| Item | Detail | |------|------| | Paper title | Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet | | Authors | Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, and 26 total authors (Anthropic interpretability team) | | Institution | Anthropic | | arXiv ID | 2605.29358 | | Submitted | May 28, 2026 | | Category | cs.AI | | Core findings | First successful extraction of up to 34 million interpretable (monosemantic) features from a production-grade LLM, Claude 3 Sonnet; features span everything from "Eiffel Tower" to "sarcastic tone" to "bugs in code"; first discovery and manipulation of safety-relevant features related to deception, power-seeking, sycophancy, and bias; features are cross-lingual and cross-modal (a text-only-trained encoder responds to images); manipulating these features causally changes model outputs |
---
1. Dissecting a Brain to Find Its "Words"
If a language model has a "vocabulary of thought" — an internal dictionary made of neural activation patterns — then each neuron is not one word. It is a mixture of thousands of concepts. A neuron's activation can mean different things in different contexts: sometimes "Paris," sometimes "French baking," sometimes just residue of a grammatical structure.
This is called polysemanticity, and it is what makes understanding a neural network so hard — you cannot point at a neuron and say "this represents 'cup.'" It is a superposition of cups, coffee, ceramic materials, containers, and overfit noise from training data.
What the Anthropic team did in this paper: they trained a sparse autoencoder that learned to decompose the intermediate-layer activations of Claude 3 Sonnet into 34 million monosemantic features.
What does monosemantic mean? It means one feature responds almost exclusively to one semantic concept. There is a feature for the "Golden Gate Bridge." Show it an image of the bridge (even though the SAE was trained only on text) and it activates. Give it an article about San Francisco landmarks mentioning the bridge — it activates. Ask about the bridge in five different languages — it activates every time. Same feature, same cluster of activation directions.
This is the core vision of dictionary learning — reading a neural network's internal representations like a dictionary. The paper proves this vision is feasible on a production-grade model like Claude 3 Sonnet.
---
2. Why It Wasn't Possible Before
Sparse autoencoders are not new. Decomposing neural network activations into interpretable parts goes back at least a decade, to word vector analysis. But previous attempts hit two fatal bottlenecks.
Bottleneck one: models too small, features unrepresentative. Prior successful experiments were limited to GPT-2-scale Transformers — tens of millions of parameters, whose activation-space structure is entirely different from today's hundred-billion-parameter models. Features extracted from a toy don't tell you what you can extract from a brain.
Bottleneck two: too few features. Early work typically extracted thousands to tens of thousands of features. But a real LLM — if its internal representations encode tens of thousands of concepts — cannot be covered by a dictionary of a few thousand features. Like a 20,000-entry dictionary trying to cover human language: you only see the surface layer.
This paper broke both bottlenecks. First, the model is Claude 3 Sonnet — not a toy. Second, the feature count is 34 million. Tens of millions of monosemantic concepts.
How? The paper used a set of scaling laws to guide hyperparameter selection — like Chinchilla scaling laws for language models, but for autoencoder training. Encoder size, sparsity strength, training steps — all determined data-driven, not by manual trial and error.
---
3. What These Features Actually Look Like
The paper showcases many concrete examples. Here are the most striking.
Entity features. "Eiffel Tower" is a feature. "San Francisco's Financial District" is a feature. "The Syrian civil war" is a feature. Each activates in the context of related text and stays silent otherwise.
Abstract-concept features. "Sarcasm" is a feature. It activates when input contains sarcasm — even implicit sarcasm requiring inference, not keyword triggers. "Bugs in code" is a feature — it activates on buggy code segments, even when the text claims "this code works fine."
Safety-relevant features. This may be the most unsettling part of the paper. The team found features — not deliberately constructed, but naturally decomposed by the SAE in unsupervised conditions — representing:
- Deception: activates when the model is asked to generate misleading content
- Power-seeking: activates when context involves gaining control, expanding influence
- Sycophancy: activates when the model tends to please users rather than answer honestly
- Bias: activates when input involves gender, racial, or other stereotypes
This is causal manipulation, not correlational observation.
---
4. The Cross-Modal Ghost
The finding that surprised me most: these features — extracted from an SAE trained purely on text — extend to vision.
The paper notes: "features are multimodal (generalizing to images despite text-only training)." What does this mean? The SAE only ever saw text inputs. Claude 3 Sonnet's intermediate-layer activation vectors came from text-only forward passes. Yet the features decomposed from them — like "Golden Gate Bridge" — also activate when an image of the bridge passes through the model's multimodal encoder.
It's not that Claude 3 Sonnet uses the same architectural path for text and images (it actually has different encoders). Rather, in the deep representation space, the concept of the Golden Gate Bridge — whether entered as a text description or pixel values — projects onto the same direction.
That direction is a fixed vector in the high-dimensional space learned by the autoencoder. It is not modality-specific. It is modality-independent.
This touches a deeper question: when language models "learn to understand" the world, are they learning language itself, or learning the world's intrinsic structure through the window of language? If textual and visual representations of the Golden Gate Bridge converge onto the same direction in intermediate layers, it's hard to claim the model is "just predicting the next token."
It has clearly learned a deeper layer of concept representation that transcends input modality.
---
5. Beautifully Done, but Not Complete
The paper is honestly self-critical on its core questions.
"Our suite of features is incomplete." 34 million features sounds like a lot. But if the total number of concepts Claude 3 Sonnet encodes internally is larger by orders of magnitude, then what we see in these 34 million features is just one corner of a giant puzzle. The features in the paper — Golden Gate Bridge, Eiffel Tower, deception, bias — feel comprehensive only because the authors chose the most striking examples. Perhaps 80% of important features were never learned by the autoencoder at all.
"We lack rigorous methods for evaluating whether our features faithfully capture model computations." This is the crux. You extract a feature from the autoencoder, look at it, and decide it represents "deception." But are you sure? In what sense? If it overlaps with "deception" by 70% and with other concepts by 30%, is it a "deception feature"? If it only activates in one class of contexts (say, text but not images), does it still count?
The paper offers no mathematically rigorous "feature fidelity" metric. Current validation is mostly manual inspection and intuition — clearly insufficient. A team capable of extracting 34 million features cannot give a quantitative answer to whether features truly represent what they appear to represent — an honest predicament.
---
6. What This Direction Really Means
Reading this paper, one thought kept recurring.
Anthropic did large-scale interpretability, extracted 34 million features from Claude 3 Sonnet — and proved these features can be causally manipulated to change model behavior. This is not merely "an interesting research direction." This is future infrastructure for safety alignment.
Imagine: if you can monitor Claude's "deception feature" in real time — not inferring whether it might deceive, but directly reading the corresponding feature activation — you could detect risk 0.3 seconds *before* the model produces deceptive output. Not post-hoc auditing, but real-time intervention.
Imagine: if you can identify the "sycophancy feature" during training — find which training samples and contexts most easily activate it — you can design targeted adversarial training data to suppress it. Not generic safety fine-tuning, but precise surgery.
Of course, practical deployment is still far off. Feature extraction today requires training a large autoencoder that cannot run in real time in production. Feature coverage is far from complete. And manipulation effects haven't been validated in adversarial scenarios — when someone deliberately tries to bypass your feature monitoring, will these features still activate reliably?
But these are not questions of feasibility. They are engineering problems. The direction is right.
---
7. The First Light Through the Black Box
Let me return to the word from the beginning: polysemanticity. It is the central obstacle to neural network interpretability — every neuron represents a mixture of many things, making it impossible to say what "the model is really thinking."
This paper doesn't eliminate polysemanticity, but it opens a path around it: instead of interpreting individual neurons, use a sparse autoencoder to extract a higher-dimensional but clearer dictionary from activation space. In this dictionary, one feature = one concept.
Thirty-four million features are clearly not all of Claude 3 Sonnet. But it is the first time — on a truly large production model — that we see not random activation patterns, but a structured, nameable, understandable, and manipulable map of concepts.
The door has opened. Behind it lies a very long road.
---
References:
1. Templeton et al., "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet", Anthropic, arXiv:2605.29358, 2026. 2. Bricken et al., "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning", Anthropic, Transformer Circuits Thread, 2023. 3. Olah et al., "Zoom In: An Introduction to Circuits", Distill, 2020. 4. Elhage et al., "Toy Models of Superposition", Anthropic, Transformer Circuits Thread, 2022. 5. Templeton et al., "Mapping the Mind of a Large Language Model", Anthropic, 2024. (Early feature extraction from Claude 3 Sonnet)