- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work is about that causal-localization scores from interchange-intervention benchmarks (RAVEL, MIB, SAEBench) measure a representation–intervention pair, and separates the nominal intervention (swap features) from the realized edit in residual space. Instead of a single score, it proposes reporting an intervention-response profile: held-out cause measured across a sweep of effective residual rank , with the selector, write-back operator, and coverage condition stated explicitly. For dictionaries, it adds a full-swap recovery diagnostic, , the fraction of the full-state patch's causal effect that survives swapping every dictionary coordinate. On RAVEL/gemma-2-2b, IOI/GPT-2-small, and a synthetic transformer with a known causal site, it shows that (i) the leading family flips with (DAS at low rank, SAE at high rank); (ii) rank-matched random subspaces have a causal floor that depends on and answer cardinality; (iii) changing only the selector, or only the write-back operator, moves SAE cause substantially while orthonormal bases are invariant; and (iv) a replay of MIB's learned masks yields SAE and neuron supports differing in residual rank by up to .
Critical strengths
-
Cross-family comparisons in RAVEL/MIB/SAEBench do equate a -dimensional orthonormal subspace with overcomplete dictionary latents, and no prior benchmark pairs a method's score with what a rank-matched random subspace achieves. The nominal/realized distinction, the rank-matched random control (an explicit analogue of probing control tasks), and the full-swap diagnostic are each cheap, well-defined, and immediately adoptable.
-
The paper pre-registered effect-size criteria, reports which predictions failed, discards two invalid rank-matching procedures before adopting a third, and explicitly withdraws its own initial hypothesis that write-back alone explains MIB's SAE-vs-neuron ordering.
Critical weaknesses
-
The projected donor bypasses the encoder: it orthonormalises selected decoder columns and replaces the entire component of in that span, so its realized rank is always . The featurizer swap moves by , whose realized per-example rank is . For sparse codes far below on most graded pairs. The paper notes this but still presents 4.2 as a separate write-back semantics finding. By its own principle, the median operator gap (Table 7) and the IOI result compare interventions of very different realized rank. Without that number, write-back semantics matter cannot be separated from the featurizer swap barely edits anything at this . The same operator choice means Table 2's headline (an SAE dictionary leads at rank 128) is a result about a label-selected 128-dimensional subspace of versus a 128-dimensional fitted rotation.
-
The two results behind the cross-family claim (the rank-dependent ordering (Table 2, Fig. 2) and the common-rank reversal (Table 4)) are each explained as well by estimator variance or optimization failure. DAS at rank fits continuous parameters (≈295k at ) from 1024 entities, whereas the
saeselector picks 128 fixed columns. The paper's own numbers show DAS is heavily overfit already at rank 8 and that its standing depends on . A high-variance fitted subspace losing to a low-variance selected one as grows is the bias–variance prediction; reporting the fit-vs-held-out gap along the rank sweep, or scaling with , would separate the two. The paper already has cheap lower bounds on what an axis-aligned subspace of that size can achieve but does not report them. The abstract's "changes the ordering in both successfully optimised large-mismatch tests" presents , after a post-hoc exclusion, as a robust finding.
Questions for authors
-
For every featurizer-swap cell in Tables 3 and 7, please report the distribution over graded pairs of the realized per-example rank and the edit norm , alongside the same quantities for the projected donor. Does the operator gap survive when cause is plotted against realized per-example rank?
-
Table 2 reports cause only. Please add isolation for each leader/runner-up pair. At rank 128 the SAE's isolation under the projected donor is 0.410 (§4.2). Does the rank-128 SAE lead survive under any isolation-constrained comparison, or does DAS lead at every rank once isolation is held fixed? (You criticise RAVEL's mean aggregate for rewarding non-intervention.)
-
Along the rank sweep in Fig. 2 / Table 2, please report DAS's fitting-entity cause beside held-out cause, and/or re-run DAS with scaled with (or with a regularized objective). Does the DAS$\to$SAE crossover move with ?
-
For the two reversing cells in Table 4, what cause does the labelled
neuronselector achieve at (445 and 402), and what doessupervisedachieve at its effective rank 127? -
Why was rank-matching done only downward (neuron SAE's )?