acceptodds
Under review as a conference paper at ICLR 2027

The Representation-Continuation Gap: Low-Rank States vs. Low-Rank Dynamics in Vision Transformers

Abstract

Low-rank activations make a Vision Transformer look compressible, but compression has to survive the remaining residual computation: after a state has been represented in a small subspace, the later blocks must still compute there. We show that these two requirements can diverge. A subspace can reconstruct the current state, or entry, nearly exactly while failing to support the suffix computation, a failure we call the representation-continuation gap. In eight frozen ViT settings spanning supervised, self-supervised, and CLIP pretraining, confining only the last two blocks to an almost lossless entry space reduces accuracy by 1 to 67 points (median 12). The gap persists under distribution shift and at mid depth, and it appears even though late-layer subspaces are image-specific and stable across views. The obstruction is directional rather than merely low-rank: at a 64-direction budget, replacing eight entry directions with directions read from the input's natural future raises CLIP-FT ViT-B from 3.9% to 82.5%, while random directions reach 4.9%. We explain this with two facts: a fixed continuation space must contain directions created after the entry, and whether an omitted direction matters depends on how strongly the output responds to it. We then introduce functional subspace completion (FSC), a small self-distillation of subspace directions. FSC keeps the dominant entry directions, learns a shared source of complementary directions through the frozen suffix by matching the original model's predictions, and uses no future states at test time. With only 6,144 learned parameters per ViT-B classifier (0.007% of the frozen model), FSC raises accuracy on 5,000 held-out images from 50.6% to 70.5% on AugReg ViT-B and from 4.0% to 77.6% on CLIP-FT ViT-B relative to the entry basis, against 83.0% and 84.5% for unconstrained continuation. FSC improves on the entry basis in all eight settings and transfers to shifted test sets without retraining. Low-rank spaces should therefore be evaluated not only by the states they reconstruct, but by the computation they allow to continue. Code is available in the supplementary material.

Primary area:
interpretability and explainable AI
Keywords:
Vision Transformers, Low-Rank Representations, Activation Subspaces, Representation Geometry, Subspace Interventions, Mechanistic Interpretability, Residual Stream

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_E2n9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper argues whether nearly lossless representation of the current ViT state in a subspace guarantees that the remaining layers can continue computing within that same subspace. Further, the authors introduce the representation–continuation gap. Even when the entry state is preserved almost exactly, restricting subsequent residual updates to the same subspace can degrade performance. Further, to reduce the gap, authors propose FSC, which learns a few complementary directions while keeping the original model frozen. They demonstrate that FSC improves performance across 8 .

Critical strengths
  1. Well-controlled experiments. The paper isolate entry reconstruction errors and providing convincing evidence for the representation–continuation gap.

  2. Thorough evaluation. Across 8 ViT settings, FSC keeps the dominant entry directions and learns the shared source of complementary directions while keeps parameter efficient. Results on extensive experiments demonstrate the effectiveness of their models.

Critical weaknesses
  1. Poor writing and presentation. The overall writing quality is poor, making the paper’s main ideas unnecessarily difficult to follow. Several terms and expressions in the introduction are insufficiently explained. Here are some examples: “representation sufficiency” and “continuation sufficiency” (lines 53 and 74–76), “current entry support” (line 99), and r−kr-k (line 97) and " define and test continuation sufficiency"(line 111). The paper needs substantial revision to improve its clarity, organization, and explanation of key concepts before acceptance.
Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer E2n9

Thank you for the review, and for finding the experiments well controlled and the evaluation thorough. We are sorry that the submitted text was hard to follow. Your list of the exact lines where it failed guided our rewriting of the introduction.

In short. Every term you list is now defined in the introduction, on one running example. The abstract, the introduction and the conclusion are rewritten, the main text names four comparison arms, and the changed text is in blue in the revised PDF.

Weakness. Terms in the introduction were not explained (lines 53, 74–76, 97, 99 and 111 of the submission). The running example is an image entering the last two blocks of a ViT-B, where the state is a matrix of 197 tokens by 768 channels. The revised introduction defines the terms on this example (p. 1, lines 48–53, and pp. 2–3, lines 104–108).

  • Entry. The state at which the restriction begins. In the example it is the input to the last two blocks.
  • Entry subspace. The span of the r leading right singular directions of the entry. These are r channel directions out of 768.
  • Representation sufficiency (line 53). Projecting the entry onto the subspace leaves it nearly unchanged. The entry has 197 tokens, so 197 directions span all of it, and the relative error is below 10⁻⁶.
  • Continuation sufficiency (lines 74 to 76). The output of the model is preserved when every later attention and MLP update is also projected onto the same subspace. With the same 197 directions, AugReg-B loses 15.7 points of accuracy.
  • Entry support (line 99). The entry subspace together with the all-ones channel direction, which carries the mean of each token over its channels and which our projection rule restores separately.
  • r − k (line 97). FSC has a budget of r directions. It keeps the r − k leading entry directions and fills the remaining k with learned directions. The main experiments use r = 64 and k = 8.
  • "Define and test continuation sufficiency" (line 111) now reads: we test whether the last blocks still produce the same output when their updates are confined to a subspace that represents their input (p. 3, lines 119–125).

Comparison arms. The main text compares four arms (p. 2, lines 73–78). Each starts from the same projected entry.

  • Free continuation. The later blocks then run without restriction.
  • Entry basis. Every later attention and MLP update is also projected onto the entry subspace.
  • Future basis. Some entry directions are replaced by directions taken from the image's own later states. It sees the future and serves as a reference.
  • FSC. The same entry directions are replaced by learned directions, which need no later states.

Organization. The main text now follows one line of argument. The introduction states the question on the running example. Section 5.1 (p. 6) shows that a subspace containing the whole entry still fails the last blocks. Section 5.2 (p. 7) shows that eight directions decide most of the outcome. The new Section 5.3 (pp. 7–8) shows that the loss comes from holding the subspace fixed across depth, and Section 6 (pp. 8–10) presents FSC.

Soundness. The revision adds four controls that test the result further. Refitting the subspace after every update brings accuracy within 0.6 points of free continuation in seven of eight settings, so the gap comes from holding the entry subspace fixed. In a ViT sliced by SliceGPT at 25%, the gap of the complete entry basis grows from 16.25 to 21.75 points on AugReg-B and from 33.25 to 64.40 on CLIP-FT. With more tokens than channels, the comparisons at 64 directions keep their sign. On GPT-2 and GPT-2 medium with 256 tokens, the complete entry basis moves the output by 1.06 and 1.80 nats of KL.

We hope the revised text is now easy to follow. If any passage is still unclear, please tell us where and we will fix it. If the revision resolves your concern about the presentation, we would be grateful if you would reconsider your rating.

Reviewer_JmmCOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper targets the Representation-Continuation gap in Vision Transformers. This gap is between the subspace of the output of a transformer block which is denoted as the entry\textit{entry}, in the paper often used as the output of the 10th block of a 12 block vision transformer, and the subspace of the outputs of subsequent blocks denoted as the suffix\textit{suffix}. The paper studies the dynamics of constraining the outputs of the suffix blocks by projecting them into the entry's subspace to close this Representation-Continuation gap. The baseline presented is projecting the entry on it's top rr right singular vectors. Where the almost lossless projection uses r=dr=d where dd is the original number of dimensions of the entry d=197d = 197. This highlights the loss in performance even if an almost exact projection of the entry is used as a constrain on the suffix. Another baseline is the entry basis where r=64r = 64 and this shows a significant drop in performance up to 66.666.6 percentage points of accuracy in MAE-B/linear. The paper then introduces Functional Subspace Completion, an approach where only r−kr-k directions of the entry are retained and kk directions are learned through knowledge distillation. The teacher model is the original unconstrained model while the student is the constrained model, and the KL divergence loss is calculated only on the final output and not on intermediate layers. The usage of these kk learned directions narrows the Representation-Continuation gap with the matrix AA containing the kk directions being shared among all inputs while the rest of the entry's r−kr-k directions are calculated per input.

Critical strengths
  • The paper presents a strong view of the Representation-Continuation gap problem, with different baselines such as entry-basis, free continuation, and random directions. The paper also provides a fairly simple theoretical justification. This problem is of significance in the fields of model compression and interpretability.

  • There is a wide breadth of experiments implemented in the paper and the appendices which adds to the reliability of the approach and the soundness of the motivation and hypothesis. This includes 88 models, and longer suffixes.

Critical weaknesses
  • The Representation-Continuation gap is a problem that arises when large models need to be computed more efficiently, for example when r<<dr << d however, in all experiments the full width of the transformer blocks was computed and the FSC projection computed on top, so it is unclear what is the actual gained efficiency in this case. Especially when the suffix only consists of 2 out of the 12 transformer blocks.
  • The paper did not provide real world application to the narrowed Representation-Continuation gap, given the loss in accuracy against the gained efficiency, why would someone use FSC over the original model? The paper suggests interpretability potential of FSC and demonstrating it would strengthen the paper.
Questions for authors
  • What are the differences and relationships between the retained directions r−kr-k, the dropped kk directions, the learned kk directions, and the original directions of the entry subspace? it is clear that the overlap between the dropped kk and the learned kk should be zero since all directions of rr are removed from AA. And in Appendix E.6 you experiment with allowing the dropped kk directions (57−64)(57-64) to be reconstructed in AA with no constraints, yet AA does not recover the dropped kk in a meaningful capacity in this case. So the question is what is the essential difference between r−kr-k and the learned kk, compared to the dropped kk, and how each of these components relate to the original subspace of the entry without compression.
  • FSC was evaluated only in r=32,64,or128r = 32, 64, or 128, and the gap between FSC and free continuation is not always narrowed with an increasing rr. Given that free continuation itself relies on the choice of rr, can you analyze in theory (relating to proposition 1 and appendix A.3) and in practice how the choice of rr and kk affects this gap (maybe using a joint sweep over rr and kk)? and how can this gap be closed between FSC and free continuation and the teacher model.
6 comments
AuthorsOpenReview ↗

Response to Reviewer JmmC [1/4]

(Part 1 of 4)

Thank you for the careful review and the detailed summary. We appreciate that you see the problem as significant for model compression and interpretability and find the experiments broad enough to support the hypothesis. Your two questions led to new experiments on what the learned directions are and on how r and k act.

In short.

  • Efficiency. All blocks run at full width, so the paper measures accuracy rather than speed and makes no efficiency claim.
  • Use of FSC. FSC measures which directions a subspace fixed at the entry lacks. A new experiment shows that the unmodified model uses these directions, mostly at the class token (Table R10).
  • Q1. Four results below relate the retained, dropped and learned directions (Appendices E.6 and I.5).
  • Q2. The grid over r and k is complete with 60 new runs (Tables R8 and R9). With the entry held fixed, more learned directions raise FSC in every row.

Weakness 1. The blocks are computed at full width, so the efficiency gain is unclear. All experiments evaluate the blocks at full width and project afterwards, as you note, so the paper measures accuracy rather than speed. The submitted introduction pointed to an operation count for a suffix running at reduced width. The revision removes this and states that we make no efficiency claim (p. 1, lines 46–47). The question of the paper comes before efficiency. It asks whether the last blocks can work inside a subspace that is known before they run. Appendix G (p. 40) remains as an operation count.

AuthorsOpenReview ↗

Response to Reviewer JmmC [2/4]

(Part 2 of 4)

Weakness 2. Why use FSC, and can its interpretability potential be demonstrated? FSC is not meant to replace the model. It is a static stand-in for the refit (Section 6), and its trained source measures which directions a subspace fixed at the entry lacks. The diagnostic shows that a subspace fixed at the entry fails the blocks that follow. FSC shows that the failure comes from a few missing directions and not from the size of the budget. Eight of 64 directions recover 61% and 91% of the accuracy that the entry basis loses on AugReg-B and CLIP-FT.

Its second use is the one you name. FSC turns the question of which directions the last blocks need beyond their input into something that can be measured, and the revised paper reports three results on it (Section 6, p. 8, lines 410–413; Appendix I.5, pp. 48–49). Much of what is missing comes from one source shared by all images. This source, a matrix A of 768 × 8 = 6,144 parameters, gives the recoveries above. The learned directions act mainly outside the row space of the entry, that is, outside the span of its 197 token vectors. When their component inside that span is removed, without retraining, they keep 92.1% and 97.4% of their gain in cross-entropy (CE). They concentrate on a few high-magnitude channels of the backbone, and models with the same pretrained backbone share these channels, 7 of 10 for CLIP-FT and zero-shot CLIP. Q1 below gives the details.

We also tested the learned directions in the unmodified model, with no restriction and no entry projection. We remove 8 directions from the four updates of the last two blocks. The table gives accuracy in % on the 5,000 confirmation images, the KL from the unmodified model in nats (0 means the same output), and the share of the model's own update energy that lies in the removed directions, in % (new Table R10).

Removed directionsAugReg-B acc.KLEnergyCLIP-FT acc.KLEnergy
None (unmodified model)83.980–85.040–
Learned complement of FSC81.130.1627.983.751.91310.7
Random, outside the entry support84.040.0020.685.040.0020.6
Entry directions 57 to 6483.680.0071.085.220.0031.0
Entry directions 1 to 877.960.49625.384.540.03931.1
Future directions of the image38.982.28028.753.142.85626.3
Leading directions of the update energy76.580.39023.783.482.11317.6
Highest-energy channel axes82.220.0729.984.600.0246.7

The last two blocks write 7.9% and 10.7% of their update energy into the 8 learned directions, against 0.6% for 8 random directions. Removing the learned directions lowers accuracy by 2.85 and 1.29 points and moves the output by 0.16 and 1.91 nats. Removing random directions barely changes the output (KL 0.002), and in all eight settings the learned directions cost more accuracy than random ones. Most of the effect sits at the class token. Removing the learned directions from the class token's update alone, one of 197 tokens, gives 0.14 and 1.88 nats, and removing them from the 196 patch tokens gives 0.03 and 0.003. So FSC identifies directions through which the last blocks write to the class token. The image's own future directions change the output more, as expected for directions read from its later states.

AuthorsOpenReview ↗

Response to Reviewer JmmC [3/4]

(Part 3 of 4)

Q1. Retained, dropped and learned directions. Take one image, r = 64 and k = 8.

  • Retained. The 56 leading singular directions of the entry. They are specific to the image and hold the entry state.
  • Dropped. The next 8 singular directions of the same entry. They add little to the entry and do not help the continuation. With them in place the suffix reaches 50.6% and 4.0% (the entry-basis arm).
  • Learned. 8 directions from one source shared by all images. For each image they are made orthogonal to all 64 entry directions and to the mean direction. They receive what the later blocks write.

Four results relate the learned directions to the entry.

  1. They do not return to the dropped directions when allowed. If the complement may also use entry directions 57 to 64, it places 1.2% (AugReg-B) and 1.8% (CLIP-FT) of its energy there, against 1.1% for a random subspace, and accuracy is nearly unchanged at 70.49% against 70.53% and 77.34% against 77.60% (Appendix E.6).
  2. They act mainly outside the entry, including its uncompressed row space. Removing from the trained complement its component inside the row space of all 197 entry tokens, without retraining, keeps 92.1% and 97.4% of the CE gain. Keeping only that component keeps 30.5% and 14.8% (Table 27). The later updates do write there. On the 2,000 diagnostic images the four updates of the last two blocks place 46% and 45% of their energy outside the row space of the entry.
  3. They concentrate on high-magnitude residual channels of the backbone. The learned projector puts 24.3% and 22.2% of its energy on 10 of the 768 coordinates, against 3.5% for random directions, and 11.6% to 24.3% across the eight settings. For AugReg-B and CLIP-FT, 5 and 6 of these 10 coordinates are among the 10 channels with the highest RMS at the entry. Models with a common pretrained backbone share several of these coordinates, 7 of 10 for CLIP-FT and zero-shot CLIP, and 5 of 10 for AugReg-B and AugReg2-B. A source trained from a random start has 3.6% on its top 10 coordinates before training and 21.8% and 16.6% after, with 7 of the 10 in common with the main source.
  4. They are not those channels. Using the 8 highest-RMS coordinate axes as the complement gives 50.2% and 29.9% on the 5,000 confirmation images, against 70.5% and 77.6% for FSC. The entry basis gives 50.6% and 4.0%, and 8 random other channels give 52.7% and 4.7%. Results 3 and 4 are reported in the new Appendix I.5.

In short, the retained directions hold the state, and the dropped ones are more of the same. The learned ones act mainly outside the entry's row space. They are tied to a few fixed high-magnitude channels, consistent with one shared source serving all images, and they are still not the channel axes themselves.

Reviewer_9zWkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues against a common inference in low-rank compression: activations are low-rank, so the model could run at that rank. Its point is that a subspace holding the current state need not hold what the later blocks write into the residual stream. To test this it freezes a trained ViT, fits one subspace to the activations at some depth, and projects every subsequent residual-branch output onto it. Accuracy drops sharply, and still drops when the subspace is large enough to hold the entire activation matrix, whereas a subspace fitted to the later activations costs almost nothing. The proposed method, FSC, reserves eight directions of the budget for a matrix learned once and shared across images, which recovers much of the loss. The basis is held fixed across the whole window, however, and that is what produces the failure: SliceGPT fits a separate basis to each block's own output, so what a block writes is already in the basis that follows it.

Critical strengths

A low-rank subspace that reconstructs the activations well can still be useless for running the model. Before trusting one, check that the model still works when forced to compute inside it. The paper shows the difference can be very large: eight directions out of sixty-four decide whether CLIP-FT sits at 3.9% or 82.5%.

Critical weaknesses
  • My main concern is that the setup does not correspond to what the compression methods it cites actually do. SliceGPT transforms the weights before slicing them and ASVD and SVD-LLM factorize them, so in each case the operators are rebuilt for the reduced space, whereas here the model is frozen and one basis is held fixed across the whole window. SliceGPT also sets each block's basis from that block's own output covariance, so what a block writes is already in the basis that follows. Since the failure reported here is exactly a block writing outside a basis fitted before it, I suspect much of the gap follows from freezing the basis across depth rather than from low rank. FSC would then be a static stand-in for per-block re-fitting, which fits its fall from 70.5% over two blocks to 30.4% over four. Block-varying spaces are named as future work but not tested.

  • it is not clear what the result is a result about. The eight settings are one architecture at one resolution, varying only pretraining and readout, and nothing in the propositions or the method is specific to vision. What the design does depend on is N < D, which Appendix A.1 states is what makes the full-rank control informative; a ViT-B/16 at 448 input has 785 tokens against a width of 768, so that control would be vacuous there. Five of the six works motivating the paper are language-model compression and the conclusion names language models as the next step, but none is tested, and in long context N >= D would again empty the control.

  • Several of the claims are weaker than stated. Proposition 2 is an expansion around epsilon=0 while every experiment uses epsilon=1, and Table 26 reports its leading term as "too small by a factor of 7 to 42," so the contribution claimed in the introduction, that importance is set by output sensitivity, is not supported. The 67-point upper bound in the abstract comes from MAE-B, which Appendix D attributes to its frozen-BatchNorm probe rather than to the phenomenon. And by Appendix G a two-block window is 16.7% of the model's multiply-adds with a 5.0% saving at r=64, for which the method reaches 70.5% against 83.0% unconstrained; over four blocks the saving doubles but accuracy falls to 30.4%, so the regime where it works is the regime where folding is not worth doing.

Questions for authors
  • How much of the gap survives if the subspace is re-fitted at each block, in the way SliceGPT does, rather than held fixed across the window? This looks like the cheapest experiment that would settle whether the effect is a property of low rank or of the fixed basis.

  • Does the gap appear in a model that has actually been compressed? Running the diagnostic on a SliceGPT- or ASVD-compressed model at a standard ratio would show whether the compression literature is affected by this at all.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9zWk [1/5]

(Part 1 of 5)

Thank you for a careful and precise review. Your summary states the point of the paper clearly, and the experiment you proposed in Q1 changed how we state the result. You were right that the loss comes from holding the basis of the entry fixed across depth. The revision states this in the abstract, the introduction and a new Section 5.3, and it revises the three claims you questioned.

The result now has a precise scope. Each state stays close to rank 64 for its own image, so in seven of our eight settings 64 per-image directions are enough if they are refitted after every update. A basis shared by all images needs about 576. Using the per-image rank therefore needs a per-image basis, and a computation that runs inside it must choose that basis before each block runs. The refit cannot do this, because it reads the block's full-width output. A basis chosen at the entry is the natural candidate. The paper shows that it misses directions that the following blocks write, and FSC measures which ones. We answer the two questions first and then the three weaknesses.

In short.

  • Q1. Refitting the basis after every attention and MLP update brings accuracy within 0.6 points of free continuation in seven of eight settings, over two blocks and over four. Refitting once per block leaves up to 38.6 points over two blocks and 76.0 over four. The paper now attributes the gap to holding the entry basis fixed.
  • Q2. We sliced the models with SliceGPT. Its own bases cost 3.7 and 3.0 points at the 25% ratio, and at that ratio a per-image subspace chosen at the entry loses more inside the sliced model than before slicing.
  • Block-varying complements. We tested two. Over four blocks, a complement per block adds 8.50 and 11.55 points to FSC and one full-width read at the midpoint adds 16.73 and 21.55. Among the arms we tested, only a basis refitted after every update reaches free continuation.
  • Scope. On GPT-2 the gap appears, and refitting removes 86% to 99% of the KL that a fixed basis adds at 64 and 128 directions. The comparisons at a fixed number of directions hold with N > D, including ViT-B/16 at 448 px.
  • Claims. Proposition 2 is now a heuristic in the appendix, the abstract reports the median of 12 points, and the paper makes no efficiency claim.

Q1. How much of the gap survives refitting the subspace, as SliceGPT does? Almost none, if the basis is refitted after every attention and MLP update, which is where SliceGPT places its bases. With a per-image basis of 64 directions refitted at every such site from the current state, AugReg-B reaches 82.74% and CLIP-FT 84.54% on the 5,000 confirmation images. Free continuation, where only the entry is projected, reaches 83.00% and 84.50%, and the entry basis held fixed in the same form reaches 50.76% and 58.22% (first table below). Over the last four blocks the refit reaches 83.02% and 84.96%, against 82.98% and 85.02% for free continuation.

Across the eight settings, the refit is within 0.6 points of free continuation in seven, over two blocks and over four. On MAE-B it is 2.7 and 4.5 points below, where the fixed basis loses 56.9 and 59.5 points (second table below).

Refitting once per block is not enough. After the MLP update or after the attention update, it reaches at best 66.08% on AugReg-B and 83.72% on CLIP-FT over two blocks, and 6.96% and 78.08% over four. Across the eight settings it loses 0.2 to 38.6 points over two blocks and 2.1 to 76.0 points over four (second table below).

At the complete row space, refitting is the identity, because every state lies in its own row space. With N < D that row space is smaller than the channel space, so the complete-entry control isolates the cost of holding the basis fixed. The revised paper attributes the gap to the fixed entry basis (Section 5.3, pp. 7–8, lines 359–402; Appendix I.1, pp. 43–46).

AuthorsOpenReview ↗

Response to Reviewer 9zWk [2/5]

(Part 2 of 5)

The table restricts the last two blocks with 64 directions unless stated. It gives top-1 accuracy in % and, in parentheses, the KL from the unmodified model in nats (0 means the same output). All projections use the token-norm rule, which keeps the channel mean and the norm of each projected token, so that only its direction changes. The paper's entry-basis arm and FSC project each attention or MLP output before it is added (update form). The other restricted rows project the state after each addition (state form), the form in which refitting is defined. Under raw projection the two forms are the same operation for a fixed basis, and the entry basis reaches 15.70% on CLIP-FT against 84.42% for free continuation (Table 27).

ProjectedBasis at the four sitesKnown in advanceAugReg-BCLIP-FT
–Free continuation (reference)–83.00 (0.029)84.50 (0.030)
UpdatesEntry basis (entry-basis arm of the paper)yes50.64 (2.369)4.02 (5.659)
FSC with 56 entry and 8 learned directionsyes70.53 (0.925)77.60 (0.807)
StatesEntry basisyes50.76 (2.271)58.22 (3.584)
Dataset-level PCA basis of each site (SliceGPT's choice)yes9.34 (3.756)2.56 (5.428)
The same with 576 directions (SliceGPT's 25% ratio)yes81.58 (0.087)83.78 (0.127)
Per-image basis refitted at every siteno82.74 (0.041)84.54 (0.030)

The second table repeats the comparison on all eight settings and refits less often. The protocol is the same, and the state after each site is projected. Every refits after every update, MLP and attn. refit once per block after that update, and never keeps the entry basis fixed (Tables R4 and R5 in the revision).

SuffixSettingFreeRefit everyRefit MLPRefit attn.Never
Two blocksAugReg-B83.0082.7466.0853.8050.76
AugReg2-B84.2684.3080.2882.8276.02
CLIP-FT-B84.5084.5481.3083.7258.22
AugReg-S79.4279.1670.9667.7460.60
AugReg-L84.8484.9684.3484.6083.90
DINO-B / linear71.9071.9466.5670.0664.66
MAE-B / linear57.5654.901.3018.960.64
CLIP-B / zero-shot67.8067.2835.0454.9832.08
Four blocksAugReg-B82.9883.026.960.564.86
AugReg2-B84.1684.2668.8073.289.96
CLIP-FT-B85.0284.9672.0078.0817.22
AugReg-S77.4477.3244.947.4416.40
AugReg-L84.9084.7282.8080.5677.54
DINO-B / linear74.4473.9862.7068.2448.74
MAE-B / linear59.7055.160.106.300.18
CLIP-B / zero-shot68.0467.727.9613.604.60

The same experiment shows what the fixed basis has that refitting lacks. Refitting evaluates each block at full width and takes the basis from the block's output, so it cannot be chosen before the block runs. Bases that can be chosen in advance for all images, as in SliceGPT, need about nine times as many directions. The next table applies SliceGPT's choice of basis to the last two blocks of the frozen model (Table R3, p. 44). Each site's basis is the uncentered PCA of that site's natural state over 512 calibration images, and the entry is projected onto its own PCA basis. The table gives top-1 accuracy in % on the 5,000 confirmation images, and the column heads give the number of directions. The unmodified models reach 83.98% and 85.04%.

ModelDataset-level basis64192384576
AugReg-BEntry projection only34.7473.1481.4683.22
Entry basis fixed6.6039.6070.3279.58
Bases of each site9.3444.7874.5481.58
CLIP-FTEntry projection only40.4078.4483.8284.76
Entry basis fixed3.9055.6880.3083.66
Bases of each site2.5644.0879.5883.78

At 64 directions a dataset-level basis already loses the entry, while a per-image basis keeps it. Projecting the entry onto its own 64 directions and then running the blocks freely gives 83.28% and 84.72%. Using the per-image rank requires a per-image basis, and for such a basis the two ends are fitting it after each update and fixing it in advance. The paper measures what fixing it in advance costs.

AuthorsOpenReview ↗

Response to Reviewer 9zWk [3/5]

(Part 3 of 5)

Q2. Does the gap appear in a model that has actually been compressed? Yes. We implemented SliceGPT for ViT-B following the official code: LayerNorm fusion, one PCA rotation per residual site computed block by block on 1,024 calibration images, and slicing of all twelve blocks, with no recovery fine-tuning. The result is a narrower network and not a projection inside the original one. At ratio 0 it reproduces the original outputs. On 5,000 confirmation images the sliced models reach 82.66%, 81.28%, 80.24% and 78.40% on AugReg-B at ratios of 10%, 20%, 25% and 30%, from 83.98%, and 84.22%, 82.86%, 82.02% and 80.94% on CLIP-FT, from 85.04%.

We then ran our diagnostic on the sliced model, with the sliced model as its own teacher. The table gives top-1 accuracy in % on 2,000 class-balanced diagnostic images, a different set from the 5,000 confirmation images, with the last two blocks restricted by the token-norm rule. The 0% columns are the fused and rotated model before slicing, whose arms are within 0.9 points of those of the original model.

ArmAugReg-B 0%AugReg-B 25%CLIP-FT 0%CLIP-FT 25%
The model itself85.0582.1086.1583.00
Complete entry basis held fixed68.8060.3552.9018.60
Free continuation at 64 directions84.1580.2585.6082.80
Entry basis held fixed at 64 directions (updates)54.6538.804.900.10
Per-image refit at 64 directions83.8579.8085.5082.85

With the complete entry basis the sliced AugReg-B loses 21.75 points, against 16.25 before slicing, and the sliced CLIP-FT loses 64.40 points, against 33.25. The increase is 5.50 points (95% interval 3.20 to 7.75) and 31.15 points (28.55 to 33.70). At ratios of 10%, 20%, 25% and 30% the loss is 15.90, 20.05, 21.75 and 23.80 points on AugReg-B and 50.30, 61.15, 64.40 and 67.30 points on CLIP-FT. Under raw projection the loss rises from 17.70 to 21.00 points and from 14.65 to 43.45 points at 25%, so the increase does not come from the token-norm rule. Per-image refitting stays within 0.6 points of free continuation at every ratio.

We chose SliceGPT because its basis changes at every residual site, which is the case your first weakness is about.

So the answer has two parts. SliceGPT's own bases cost 3.7 and 3.0 points at the 25% ratio. Inside the sliced model, a per-image subspace chosen at the entry loses more than before slicing, from the 20% ratio on in both models (Section 5.3; Appendix I.2 and Table R6, pp. 45–47).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response and Revision of the Paper

We thank the three reviewers for their careful reading and constructive reviews. All three recognize the central observation of the paper. A subspace that reconstructs a ViT state almost exactly can still fail the blocks that follow. Reviewer E2n9 finds the experiments well controlled and the evidence for the gap convincing, Reviewer JmmC sees the problem as relevant to model compression and interpretability and the experiments as broad, and Reviewer 9zWk writes that before trusting such a subspace one should "check that the model still works when forced to compute inside it", and that the paper "shows the difference can be very large".

Presentation. We apologize that the submitted version was hard to follow. Reviewer E2n9 listed the terms that the introduction used without explaining them. We have uploaded a revised PDF. Its abstract, introduction and conclusion are rewritten, the terms are defined in the introduction on one running example, and changed text is in blue. Pages and line numbers below refer to this PDF. We have also uploaded the supplementary code package, which now includes the code of the new experiments.

What the new experiments show. Every state along the suffix stays close to rank 64 for its own image, but the computation does not stay inside the subspace of its entry. Reviewer 9zWk asked how much of the gap survives when the subspace is refitted, as SliceGPT does. When a per-image basis of 64 directions is refitted after every attention and MLP update, which is where SliceGPT places its bases, accuracy stays within 0.3 points of free continuation (the entry is projected and the blocks then run without restriction) on AugReg-B and CLIP-FT. It stays within 0.6 points in seven of our eight settings, over two blocks and over four. Refitting once per block is not enough. The per-image basis chosen at the entry and held fixed loses 32.2 and 26.3 points on AugReg-B and CLIP-FT when it projects the states as the refit does, and 32.4 and 80.5 points when it projects the updates as in the main arm of the paper. Low-rank states therefore do not give low-rank dynamics, which is the distinction in the title. The refit reads each block's full-width output, so its basis is known only after the block has run. Bases fixed in advance for all images, as in SliceGPT, need about 576 of the 768 directions to come within 2 points of free continuation. The revision states the result for a per-image subspace chosen at the entry and held fixed, and it states how this setting differs from SliceGPT, ASVD and SVD-LLM, which rebuild the operators for the reduced space.

What this is for. For low-rank methods, the result gives a simple test. A subspace chosen before the blocks run should be checked by running the blocks inside it, because its reconstruction error does not reveal the gap. A method that computes at reduced width needs to refit its basis after every update or to supply the directions that the entry lacks. FSC shows that a few learned directions, from one source shared by all images, recover most of what a fixed entry subspace loses over two blocks.

Summary of the revision.

  1. Clarity (all reviewers). The abstract (p. 1, lines 13–32), the definitions and the four named comparison arms (pp. 1–2, lines 48–78), the contributions (p. 3, lines 119–125) and the conclusion (p. 10, lines 523–539) are rewritten.
  2. Refitting the basis (9zWk Q1). New Section 5.3 and Table R1 (pp. 7–8, lines 359–402). All eight settings in Table R4, and refitting once per block in Table R5 (Appendix I.1). The difference from SliceGPT is stated in the introduction (p. 2, lines 86–94), and from SliceGPT, ASVD and SVD-LLM in the related work (p. 3, lines 146–149).
  3. A model compressed with SliceGPT (9zWk Q2). Section 5.3, Appendix I.2 and Table R6 (pp. 45–47).
  4. More tokens than channels (9zWk). Section 5.1 (pp. 6–7, lines 322–325) and Appendix I.3 (pp. 46–47), with Table R7 for ViT-S/16 at 384 px and the text for ViT-B/16 at 448 px.
  5. The grid over r and k (JmmC Q2). Section 6.2 (p. 9, lines 478–479), Appendix I.4 and Tables R8 and R9 (pp. 47–48).
  6. What the learned directions are (JmmC). Section 6 (p. 8, lines 410–413), Appendix I.5 and Table R10 (pp. 48–49).
  7. How the subspace moves with depth. Appendix I.6 and Table R11 (pp. 49–50).
  8. A language model (9zWk). GPT-2 and GPT-2 medium, Appendix I.7 and Table R12 (pp. 50–51).
  9. Block-varying complements for FSC (9zWk). Appendix I.8 and Table R13 (pp. 51–52).
  10. Claims (9zWk, JmmC). Proposition 2 is now a heuristic (Section 4, p. 5, lines 249–263; Appendix A.4). The abstract reports the median of 12 points, and Section 5.1 explains the MAE-B value (p. 6, lines 301–304). The paper makes no efficiency claim (p. 1, lines 46–47; Appendix G).

The replies to each reviewer give the numbers. We are glad to answer any further question before the discussion closes.

AuthorsOpenReview ↗

Summary of the Paper and the Discussion

We thank the Program Chairs, the Senior Area Chair, the Area Chair and the three reviewers for their time, and the reviewers for their detailed reviews. As the discussion closes, this comment summarizes what the paper finds, what it contributes, and how the revision answers the reviews.

What the paper finds.

  1. A subspace that holds the state can still break the computation. Projecting a ViT state onto the 197 directions spanned by its tokens changes it by less than 10⁻⁶ in relative error. Yet confining the updates of the last two blocks to the same subspace lowers accuracy by 15.7 points on AugReg ViT-B and 33.2 on CLIP-FT ViT-B, with a median of 12 points over eight settings.
  2. The low rank is not the problem; holding the subspace fixed is. Each state stays close to rank 64 for its own image: refitting 64 directions after every attention and MLP update stays within 0.6 points of leaving the updates unrestricted in seven of eight settings. Low-rank states therefore do not give low-rank dynamics. Bases shared by all images need about 576 of the 768 directions.
  3. The missing directions are few and can be learned. Eight of 64 directions decide most of the outcome. Taken from the image's own later states, they raise CLIP-FT from 3.9% to 82.5%, while eight random directions reach 4.9%. FSC learns such directions once, as a source of 6,144 parameters shared by all images. It improves all eight settings and recovers 61% and 91% of the accuracy that the fixed subspace loses on AugReg-B and CLIP-FT. The unmodified model writes to these directions, mostly through the class token.
  4. The gap also appears outside the main setting. Inside a ViT sliced with SliceGPT at 25%, it grows from 16.25 to 21.75 points on AugReg-B and from 33.25 to 64.40 on CLIP-FT. On GPT-2 and GPT-2 medium it moves the output by 1.06 and 1.80 nats of KL.

What the paper contributes. A controlled diagnostic that separates how well a subspace holds the state from whether the blocks can compute inside it. From it follows a test for low-rank methods: a subspace chosen before the blocks run should be checked by running the blocks inside it, because its reconstruction error does not reveal the gap. FSC adds a measurement of which directions the computation needs beyond its input.

How the revision answers the reviews.

  • Reviewer E2n9 asked for a clearer presentation and listed the terms that the introduction left unexplained. The revision defines each of them on one running example and rewrites the abstract, the introduction and the conclusion.
  • Reviewer 9zWk proposed two experiments: refitting the basis as SliceGPT does, and running the diagnostic on a compressed model. We ran both. The refit sharpened the claim to a per-image subspace held fixed from the entry (new Section 5.3), and the gap persists in the sliced model. We also ran the other settings the reviewer raised and corrected the three claims the reviewer questioned.
  • Reviewer JmmC found that the sweep over r and k addressed the second question and that the analysis of the dropped and learned directions is insightful. On the significance of the result without the efficiency claim, our follow-up explains, with the findings above, why the result bears on low-rank compression.

The new experiments are in Appendix I (Tables R1 to R13), and their code is in the supplementary package. We thank the reviewers again. Their questions made the claims of the paper more precise.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.