Transcoders are auxiliary sparse models designed to explain a neural transformation rather than merely reconstruct one activation state. A transcoder takes the input to a component such as an MLP, produces sparse latent features, and reconstructs that component’s output, creating a feature vocabulary tied directly to computation.
In mechanistic interpretability, transcoders address a limitation of ordinary sparse autoencoders. An SAE can reveal sparse factors that reconstruct a state; a transcoder asks which sparse factors explain how one state becomes another across a nonlinear module.
This Cognitive Art owner explains transcoder architecture, sparse latent inference, MLP replacement, feature dictionaries, reconstruction fidelity, sparsity, feature splitting, attribution edges, residual error, circuit analysis and causal validation.
1. From sparse states to sparse transformations
The Sparse Autoencoders owner reconstructs x from sparse z. A transcoder receives x and predicts f(x), where f is the component being interpreted.
2. Why MLPs are natural targets
Transformer MLPs combine residual features through a nonlinear transformation and write new directions. A sparse surrogate can expose a smaller set of active transformations.
3. A transcoder is a surrogate, not ground truth
It approximates the original component under a sparsity constraint. Scientific claims must track approximation error because a sparse circuit can explain the surrogate better than the base model.
4. Fidelity is behavioural as well as geometric
Measure vector reconstruction and replace the original component with the transcoder inside the model. Task loss and behavioural preservation reveal whether apparently small reconstruction errors matter downstream.
5. Sparsity creates a manageable active vocabulary
Only a small subset of latents should fire for one input. Report reconstruction-versus-sparsity trade-offs rather than celebrating one operating point.
6. Encoder and decoder have different meanings
The encoder describes evidence that makes a latent fire; the decoder describes what contribution that latent writes into the target output space.
7. Sparse transformations support attribution graphs
A small active latent set can connect upstream conditions to downstream features, which is why transcoders are useful for Attribution Graphs.
8. Dead, dense and split features remain diagnostic
Track latent density, dead capacity and feature splitting across dictionary widths. Transcoder interpretability inherits many dictionary-learning failure modes.
9. Residual error must be audited
Probe and intervene on f(x)−f̂(x) for variables central to the claimed circuit. Important computation left in the residual limits completeness.
10. Causal validation returns to the original model
Use Causal Tracing and activation interventions to test whether predicted sparse routes actually control the base model.
11. Replacing many components tests cumulative fidelity
A single transcoder can be accurate while a stack of replacements accumulates error. Measure behavioural degradation as more original modules are substituted.
12. Failure signatures
Surrogate worship, reconstruction-only evaluation, showcase features, residual blindness, feature-label leaps and causal claims without base-model interventions are the main warning signs.
13. Evaluation blueprint
Measure held-out reconstruction, sparsity, density distributions, replacement loss, semantic coherence, residual information, cross-run stability, graph faithfulness and causal intervention selectivity.
14. The proof standard
The strongest transcoder is not the sparsest diagram. It is the sparse surrogate that preserves the computation closely enough that its features and edges keep predicting what the original model will do when intervened upon.
World Return
A transcoder becomes scientifically valuable when sparse explanatory structure survives the return trip to the original neural computation.