Determine whether selection or controller design causes observed performance differences

Determine whether differences between sparse-autoencoder probing and control results are caused by feature-selection procedures or by static versus gated controllers, while controlling the SAE dictionaries, supervision, data splits, tuning budgets, and comparator methods.

Background

The survey identifies disagreement between probing/classification outcomes and advances involving conditional control. It proposes comparing SAE dictionaries under shared supervised or unsupervised selection rules and under static or gated controllers, while including dense probes and direct activation-space directions as matched comparators.

The causal attribution remains unresolved because selection and gating have not yet been varied independently under controlled conditions. The proposed target–retention frontier is intended to distinguish performance gains due to the selected SAE features from gains due to the controller or intervention policy.

References

The probing/classification disagreement and conditional-control advances motivate this design; attributing their differences to selection or gating remains a hypothesis until those factors are controlled.

— Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey of Recoverability, Interpretability, and Utility  (2610.20139 - Lamsiyah et al., 14 Oct 2026) in Section 6.1, “Three testable comparisons”