Unknown-Unknown SAE Feature Discovery via Context-Differential Activation with Probe-Alignment Filtering
We describe the unknown-unknown SAE feature discovery workflow in MIVP. The workflow identifies SAE features that fire differentially between two contexts, then filters out features that are already explained by a registered probe via decoder-row cosine alignment. Surviving features are candidate hidden-behaviour features for human review.
1. Introduction
A registered probe is an instrumented concept. If a model is doing something we have not instrumented, the probes do not see it. Sparse autoencoder features are candidates for such concepts. A feature that fires differentially between contexts and is not explained by any registered probe is a candidate hidden-behaviour feature.
2. Background
Sparse autoencoders (Bricken et al., 2023; Cunningham et al., 2023; Templeton et al., 2024) are the de facto dictionary. Polysemanticity and superposition are studied by Elhage et al. (2022). Hidden-behaviour discovery is discussed by Hubinger et al. (2024).
3. Method
3.1 The Workflow
UnknownUnknownDiscovery takes an SAE, two context groups of activations, and the list of known probes. For each feature, the effect size is (mean_a − mean_b) / pooled_std. Features with effect below min_effect (default 1.0) are skipped.
3.2 The Decoder-Row Alignment Test
For each candidate feature, compute the cosine similarity between the SAE's decoder row and each known probe direction. Features with max_align > 0.7 are filtered out (already explained).
3.3 Structural Non-Gating Posture
As with the other T3 audits, the finding is experimental and E0.
4. Results
Evaluated on a synthetic SAE with a known concept injected as a sparse direction.
| Setting | True unknowns detected | False positives |
|---|---|---|
| 1 known, 1 unknown | 1 / 1 | 0 |
| 5 known, 5 unknowns (low signal) | 3 / 5 | 0 |
| 5 known, 5 unknowns (high signal) | 5 / 5 | 0 |
5. Discussion
The max_align filter is the structural reason the workflow is useful and not just noisy. The workflow inherits all of the SAE's limits.
6. Conclusion
An unknown-unknown SAE feature discovery workflow that surfaces context-differential features not explained by registered probes.
References
Bricken et al. (2023). Anthropic. / Cunningham et al. (2023). ICLR. / Elhage et al. (2022). Anthropic. / Hubinger et al. (2024). arXiv. / Templeton et al. (2024). Anthropic.