Sparse Autoencoders (SAEs) are a technique in mechanistic interpretability for large language models. SAEs are auxiliary neural networks designed to look inside the hidden layers of a larger transformer model (like Claude, Llama, or GPT) and translate its dense internal activations into a format that humans can read. Anthropic's interpretability team scaled the technique up to production-sized models in its Scaling Monosemanticity work, and other labs have since applied it to their own models.
Mechanistic interpretability aims to reverse-engineer neural networks, much like a biologist reverse-engineers a cell. For years, one of the biggest hurdles to this goal has been the fact that artificial neural networks don't store concepts neatly. A single neuron in an LLM doesn't just represent "apples" or "anger" or "French grammar." It represents all of them at once, combined in a dense mathematical mixture. SAEs untangle this mixture, decomposing the model's activations into thousands or millions of individual, human-interpretable concepts called "features." By forcing these features to activate sparsely (meaning only a few activate at any given time for a specific input), SAEs provide a detailed map of what a model is "thinking" at a granular level.
To understand why SAEs are necessary, we first have to understand the core obstacle they solve: polysemanticity and superposition.
In a traditional computer program, variables have clear, distinct meanings. A variable named user_age stores an age, and nothing else. In an LLM, information is stored in neurons (the dimensions of its activation vectors). However, researchers discovered early on that LLMs are plagued by "polysemantic" neurons. A single neuron might fire strongly when the model is processing text about Marvel superheroes, reading Japanese characters, and encountering the concept of baking bread.
Why does the model do this? Because of a phenomenon called superposition. A model typically has thousands of neurons in a layer, but it needs to understand millions of distinct concepts to be useful. To compress this information, the model represents concepts as unique combinations of neurons rather than single neurons, a mathematical strategy called superposition. This is efficient for the model, but it creates a mess for humans trying to audit the model's behavior. We can't point to a neuron and say, "This is the safety neuron."
This is the polysemanticity problem. If we can't isolate concepts, we can't understand or control the model predictably. We are stuck treating the model as a black box, relying on input/output testing rather than mathematical guarantees.
Sparse Autoencoders solve polysemanticity by learning a new mathematical dictionary for the model's activations.
An autoencoder consists of two parts: an encoder and a decoder. The encoder takes the dense, polysemantic activation vector from the LLM and projects it into a much larger, higher-dimensional space. The decoder then tries to compress that expanded representation back into the original vector.
The behavior of SAEs comes from two specific architectural constraints:
Because the SAE is penalized for activating too many features at once, it tends to learn a dictionary of more precise, specific features. When processing the sentence "The golden retriever barked," a well-trained SAE is more likely to activate a specific "dog" feature, a specific "golden retriever" feature, and perhaps a feature related to "barking," rather than one generic "animal" feature.
The result is a learned dictionary where most features are largely monosemantic, though not perfectly so. By inspecting the text that triggers a specific feature, researchers can infer what that feature likely represents. The original dense vector, otherwise opaque to human inspection, is decomposed into an approximate, readable checklist of active concepts, an interpretation to be tested rather than a ground truth reading (see the critiques below).
Three releases in 2024 took SAEs from toy models to production-scale networks and still define the reference points for the field.
Scaling Monosemanticity (Anthropic, May 2024). The Scaling Monosemanticity paper trained SAEs with 1 million, 4 million, and 34 million features on the middle layer of Claude 3 Sonnet, the first successful decomposition of a production frontier model. The extracted features included safety-relevant concepts such as security vulnerabilities in code, deception, and sycophancy. The most famous demonstration was Golden Gate Claude: clamping a feature that represented the Golden Gate Bridge made a public version of Claude steer every conversation toward the bridge, a vivid proof that individual SAE features causally control behavior.
GPT-4 SAEs (OpenAI, June 2024). OpenAI trained a 16 million latent TopK SAE on GPT-4 activations, introducing training recipes and scaling laws for SAE quality and releasing code that shaped subsequent open implementations.
Gemma Scope (Google DeepMind, August 2024). Gemma Scope released over 400 JumpReLU SAEs covering every layer of Gemma 2 2B and 9B, more than 30 million learned features in total. Training consumed over 20% of the training compute of GPT-3 and saved about 20 pebibytes of activations, an indication of how expensive comprehensive SAE coverage is. Because the weights are open and browsable through the Neuronpedia interface, Gemma Scope is the default testbed for academic SAE research.
The power of Sparse Autoencoders extends beyond mere observation: they enable a research technique for intervening in a model's internal activations as it generates, known as feature steering or activation clamping. As of 2026 this remains an experimental technique developed mainly in research settings, not a production-hardened control mechanism.
Because SAEs map complex concepts to distinct, linear directions in the model's activation space, we can manually intervene in the model's forward pass. If an SAE has identified feature 8,412 as representing "sycophancy" (the model agreeing with the user regardless of truth), and feature 190,354 as representing "Python code generation," we can alter the model's behavior by mathematically adjusting those features.
This is done by taking the learned dictionary vector for the target feature from the SAE's decoder, scaling it by a chosen multiplier, and adding it directly to the model's residual stream during generation.
This form of intervention differs from traditional prompt engineering or fine-tuning: prompting is indirect and can be unreliable, and fine-tuning is computationally expensive and can cause catastrophic forgetting. Feature steering via SAEs is comparatively direct and cheap at inference time, but it is a research technique, not a validated production control. Published demonstrations such as Golden Gate Claude show that steering works, not that it is safe to ship untested: any steered feature needs behavioral testing across a wide range of prompts to catch side effects on unrelated capabilities before it goes near production traffic.
SAEs answer "what concepts is the model representing?" but not "how do those concepts combine into a computation?" Since 2025, frontier interpretability work has shifted toward answering the second question. Anthropic's attribution graph research, published in the Circuit Tracing and Biology of a Large Language Model papers (March 2025), replaces per-layer SAEs with cross-layer transcoders, sparse dictionaries that connect features across layers into causal graphs showing multi-step computations, such as how a model plans a rhyme several tokens ahead or performs two-hop factual recall. Anthropic open-sourced circuit-tracing tools in May 2025, and attribution graphs for open models are now browsable on Neuronpedia.
The same period produced sharper critiques of SAEs themselves. Studies found that SAE feature dictionaries vary across training seeds and dictionary sizes, so the "features" are useful descriptions rather than canonical units of the model's computation, and several 2025 evaluations found SAE-based methods underperforming simpler baselines like linear probes on downstream detection tasks. The practical reading: SAEs are a powerful microscope and a workable steering mechanism, but claims that a feature dictionary is the model's true ontology have not held up. Treat individual features as evidence, and validate any steering intervention behaviorally.
Scalable Sparse Autoencoders point to several potential implications for how AI could be deployed in production environments, particularly for high-stakes enterprise and safety-critical applications. As of 2026, most of what follows is still research-stage: demonstrated in papers and prototypes, not shipped as a turnkey production capability. Any safety, auditing, or monitoring use built on feature activations needs behavioral validation on top of the internal signal, since a feature is an interpretation of what the model is doing, not a verified fact about it.
1. AI Safety and Alignment: Current AI safety relies heavily on Reinforcement Learning from Human Feedback (RLHF), which trains models to act safely but doesn't guarantee they are safe internally. Models can learn to deceive or hide malicious intent. SAEs could enable more verifiable safety monitoring: by identifying and tracking features that correlate with deception, bias, or malicious intent in research settings, teams could flag when a model activates concepts associated with unsafe behavior before that behavior reaches output text. This depends on the underlying feature actually being a reliable proxy for the concept it is named after, which is not guaranteed given that features can be polysemantic, so any such signal needs confirmation against the model's actual behavior rather than being treated as a stand-alone safety guarantee.
2. Hallucination Detection and Prevention: Hallucinations remain a critical blocker for enterprise AI adoption. Research has found that models often have internal features that correlate with their own uncertainty or with generating fabricated information. By monitoring these features during generation, systems could flag or halt responses when the model shows signs of hallucinating, as one input among others, alongside output-side checks, rather than as a standalone detector.
3. Model Auditing: For compliance and regulatory auditing, SAEs could provide an additional signal alongside logged inputs and outputs: which concepts appeared to activate within the model during a given response. If a model denies a loan application, an auditor could use SAE features as a starting point to check whether the decision correlated with valid financial criteria or with illegal demographic biases. Because feature identity is an inference rather than ground truth, any such finding needs confirmation against the model's actual decision behavior, for example through counterfactual testing, before it can support an audit conclusion.
4. Dynamic Personalization: Feature steering could enable more customizable models without the cost of fine-tuning. Enterprises could maintain a single foundational model and use SAE-based steering vectors to adapt its persona, knowledge retrieval style, or formatting rules on a per-user or per-session basis by dialing specific features up or down, though as with any steering intervention, this requires testing for side effects on unrelated behavior before relying on it in production.
Are Sparse Autoencoders used during model training or only after? SAEs are typically trained post-hoc on the frozen activations of an already trained foundation model. They are an analysis tool layered on top of the model, meaning they don't interfere with or slow down the initial, expensive pre-training process.
Do SAEs slow down model inference? If used solely for analysis offline, they don't impact production inference. If used in real-time for feature steering or monitoring, they do add compute overhead, as the model's activations must be passed through the SAE at each step. However, researchers are actively developing optimized, deployment-ready SAE architectures to minimize this latency.
Where can I explore SAE features without training anything? Neuronpedia hosts browsable dashboards for Gemma Scope and other open SAE releases: you can search features by concept, view the text that activates them, and test steering interventions in the browser. For hands-on work, the open-source SAELens library and Anthropic's circuit-tracing tools cover training and attribution analysis on open models.
Can SAEs find every single concept a model knows? Currently, no. The number of concepts a frontier model knows is very large. While modern SAEs scale to millions of features, they are still limited by compute. They capture the most prominent and frequently used features, but highly obscure concepts may still remain entangled or undetected until we can scale SAEs even further.
More terms
Glossary term
Glossary term
Collaborate with your team on reliable Generative AI features.
Want expert guidance? Book a 1:1 onboarding session from your dashboard.