Ways a dictionary misleads

Shrinkage was the first known failure of the method, and the architectures in piece 8 largely fixed it. The ones below are harder to fix, because they come from the sparsity objective itself rather than from a particular penalty.

Absorption

Take a latent that seems to mean “the token starts with S”. Chanin and colleagues found such latents in Gemma Scope that failed to fire on some S-words, such as short. A token-specific latent for short had absorbed the “starts with S” direction. Sparsity rewards this: if “short” always implies “starts with S”, one latent is cheaper than two. No Gemma Scope SAE matched a logistic-regression probe on the first-letter task, and the authors consider their absorption measure likely an underestimate.Chanin et al., A is for Absorption, 2024.

The consequence is that a latent’s label can be accurate wherever it fires and still miss part of the concept, because it fails to fire exactly where a more specific latent has taken over.

Hedging

When the dictionary has too few latents for a set of correlated features, the SAE mixes them, and each latent carries a little of its neighbours. The authors suspect this is one of the main reasons SAEs underperform supervised baselines.Chanin, Dulka and Garriga-Alonso, Feature Hedging, 2025. Hedging is a problem for narrow dictionaries and absorption for wide ones, so increasing width does not escape both.

Composition

A meta-SAE, trained on the decoder directions of another SAE, splits an “Einstein” latent into “scientist”, “Germany” and “famous person”.Leask et al., Sparse Autoencoders Do Not Find Canonical Units of Analysis, 2025. Latents are not atomic. What counts as one feature depends on the size of the dictionary, as feature splitting already suggested in 2023.

Dark matter

About half of the SAE’s error vector, and more than 90% of its norm, can be predicted linearly from the input activation. Larger SAEs fail on mostly the same contexts.Engels, Riggs and Tegmark, Decomposing the Dark Matter of Sparse Autoencoders, 2024. The error is not just noise that more latents will eventually absorb. It has structure that the current objective does not capture.

Seeds

Train two 131K-latent SAEs on the same Llama 3 8B MLP, with different random seeds: “only 30% of the features were shared”. TopK SAEs depend more on the seed than L1 SAEs.Paulo and Belrose, Sparse Autoencoders Trained on the Same Data Learn Different Features, 2025. Song et al. (2025) report a pairwise consistency of 0.80 for TopK under different conditions. If the features were the model’s own, two runs ought to find them both times.

Random networks

SAEs trained on randomly initialised Pythia models score about as well on standard interpretability metrics as SAEs on trained ones. The metrics “do not distinguish trained and random transformers”.Heap, Lawson, Farnik and Aitchison, 2025. Reported in the bitfilter review. Part of what an SAE finds is structure in the data and the architecture rather than in what training learned.A 2026 synthetic study (Korznikov et al., reported in the review) found SAEs recovered 9% of known features while explaining 71% of the variance. On real activations, frozen random baselines came close to trained SAEs: 0.87 against 0.90 on interpretability, 0.69 against 0.72 on probing, and 0.73 against 0.72 on causal editing.

Features that are not lines

Days of the week and months of the year lie on circles in GPT-2 and Mistral 7B. An SAE, which can only represent lines through the origin, splits each circle across several latents.Engels et al., Not All Language Model Features Are One-Dimensionally Linear, 2024. More generally, any feature that lives on a curved manifold wastes latents and appears as several unrelated pieces.

What this adds up to

None of these results says the features are fake. Each says something narrower: the dictionary we get depends on its width, its seed and its objective, and nothing in the method picks out one dictionary as the model’s own. In the words of one of the papers above, SAE features are “a pragmatically useful decomposition.”