Source
Kantamneni et al. (2025), LIT-568. Read as NOTE-309.
The practice
- Default to logistic regression on the residual stream, at the layer where it probes best (layer 20 of Gemma-2-9B’s 42, in the paper). It is the paper’s reference baseline, and SAE latents added nothing to a toolkit built around it
- Do not switch to SAE probes for the hard cases. Few examples, imbalanced classes and noisy labels were where the interpretable basis was expected to help, and it did not
- When claiming an SAE-based result, give the baseline the same affordances. Pool raw activations across tokens (attention pooling) if the SAE probe pools, and choose among methods on validation data, never on test. Both of this paper’s apparent SAE wins disappeared under those rules
The recommendation comes from Kantamneni et al., LIT-568, who added SAE probes to a toolkit of raw-activation probes and let validation AUC choose per task. On Gemma-2-9B layer 20 across 113 datasets, SAE probes were chosen for 14 tasks and changed mean test AUC by −0.003 ± 0.002 (Figure 4); under data scarcity, class imbalance and label noise there was no average gain at any setting, and under covariate shift SAE probes generalized worse. Their multi-token SAE win rate fell from 19.6% against a last-token baseline to 8.7% against an attention-pooled one, and the result held on Llama-3.1-8B with Llama Scope SAEs.
Conditions
- Binary probing with linear heads, on Gemma-2 and Llama-3.1. It does not rule out SAEs for circuit analysis, steering or unlearning
- Public SAEs. A better SAE could change the result, though eight architectures over two years moved it less than the spread