Sparse Autoencoders Often Encode Diffuse Features
Sparse autoencoders frequently fail to recover distinct features, converging instead to diffuse representations that misalign with the underlying data structure.
Sparse autoencoders are widely used to identify interpretable features within large language models, but they often struggle to recover ground-truth features in controlled benchmarks. Instead of isolating single concepts, these models frequently collapse into a diffuse phase where learned dictionary atoms are indistinct and essentially represent smeared combinations of multiple true features. These findings challenge the reliance on standard dictionary learning for mapping neural activations to modular concepts.
The Geometry of Diffusion
In an ideal sparse recovery, each dictionary atom should map one-to-one to a ground-truth feature. The autoencoder objective consists of a reconstruction term—usually L2 error—and a sparsity penalty. During backpropagation, the model discovers that it can minimize L2 reconstruction error by combining several true features into a single dictionary atom. While a sparse, individual representation is the intended target, the reconstruction loss treats a linear combination of features as effectively equivalent to the sum of individual activations. Because the autoencoder finds that 'smearing' these components reduces the total reconstruction error more reliably than finding the exact sparse basis, it optimizes toward this dense, diffuse state. It effectively treats the sparsity penalty as a weak constraint that can be bypassed by distributing information across more atoms than strictly necessary.
Researchers tested this phenomenon using the MAIS-O43 protocol, performing 3,300 fits across a 165-cell parameter grid that varied the nesting fraction and sparsity penalty. Across every configuration, the models consistently avoided the global optimum of the sparse-coding objective. Instead of reaching the sparse recovery or the known merging behavior for nested features, the training process forced the dictionary atoms into a persistent, suboptimal diffuse phase.
| Metric | Global Optimum | Trained SAE Outcome |
|---|---|---|
| Feature Recovery | High (Matches ground truth) | Low (Median cosine 0.5-0.7) |
| Learned Code Sparsity | Matches ground truth | ~10x denser than target |
| Feature Merging | Occurs in nested cases | Absent (Diffuse phase prevails) |
Limitations of the Optimization Path
When interpretability tools assume a one-to-one mapping between atoms and semantic concepts, they are likely interpreting a blurred approximation of the truth. The failure of these models to reach the global optimum points to a fundamental divergence between the objective function and the training dynamics. In particular, adaptive optimizers like Adam, which adjust learning rates for each parameter, may exacerbate this issue. If the model enters a region of the loss landscape where the sparsity penalty is overshadowed by the reconstruction gradient, Adam’s per-parameter scaling might suppress the corrective updates needed to push weights back toward sparsity. The optimizer essentially maintains the weights in a non-sparse, locally stable basin because the gradient signal for individual features is weaker than the combined reconstruction signal.