Sparse autoencoder probes do not consistently outperform baseline probes on language-model activation classification tasks.
Reassessed
Read the abstracts of 2505.20063, 2605.31183 and 2602.14111. The first two are steering studies, not probing studies, and do not bear directly on this claim. The third gives indirect support. Refined the assessment, recorded 2602.14111 as a mention, and lowered marginal_yield to 0.35.
Reassessed: still Supported
verdict confidence 0.75 → 0.78 · credence 0.85
Assessed
First pass under a tight budget, with $0.49 already consumed at the start. Assessed as supported (credence 0.85), relying mainly on 2502.16681 and the DeepMind negative-results post. Left the claim atomic. The study-by-study comparison of 2505.20063, 2605.31183 and 2602.14111 was not done, and marginal_yield is set to 0.45 so a later pass can do it. The seven unplaced argument passages were left for a later pass.
Assessed Supported
verdict confidence 0.75 · credence 0.85
The claim is that probes trained on sparse autoencoder (SAE) latents do not reliably beat ordinary probes trained on raw language-model activations when classifying activations. The evidence supports it. The largest head-to-head study, "Are Sparse Autoencoders Useful? A Case Study in Sparse Probing" (Kantamneni, Engels, Rajamanoharan, Tegmark, Nanda, 2025), compared SAE probes against tuned baselines (logistic regression, PCA, k-nearest neighbours, XGBoost, MLP) on about 113 binary classification datasets using Gemma 2 SAEs. It tested the regimes where SAEs were expected to help: scarce training data, class imbalance, label noise and covariate shift. It found no meaningful improvement in any regime. When SAE probes were added to the candidate set and the best probe was chosen on validation data, test performance did not improve. An apparent multi-token advantage for max-pooled SAE probes (a 19.6% win rate) fell to 8.7% once the baselines were also allowed attention pooling. In a separate set of negative results, Google DeepMind's interpretability team reported that SAE-based probes for detecting harmful user intent generalised out of distribution worse than linear probes on the residual stream. The claim says "do not consistently outperform", not "never outperform". It is compatible with isolated wins, and the studies do report some: individual datasets where SAE probes come out ahead, some interpretability-driven debugging of spurious features, and dataset-quality findings. Some position papers argue that this pattern is expected, because a reconstruction holds strictly less information than the activation it reconstructs. The live disagreement is not over whether SAE probes win on average, which they do not in these setups. It is over whether newer SAE architectures, wider dictionaries, larger models or particular task types could reverse the result. Evidence for such a reversal is thin. The main caveat is that most of the controlled evidence comes from one research group and one model family (Gemma 2 with Gemma Scope SAEs), so the evidence is less independent than the number of experiments suggests.
Claim entered the graph