acceptodds
Withdrawn from ICLR 2027

Turn the Bottleneck Knob: Token-Valve Path Explanations for Vision Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Turn the Bottleneck Knob (TBK) is a post-hoc explanation method for frozen Vision Transformers. It combines shared token gates on selected pre-attention branches, finite-step target cross-entropy optimization, and integration of absolute gate sensitivities along a linear opening path. The resulting scores measure unsigned sensitivity per unit gate coordinate; they do not identify minimal or necessary tokens. We give a shared-parameter derivative and a loss-change upper bound, distinguish these properties from causal faithfulness, and clarify why a finite-noise information-bottleneck motivation does not establish deterministic information compression. Existing ImageNet-1K results on ViT-B/16 and DeiT-B/16 show a competitive but mixed proxy-metric profile. A separate ViT-Base follow-up supports an Insertion gain over CoIBA with a paired confidence interval excluding zero; its ROAD, Deletion, and Infidelity intervals include zero. Smoothing materially affects the Insertion comparison. The evidence supports this particular attribution pipeline in the evaluated setting, with unresolved implementation and protocol details stated explicitly.

Primary area:
interpretability and explainable AI
Keywords:
Vision Transformer interpretability, token attribution, path explanations, token gating, integrated gradients, post-hoc explainability, explanation faithfulness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yvCHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes TBK, a post-hoc ViT explanation method using shared token gates and unsigned path-integrated gate sensitivities. The experiment results show only mixed gains on proxy metrics.

Critical strengths

I do not see clear strengths sufficient to support acceptance.

Critical weaknesses
  • The paper seems AI-generated and not carefully reviewed by humans. The manuscript contains many awkward audit-like sentences, e.g., "Unspecified records a missing fact, not a default value", "This is static source verification, not execution verification", and "no experiment was rerun". These passages make the paper read like a experiment log rather than a coherent research paper.

  • I found the mechanism difficult to follow throughout the paper. The motivation and interpretation of the designs need clearer explanation. In particular, it remains unclear (1) why the paper shares gates across layers rather than using independent layer-wise gates? (2) how branch-level sensitivity relates to token importance when residual pathways remain open? (3) why the paper uses unsigned sensitivities without endpoint weighting rather than signed or endpoint-weighted alternatives? (4) which loss and implementation produced the main results, given the discrepancy between logit-based and probability-input cross-entropy? (5) how much of the reported improvement comes from the attribution method rather than postprocessing (standardization, clamping, smoothing, resizing, min–max)?

  • The empirical gains are also weak and mixed.

  • The qualitative evidence is limited. Only a few selected examples are shown.

Questions for authors

See weaknesses.

Reviewer_rzNtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TBK, a post-hoc explanation method for frozen ViTs. A shared token gate α\alpha is multiplied onto the pre-attention LayerNorm output of blocks 4–11, a finite number of target-CE optimization steps yields a gate endpoint, and the absolute value of the gate gradient is integrated along a linear path to produce an importance map. Mathematically, the paper provides a triangle-inequality upper bound while explicitly stating that neither completeness nor faithfulness is guaranteed. On ImageNet-1K with ViT-B/16 and DeiT-B/16, TBK reports the highest mean ROAD and Insertion, but in the paired bootstrap only Insertion is significant, and this advantage depends substantially on smoothing post-processing. The implementation that generated the main results is absent from the archive, and no new experiments were conducted.

Critical strengths
  • The mathematical properties of the method (shared-gate derivative, path upper bound, the gradient difference between LlogitL_\text{logit} and LprobL_\text{prob}, dependence on coordinate units) are stated accurately and accompanied by counterexamples.
  • The paper does not hide discrepancies between the implementation and the theory (truncated trapezoidal rule, double softmax, the effect of post-processing) and describes the scope of its results without overstatement.
  • Ablations on smoothing, initialization, layer selection, and repeated runs are provided, allowing readers to judge which part of the reported performance is due to post-processing.
Critical weaknesses

[W1] The contribution cannot be identified. The authors themselves acknowledge that shared gates were already used by CoIBA and deterministic gating by NIB (Sec. 2), and the information-bottleneck motivation in the title is invalidated by the authors in Appendix A. The remaining contribution is the combination of "CE endpoint optimization + unsigned path integration", yet there is no experiment comparing the central design choice, the unsigned score sis_i, against signed IG or the endpoint-weighted score α^isi\hat{\alpha}_i s_i. The authors state in Sec. 3.3 and Appendix F that this comparison "remains necessary". The paper thus defers the justification of its own contribution to future work.

[W2] The methodology does not define a single method. Instead of describing a design, Sec. 3 reports the result of statically reading a "recovered server source". The objective is never fixed: the paper does not commit to LlogitL_{\text{logit}} or LprobL_{\text{prob}} and writes "We use LL for the scalar loss actually chosen", while Sec. 4.1 proves that the two gradients are generally different vectors. Two integration rules coexist, the "intended" full-interval right-endpoint rule and the "recovered" truncated trapezoidal rule over [1/m,1][1/m, 1]. Algorithm 1 is titled "recovered numerical ordering". A reviewer cannot determine which method is being proposed, and the referenced source cannot be inspected.

[W3] The main results are irreproducible, and the authors admit this. Statements such as "The original archive contains no main-result implementation", "No experiment was rerun", and "code-to-result bindings remain unresolved" recur throughout the text. The submitted tbk.py only generates attribution maps, hardcodes opt_steps=20 which differs from the Table 2 setting (T=50T=50 for ViT, T=30T=30 for DeiT), contains no evaluation code or baseline implementations, lacks the input tensors img_batch.pt and label_batch.pt, and generate_attributions.py references an undeclared args.dataset so that execution fails immediately with an AttributeError. The README merely states that this was "preserved". Moreover, the follow-up setting described in Sec. 5.2 (batch size 8, batch-mean CE) does not match the submitted code (batch size 1), suggesting that the source referenced in the paper and the submitted source are yet another pair of different versions.

[W4] The empirical evidence is very weak. The main table reports no standard deviations and no image count. In the paired bootstrap (Table 3), Insertion is the only metric whose confidence interval excludes zero, and according to Table 4 the raw-map Insertion is 0.5760.576, only 0.00560.0056 above CoIBA (0.5710.571), with no confidence interval provided for this raw difference. The only significant result therefore arises from the 3×33 \times 3 mean-smoothing post-processing. Furthermore, Table 5 shows that the all-open gate with no optimization at all (ROAD 0.2720.272, Insertion 0.5820.582) is only marginally behind the learned gate (0.2850.285, 0.5940.594), casting doubt on whether Stage A is needed. LeGrad, TIS, MDA, TokenTM, and LibraGrad, all cited in related work, are absent from the numerical comparison.

[W5] The theoretical results do not support the method. Lemma 1 and Theorem 1 are at the level of the chain rule and the triangle inequality, and the authors' own counterexamples (L=C+α1−α2L = C + \alpha_1 - \alpha_2 with ΔL=0\Delta L = 0 but s1+s2=2s_1 + s_2 = 2, and L=C+εsin⁡(kα)L = C + \varepsilon \sin(k\alpha) with arbitrarily large unsigned mass) show that the unsigned score can grow independently of the loss change. The bound holds "for any qualifying endpoint, including random ones", so it says nothing about TBK's particular endpoint. Notation is also unclear: the shape of h(ℓ)h^{(\ell)} is not specified, and it is used as a matrix in Eq. (1) but as a per-token vector hi(ℓ)h_i^{(\ell)} in Eq. (5).

[W6] Readability and maturity. The paper is very difficult to read. The methodology section is filled with code archaeology and the experiments section with disclaimers about evidence of uncertain provenance, so that a reader struggles to the end to determine what the proposed method is and what has been shown. The repeated vocabulary of "existing", "retained", "recovered", "unverified", and "unresolved" conveys that the authors are not in control of their own implementation and results. The manuscript reads as a revision in which overstated claims from a previous version were retracted, leaving essentially nothing behind. The candor about limitations is acknowledged, but candor does not substitute for a contribution. Overall this is a highly immature manuscript that does not reach the standard of a conference submission. It should be resubmitted only after providing a single definitive method specification, a reproducible implementation, and controlled experiments on the core design choices.

For these critical reasons, I regret that my recommendation is to reject.

Questions for authors

Please refer to the Weakness section.

Reviewer_z3zdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a post-hoc explanation method for frozen vision transformer (ViT) backbones, which is based on a combination of bottleneck attribution and path-integration methods. The bottleneck attribution is achieved via the introduction of learnable token gates (in a minimalistic configuration denoted as "shared gates"), which are trained on a given image to minimize cross-entropy loss (stage A) and the resulting framework serves as an input for integrating path gradients producing attribution scores for patch tokens (stage B). The paper evaluates this framework on ImageNet-1K dataset for two ViT models using faithfulness metrics and compares the results to other existing feature-attribution methods. The considered method shows the best results in terms of ROAD and Insertion metric mean values across the dataset on both ViT models. While the paper gives a mathematical description of the method, the main contribution rests on empirical comparisons, as is noted by the authors themselves.

Critical strengths

To my knowledge, the paper introduces for the first time the combination of learnable bottleneck attributions with path-integrated gradients and demonstrate empirically that it can be competitive with other explanation methods.

Critical weaknesses

I have three major concerns that prevent from recommending acceptance in its current form.

  1. It is not clear from the text, which experiments/scripts/protocols/calculations the authors are reproducing. There are numerous references in the text to "source manuscript" (line 091), "server source" (lines 093, 323), "recovered source" (lines 117, 334, 353), "immutable ZIP" (line 400), etc. without a proper reference. The paper lacks "Reproducibility Statement" section to clarify these details. If the authors refer to their own code or other works I would suggest to at least add a note in the introduction, explaining what these sources are while remaining anonymous prior to publication. If the authors are using texts or materials authored by other people it requires proper citations or references.

  2. The empirical evidence seems limited and partly confounded. Table 2 does not report any variance. The paired test results (Table 3) are available only for ViT (not for DeiT) and one competitor (CoIBA) and support a modest gain in Insertion metric, while the confidence intervals for ROAD, Deletion and Infidelity include 0. The results in Table 4 suggests that a significant portion of this gain depends on additive smoothing. It also seems that the baselines did not receive the same postprocessing as TBK.

  3. I am not convinced that Stage A is properly justified theoretically or empirically. The authors claim that the protocol in stage A does not establish necessity and minimum sufficiency, which does not align well with the bottleneck information principle as far as I understand it, although this is a minor issue. Theorem 1 holds for any endpoint, even a randomly assigned one, and this theorem, as well as the considerations around it do not seem to justify the optimized endpoint. The results in Table 5 suggest that the learned endpoint configuration beats all-open or random gate configurations in terms of mean ROAD, however a better justification requires paired tests of trained and non-trained endpoints. Furthermore, it would be nice to see the results of mean-matched uniform endpoint (T=0).

Questions for authors

The first 5 questions follow up on the criticism described in "Critical Weaknesses" section.

  1. What are the various entities in the text described as "source manuscript", "server source", "recovered source", etc?
  2. Please report paired tests for all the metrics for both ViT models against a few strongest baseline methods in addition to CoIBA.
  3. If possible, please evaluate all (or at least several strongest) baseline methods with identical postprocessing.
  4. Please add a mean-matched uniform endpoint (T=0) results to Table 5.
  5. Please add the results for paired tests between learned and fixed endpoint configurations to further justify that gate optimization improves faithfulness.
  6. What is the definition of FPS used for efficiency comparison (Table 8)?

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

Related papers

Discussion (0)

Sign in to comment.