Transcoder

A transcoder learns to decompose the dense, polysemantic output of a target component (e.g., an MLP layer) into a sparse linear combination of interpretable feature vectors. Unlike a sparse autoencoder applied to activations, a transcoder is trained to reconstruct the output of a function, not its input, using a sparsity penalty on the latent coefficients. This forces the model to represent the component's computation as a sum of a small number of distinct, meaningful features.

A trained transcoder is designed to functionally approximate the original model component it was trained on. Once trained, the transcoder can be swapped in for the original MLP layer or attention block. The model's downstream computation then operates on the transcoder's sparse, interpretable feature activations instead of the original dense, opaque activations, enabling direct causal analysis of how specific features influence the final output.

Transcoders apply the principles of dictionary learning to the function itself. The learned decoder weights form a dictionary of feature vectors, and the encoder produces a sparse set of coefficients. This contrasts with standard dictionary learning on static activations by targeting the dynamic, input-dependent output of a specific computational submodule, isolating the features that the component actively contributes to the residual stream.

Because a transcoder replaces a component, it enables causal experiments on discovered features. Researchers can ablate a specific feature by clamping its coefficient to zero or amplify it by increasing its weight, then observe the direct effect on the model's final logits or behavior. This moves beyond correlational feature visualization to establish a direct causal link between a feature and the model's output.

While both produce sparse decompositions, they target different objects:

• Sparse Autoencoder (SAE): Decomposes a layer's input activations into sparse features.
• Transcoder: Decomposes a layer's output function into sparse features. A transcoder answers 'what features does this component compute and add?' rather than 'what features are present in the residual stream at this point?' This makes transcoders more directly suited for circuit analysis of specific MLP and attention computations.

The transcoder training loss combines a reconstruction error term with a sparsity penalty (typically an L1 loss on the latent coefficients). Given an input to the target component, the transcoder encodes it, applies a non-linearity, and decodes it to reconstruct the component's output. The sparsity term ensures only a few features are active for any given input, promoting monosemanticity in the learned feature dictionary.

The broader mathematical framework that transcoders instantiate. Dictionary learning decomposes a model's activations into a sparse linear combination of learned basis vectors, each representing a distinct, interpretable feature. A transcoder learns an overcomplete dictionary of features that reconstructs the output of a specific model component, effectively creating a sparse, interpretable replacement for that component's dense computation.

The core problem that transcoders are designed to solve. Polysemanticity is the observed phenomenon where a single neuron or feature direction responds to multiple, seemingly unrelated input concepts. For example, a single neuron might activate for both 'academic citations' and 'Hebrew text'. Transcoders decompose these mixed signals into a sparse set of monosemantic features, each corresponding to exactly one human-interpretable concept.

The ultimate object of study that transcoders help reveal. A circuit is a sparse, interpretable subgraph of a neural network consisting of connected attention heads and MLP neurons that implement a specific, human-understandable algorithm. By replacing an MLP layer with a transcoder, researchers can identify which sparse features are activated and how they compose with other components, making it possible to map out the computational graph of a behavior.

A causal intervention technique used to validate the features learned by a transcoder. Activation patching replaces a model's internal activation at a specific layer and position with a value from a corrupted or alternative forward pass. This is used to test whether a specific transcoder feature is causally responsible for a behavior by patching it in or out and observing the change in output.

The hypothesized compression phenomenon that makes transcoders necessary. Superposition posits that a neural network represents more independent features than it has dimensions in a given layer, compressing sparse features into a lower-dimensional space. Transcoders are designed to decompress this superposition by expanding the representation into a higher-dimensional, sparse feature space where individual concepts can be isolated and interpreted.