- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Instead of traditionally labelling SAE concepts after training, authors assign labels to the decoder directions before training. This allows to learn specific pre-defined concepts directions while also leaving space for unsupervised concepts learning like in traditional SAEs. They evaluate their method on ImageNet with CLIP and SigLIP models, comparing against unconstrained, pretrained and SpLiCE SAEs.
Critical strengths
I appreciate the honesty of authors about for example the 15 class latents with zero precision and about the sibling classes confusion.
Critical weaknesses
- The whole evaluation is centered on ImageNet and a better coverage on ImageNet of the author’s method could be due to this. Authors do not test distribution shifts or transfer to other datasets. Authors should show results on different datasets from ImageNet and also OOD concepts performance.
- Authors target only post-projection embedding while normally SAEs are trained also on intermediate layers. So if their goal is also to understand internal computations of models they should also target intermediate layers in their analysis.
- Concept-SAE (Ding et al., 2026) and CB-SAE (Kulkarni et al., 2026) are the closest prior work. Can you run and compare them to your method? Right now it's hard to tell how much of the gain over plain SAEs is unique to anchoring.
- R squared seems to suffer a lot as compared to unconstrained SAEs (Table 1). As the primary purpose of SAEs is to faithfully reconstruct the activations, this drop is not a good sign. It would be interesting to see if this drop comes from named or unnamed concepts? This could potentially be problematic for generalisation too.
Questions for authors
See Weaknesses
7 comments
Response to Reviewer 3 Part 2
W3: Comparison with Concept-SAE and CB-SAE. We agree that citing these methods without an empirical comparison leaves the distinction unclear. The central difference is when and how a latent acquires its semantic identity.
| NAMED-SAE | Concept-SAE | CB-SAE | |
|---|---|---|---|
| Identity assigned | before training | during training | after training |
| Dictionary | anchored, adaptive | supervised | pretrained, pruned |
| Supervision | text anchors | image scores and masks | concept set |
| Per-image concept targets | no | yes | no |
NAMED-SAE fixes identities from text before dictionary learning. Concept-SAE supervises designated concepts during training using per-image concept scores and localization masks. CB-SAE modifies and names a pretrained SAE after its dictionary has formed.
CB-SAE. CB-SAE has no released implementation, so we cannot reproduce the complete method. We instead evaluate its central post-hoc mechanism using the unconstrained-SAE control in Table 1. The backbone, training data, vocabulary, and alignment are the same as for NAMED-SAE; the principal difference is whether identities are assigned after dictionary learning or constrained during it.
Post-hoc assignment reaches 0.559 top-1, 0.461 precision@16, and 0.36 causal removal, compared with 0.713, 0.853, and 0.89 for NAMED-SAE, despite reconstructing better ( versus 0.710). This control does not reproduce CB-SAE's pruning stage, but directly tests the part relevant to our claim: applying the same semantic alignment after training does not reproduce the precision or causal control obtained by anchoring identities during training.
Concept-SAE. Concept-SAE cannot be directly ported because it operates on spatial token maps using per-image concept scores and CLIPSeg masks, whereas our representation is a pooled embedding. We therefore construct a pooled Concept-SAE-style control that isolates the source of the named directions.
For each of the 1,000 ImageNet classes, the Concept-SAE-style control estimates a supervised class-presence direction and freezes it as a named decoder atom. The matched text-anchor control instead freezes each atom to the aligned text embedding of its class name. Both use 1,000 named atoms, 500 free atoms, and the same encoder, sparsity, training data, reconstruction objective, and freezing. Only the source of the named directions differs.
| Named directions | Free | Class hit | Top-1 | Named share | |
|---|---|---|---|---|---|
| Supervised class presence | 500 | 0.594 | 0.944 | 0.757 | 0.274 |
| Aligned text anchors | 500 | 0.638 | 0.960 | 0.732 | 0.578 |
Named share is the fraction of the combined reconstruction magnitude carried by named atoms.
Text anchors improve by 0.044, class hit by 0.016, and named share from 0.274 to 0.578. Supervised presence directions improve top-1 by 0.025, as expected because they are estimated specifically for class discrimination.
For this controlled comparison, both arms receive the full ImageNet labels, but use them differently. The supervised control estimates a separate direction for every class, whereas the text-anchor control uses the labels only to estimate one shared alignment. Supervised directions also require labelled examples for every named concept, limiting this control to the 1,000 classes. Text anchors require only a concept name and extend to all 9,866 vocabulary concepts, including 8,866 without ImageNet labels. The standard NAMED-SAE configuration estimates its shared alignment from only ten images per class.
Thus, the post-hoc control provides the direct evidence that the gain over plain SAEs comes from anchoring identities during dictionary learning: applying the same alignment afterward does not recover NAMED-SAE's precision or causal control. The Concept-SAE-style control addresses a complementary question, showing that text anchors preserve more reconstruction and named-bank utilization than class-specific supervised directions, while the latter provide moderately higher class discrimination.
Response to Reviewer 3 Part 3
W4: Whether the reconstruction cost falls on named or unnamed concepts. We decompose reconstruction by decoding with only the named bank or only the free bank. Following Appendix C, variant B fixes each named decoder row to its text anchor, whereas variant C adds a bounded residual that allows controlled adaptation. Our reported configuration is variant C with .
Every model has the same 9,866 named slots and permits 16 active named atoms per image. We vary the named residual, free-bank width, and free-bank sparsity budget. The first two rows isolate the residual at fixed free capacity; the remaining rows show how additional free capacity changes the allocation of reconstruction. All models are evaluated on the full ImageNet validation set.
| Named bank | Free | Sparsity | Total | Named | Free | Named share |
|---|---|---|---|---|---|---|
| No residual (B) | 500 | 16+4 | 0.647 | 0.357 | 0.228 | 0.554 |
| Bounded residual (C) | 500 | 16+4 | 0.714 | 0.478 | 0.177 | 0.611 |
| Bounded residual (C) | 2,000 | 16+8 | 0.762 | 0.362 | 0.394 | 0.494 |
| No residual (B) | 4,000 | 16+16 | 0.796 | 0.229 | 0.573 | 0.386 |
Sparsity gives the named and free per-image budgets. Named share is the fraction of combined reconstruction magnitude carried by named atoms. Per-bank is measured by decoding with that bank alone and is not additive.
The exact-width control in our response to Reviewer 1, Q3 establishes that the reconstruction gap is the genuine cost of anchoring. At the same width (10,366 atoms) and sparsity (16+4), the unconstrained SAE reaches , compared with 0.7135 for NAMED-SAE. Anchoring therefore costs 0.0917 in at fixed capacity, close to the 0.097 gap in Table 1.
The bank decomposition clarifies where reconstruction is carried. At fixed free capacity and sparsity, adding the bounded residual raises named-bank from 0.357 to 0.478, named share from 0.554 to 0.611, and total from 0.647 to 0.714. The residual therefore recovers reconstruction within the named bank by adapting its concept directions to the activation distribution.
Additional free capacity recovers further reconstruction by moving information into unnamed features. Increasing the bounded model's free bank from 500 to 2,000 atoms raises total to 0.762, while named-bank falls to 0.362 and free-bank rises to 0.394. Similarly, the no-residual model reaches with 4,000 free atoms, but its named share falls to 0.386. This larger model is not capacity-matched to the baseline; it shows that additional width and support can buy back reconstruction by delegating more information to the free bank. Thus, the free bank does not cause the reconstruction deficit: it compensates for the constraint imposed on the named bank.
Generalization. The transfer results in W1 show that reduced reconstruction does not necessarily destroy semantic identity. Across ten ImageNet-derived shifts, falls to 0.507–0.674, while precision@16 remains 0.670–0.979 and NAMED-SAE continues to outperform post-hoc naming. However, transfer to a different label space is limited: on CIFAR-100, class-latent accuracy falls to 0.485, although precision@16 remains 0.942. We therefore conclude that anchored identities remain stable under the shifts tested, not that reconstruction loss is irrelevant to every form of generalization.
Thus, the reconstruction deficit is the genuine cost of constraining the named bank at fixed capacity. Bounded residuals recover part of that cost within the named bank, while additional free capacity recovers more reconstruction by moving information into unnamed features. The identities remain stable across the evaluated shifts, although transfer to substantially different label spaces remains limited.
Response to Reviewer 3 Part 1
W1: Distribution shift, cross-dataset transfer, and OOD concepts. We thank the reviewer for raising this concern. The submission evaluates all three settings, but compresses the discussion into lines 449–459 and places most evidence in Appendix E.
Distribution shift. We evaluate ten ImageNet-derived shifts without retraining or renaming: ImageNet-V2, ImageNet-Sketch, ImageNet-R, ImageNet-A, and three ImageNet-C corruptions at severities 2 and 4. Table 11 reports the complete results.
NAMED-SAE outperforms post-hoc naming on all ten shifts. Precision@16 remains 0.670–0.979 even as R2 falls to 0.507–0.674, indicating that reconstruction degrades before latent identity. This ordering is stable across three seeds, with top-1 varying by at most 0.007 and R2 by at most 0.008, and is reproduced with SigLIP.
The code also identifies the shift itself. The latents illustration and monochrome fire on 79% and 42% of ImageNet-Sketch, while fog and smoke fire on 89% and 55% under fog corruption. All four activate on less than 0.1% of clean ImageNet. In contrast, on four of six examined shifts, the most frequent post-hoc latent receives a semantically unsupported name.
Cross-dataset transfer. Transfer to a different label space is more difficult. On CIFAR-100, NAMED-SAE reaches class-latent accuracy of 0.485, compared with 0.403 for post-hoc naming, while precision@16 remains 0.942. A 32×32 ImageNet control shows that the decline is not explained by resolution alone. Instead, the chosen vocabulary and 20-sparse code do not capture every distinction required by the new label space.
OOD concept evaluation. We evaluate 43 vocabulary concepts on iNaturalist taxa. Specific order- and family-level concepts achieve mean AUROC 0.833, including cactus (0.994), moss (0.996), and owl (0.950). Broad kingdom- and class-level concepts average 0.514.
This failure appears to arise from sparse competition with more specific atoms. Aggregating nearby anchors at readout raises AUROC from 0.51 to 0.98 for bird, 0.50 to 0.98 for insect, and 0.55 to 0.99 for fish. Because this readout applies to any named representation, we present it as evidence that the hypernym information survives transfer, not as an advantage over post-hoc naming. It also motivates a hierarchical NAMED-SAE in which activation from specific concepts propagates to their hypernyms.
Thus, the learned identities remain stable across substantial distribution shifts and transfer to specific concepts on a different dataset. Transfer to a new label space and broad hypernyms remains limited. We will make both the positive results and these limitations more prominent in the main text.
W2: Intermediate-layer activations. Applying NAMED-SAE inside the network requires aligning the 768-dimensional text space with an activation space of different dimension. We therefore trained it on the 1,024-dimensional layer-22 post-MLP residual stream of CLIP ViT-L/14-336 and evaluated on the full ImageNet validation set, with the text anchors entering through a learned semi-orthogonal embedding. Causal removal is unavailable because our saved layer-22 representation holds only the CLS token and cannot be propagated through the remaining vision layers.
†Second-seed R2 is from the training held-out split. The layer-22 model uses 2,000 free atoms and r=0.5, against 500 and r=0.3 at the output, so R2 is not comparable across layers.
At layer 22, precision@16 remains high at 0.820 against 0.863 at the output, while class top-1 falls from 0.717 to 0.626. Of the 1,000 class slots, 989 remain active against 994, though the dead fraction over the full vocabulary rises from 15% to 35%. Moving inward therefore affects vocabulary utilization and classification more than the semantic precision of active class latents. The second seed agrees within 0.005 precision@16 and 0.004 top-1.
Thus NAMED-SAE extends to an intermediate representation: nearly all class slots stay active and their semantic precision stays high. This establishes feasibility inside the network, though not a complete layerwise or spatial-token analysis. The full layer-22 ladder and alignment ablations will be added to Appendix B.