaves the attention edge largely intact.

median 0.27 across the other 51 African countries

See it in the paper

“Swapping in the other 51 African countries, attention mostly stays well above that (median 0.27).”Same-Set Suppression

QK and OV attributions in HeadVis are computed against a large shared feature basis.

10M-feature weakly causal crosscoder (WCC)

See it in the paper

“the features come from the same 10M-feature weakly causal crosscoder (WCC) used in .”Fuzzy Induction

Claim ↔ evidence

What they assert, beside what they showed

Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.

The claim

The line-width head is genuinely polysemantic: one head cleanly implementing three unrelated behaviors.

“The clean separation across Q, K, and O — together with the token-level patterns — makes us fairly confident this head is polysemantic: a single head implementing several unrelated behaviors.”

The evidence

“For this head, all three PCAs separate the behaviors into distinct clusters in the first few principal components.”

A Polysemantic Attention Head

Mind the gap: The authors flag this gap themselves: a direct test of polysemanticity would split the head into separate heads and show they reproduce the original, which they explicitly defer to future work rather than having done here.

The claim

The answer-selection head implements a general 'about to repeat marked content' behavior rather than a multiple-choice-specific mechanism.

“We found that "answer selection" was better described as a more general head behavior that helps the model repeat previous context by identifying what entity is about to be mentioned.”

The evidence

“The features that align most strongly are not specifically about multiple choice, but they are other instances of "this is the relevant content" — these features are near-orthogonal in the residual stream and the head's K projection maps them close together.”

Answer Selection

Mind the gap: The authors immediately qualify the claim's confidence: "Our best guess is that this head is monosemantic, but we aren't fully convinced... we don't have a crisp enough hypothesis for this head's behavior to validate it."

The claim

Same-set suppression — a head suppressing attention between an item and other members of its own group — is a real mechanism, not just an artifact of the hand-built test sentences.

“The geometry and the 98% result on custom sentences are evidence that same-set suppression is real.”

The evidence

“A country's query vector has strongly negative cosine similarity (mean −0.47) with the key vectors of its own cities and near-zero (mean +0.01) with cities elsewhere.”

Same-Set Suppression

Mind the gap: The authors state directly that this evidence does not extend to a full account of the head's ordinary dataset behavior: "going back to this head's dataset examples in HeadVis with the pattern in mind, it isn't clean... not a full account of the place examples either."

The claim

PCA clustering of a head's Q/K/O activations is a useful hint for whether the head has distinct sub-behaviors.

“If the top components separate into clusters, that's a hint the head has distinct behaviors worth investigating separately.”

The evidence

“In this head, the PCA does not cluster activations, so it's not useful for understanding if this head implements multiple behaviors.”

Answer Selection

Mind the gap: The authors explicitly bound how much weight this method can bear: "PCA only projects activations onto a small subspace that captures the variance in activations, but different attention behaviors could seem separable in high variance dimensions but not actually impact the attention score... In principle a head could be structured to make PCA maximally misleading."

The claim

A head's behavior on the full data distribution rarely matches what a narrow task or dataset suggests.

“One theme emerges across both prior work and our own: a head's behavior on the full distribution rarely matches what a narrow task suggests, and the difference isn't always easy to characterize.”

The evidence

“found a head implementing "succession" (predicting the next element in an ordinal sequence) on a custom dataset. When they analyzed that head on a more general dataset, they noticed distinct behaviors including copying, comparison, and acronym production.”

Introduction

Mind the gap: This synthesizing claim is built from three cited prior findings plus the paper's own two case studies (line width, answer selection) — an illustrative pattern across a handful of examples rather than a systematic survey of heads.

Discussion & after

How they frame it, and what they want next

Their framing

The Discussion frames open problems in attention interpretability by direct contrast with the more mature MLP/transcoder case, naming four concrete named obstacles and pointing to each case-study head as a worked example of one obstacle rather than a solved problem. The tone stays optimistic about eventual tractability while being explicit that no decomposition method for attention analogous to MLP transcoders currently exists.

Register: Concrete and confident about the specific numeric results from their own case studies (cosine similarities, attention values, percentages), but consistently tentative when generalizing toward a full theory of attention decomposition, with frequent 'we suspect,' 'we don't know,' and 'our best guess' framing.

Where they hedge

“We don’t know how to do the analogous thing for an attention layer.”Discussion

“Our best guess is that this head is monosemantic, but we aren't fully convinced.”Answer Selection

“We have not determined which our line width head is, but a decomposition method must handle both.”Head Polysemanticity

“We suspect some attentional features are spread across multiple heads, but we have no confirmed example in a real LLM.”Attention Superposition

What they say it means

A single attention head can be polysemantic without weight superposition, unlike a single MLP neuron, so decomposition methods built for MLPs will not transfer to attention unchanged.
the paper’s words
“We also want to make a theoretical point explicit: an attention head can be polysemantic without superposition.”Head Polysemanticity
Because some attentional features are likely high-rank, any successful decomposition method will need to interpret multi-dimensional objects rather than single directions the way rank-1 MLP features work.
the paper’s words
“Some attentional features are likely high rank, making them much harder to interpret than rank-1 MLP features.”Discussion
Existing rank-1 decomposition methods target polysemanticity and superposition but still leave the high-rank interpretability problem on the QK side unsolved.
the paper’s words
“These approaches also inherit the high-rank interpretability challenge, since their QK circuits are still high rank.”Existing Decomposition Approaches

What they call for next

Researchers who want the full interactive backend should use Claude to implement it against the published frontend/backend interface specification.
the paper’s words
“We suggest pointing Claude at the repository to implement the backend pieces for your setup.”Open Source HeadVis
Anyone researching attention should spend a few hours exploring heads in HeadVis on an open model to build intuition.
the paper’s words
“we'd encourage anyone researching attention to spend a few hours in HeadVis on an open model.”Attention Biology
Future work should extend HeadVis-style investigation to heads that attend over long contexts.
the paper’s words
“We'd be excited to see work on HeadVis for long contexts.”Attention Biology
Future work should try splitting a polysemantic head like the line-width head into separate heads that each isolate one behavior and check that together they reproduce the original.
the paper’s words
“We'd be excited to see a single polysemantic head split into monosemantic units.”Head Polysemanticity
Existing rank-1 decomposition methods should be tested against a small number of early-layer heads that HeadVis already makes understandable, as an easy way to gauge their performance.
the paper’s words
“We'd be excited to see these decompositions applied to a small number of early-layer heads that are understandable with HeadVis”Existing Decomposition Approaches

Limitations they state

“Custom-sequence input is not included in the open-source release; see Open Source HeadVis .”Same-Set Suppression

“This functionality is not included in the open-source release; see Open Source HeadVis .”Same-Set Suppression

“Some features we found useful internally are not included in this open source release:”Open Source HeadVis

“So far we have only studied individual attention edges (a single key and query token) at a time, so any logic that depends on competition among keys would be invisible to us.”Attention Biology

“We have not determined which our line width head is, but a decomposition method must handle both.”Head Polysemanticity

“Past toy models, including our own , produce forms of attention superposition, but it isn't clear they capture a phenomenon that occurs in real heads.”Attention Superposition

For your own writing

Moves worth stealing

Structures the entire Discussion as an explicit analogy-by-contrast to a better-understood case (MLP superposition and transcoders), naming precisely where the analogy breaks down instead of just gesturing at similarity.

“It's useful to frame the open problems in the study of attention by contrast with MLPs.”

Names its own biggest open obstacles as a short checklist before discussing each one, giving the reader a map of the argument before the detail arrives.

“Four obstacles stand in the way:”

Closes a multi-case-study section by explicitly naming which tool view solved which case, turning four separate anecdotes into one legible method table in prose.

“Each step in this section used a different HeadVis view — top-token rankings to see what dominates, custom sequences to test a guess, QK attributions to read a confusing example, head-ranking on a (query, key) pair to trace composition.”

States its own confidence level in plain, hedged language directly beside the evidence, rather than letting tone alone signal how strong the finding is.

“Our best guess is that this head is monosemantic, but we aren't fully convinced.”

Connected

Where else this leads

Same people

Characterizing interference weights in a tiny language model Anthropic
shares Joshua Batson
Verbalizable Representations Form a Global Workspace in Language Models Anthropic
shares Joshua Batson
Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations Anthropic
shares Joshua Batson
Teaching Claude Why Anthropic
shares Adam Jermyn

Same territory

Characterizing interference weights in a tiny language model Anthropic
mechanistic interpretability
Verbalizable Representations Form a Global Workspace in Language Models Anthropic
mechanistic interpretability

Published alongside it

The nearest publications in time, across all three labs.

Model Spec Midtraining: Improving How Alignment Training Generalizes Anthropic
2026-05-15
Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations Anthropic
2026-05-15
SLEIGHT-Bench: Finding Blind Spots in AI Monitors Anthropic
2026-05-15
Teaching Claude Why Anthropic
2026-05-15

What this page was built from

Working from a full plain-text extraction of the published Transformer Circuits web page, including Related Work, Discussion, and Open Source sections; interactive figures are represented only by their captions and surrounding prose, and several inline citations were stripped of their author names during text extraction (e.g. 'found a head implementing succession...' with no named subject), so those specific prior-work attributions are not recoverable from this text.