acceptodds
Under review as a conference paper at ICLR 2027

Gauge-Equivariant Conductance

Abstract

Internal attribution explains predictions by assigning credit to hidden units. Yet a network function does not determine hidden coordinates: a compensated invertible affine change can preserve outputs while changing coordinate scores. We formulate this as an identifiability problem for function-level interventions. For conductance, we derive the similarity law under the full affine gauge and show that, for nonscalar conductance matrices, every trace-compatible coordinate attribution vector is attainable under a function-preserving gauge; completeness therefore constrains only the trace. We replace coordinate credit with isolated spectral subspaces defined by Riesz projectors, whose interventions commute with the hidden gauge, and evaluate this criterion with a frozen protocol across vision and language modalities. Specifically, gauges change conventional coordinate winners, while spectral scores remain invariant and dominant clusters yield larger target-conditioned effects than matched-rank comparators; they also beat the entire complement on a text holdout. Attack paths show conditional magnitude but mixed direction, with near-modulus ties failing the isolation gate. Thus, spectral subspaces provide function-compatible intervention targets, while coordinate conductance remains parameterization-relative.

Primary area:
interpretability and explainable AI
Keywords:
gauge equivariance, spectral interventions, adversarial attacks.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ojhEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors show that function-preserving changes of coordinates can redistribute per-neuron conductance arbitrarily, subject to its fixed total. As a replacement, they propose intervening on the dominant eigenspace of the conductance matrix. This intervention is coordinate-invariant. In small models it has a larger logit effect than a same-rank spectral comparator.

Critical strengths
  1. I find the matrix formulation useful. Writing conductance as the diagonal of a path-integrated matrix turns completeness into a trace identity and makes the omitted cross-coordinate terms explicit. The similarity law then lets the authors apply Fillmore's theorem to characterize all attainable coordinate attributions for nonscalar CC. This gives a precise account of the ambiguity, rather than just an example of scores changing under a particular transformation.

 

  1. I also appreciate the care in the experimental protocol. The selection rules, comparators, and thresholds are fixed in advance, and the authors use held-out seeds and compute uncertainty over paths rather than treating gauge draws as independent evidence. Tracking eigengaps, projector norms, and conditioning makes the numerical results easier to assess. The mixed and negative results are reported too, including inconsistent directions on attack paths and the two cells that fail the isolation criterion.
Critical weaknesses
  1. My main conceptual reservation is the choice of equivalence class. The analysis allows any M∈GL(d)M\in\mathrm{GL}(d) at the hidden cut. As the authors acknowledge in Section 3.2, doing this at a nonlinear layer may require inserting an affine interface; it need not be a reparameterization within the original architecture. I accept the theorem for this enlarged class, but I am not yet convinced that it is the right class for assessing claims about the network's native neurons.

    Under the usual permutation and positive-rescaling symmetries of ReLU networks, corresponding neurons retain their conductance scores. For a permutation matrix Π\Pi and a positive diagonal matrix DD,

    = \Pi\,\operatorname{diag}(C).$$ So the coordinate ambiguity established under full $\mathrm{GL}(d)$ does not arise under these transformations. I also think there are substantive reasons to study the native basis. Godfrey et al. (2022) relate neuron axes to the activation's symmetry group, and their Section 6, together with the results of Bau et al. (2017), shows that rotating post-activation representations can reduce unit interpretability. Whittington et al. (2023) give another reason to expect axis-aligned structure: single-factor selectivity can arise under nonnegativity and energy constraints. There is empirical evidence of reproducible native units across models as well (Li et al., 2016; Dravid et al., 2023). More recently, Dravid et al. (2026) report that the shared Rosetta population becomes more selective with scale even as its fraction of all neurons decreases. None of this contradicts the affine theorem. It does make native-neuron organization a reasonable object to explain, though. I would like the authors to say which uses of neuron attribution need full affine invariance, and why invariance under the architecture's own symmetries would not suffice. The same question applies to the proposed replacement: smooth invertible nonlinear recodings can preserve $F$ while generally changing the spectral scores and projector intervention. The spectral construction therefore also depends on a choice of equivalence class, and I think that choice needs more motivation.

 

  1. I am not sure the utility experiments distinguish identifying a useful internal feature from simply undoing most of the displacement from baseline. For a straight hidden path, writing gˉ\bar g for the gradient averaged along that path gives

    C=Δz gˉ⊤.C = \Delta z\,\bar g^\top.

    If tr⁡(C)≠0\operatorname{tr}(C)\ne 0, the projector onto its nonzero eigenvalue is

    \qquad P\Delta z = \Delta z.$$ The intervened state $z_0+(I-P)\Delta z$ is then exactly $z_0$. In this case, a rank-one intervention removes the entire change from baseline to input. The paper notes the related rank-one case for affine downstream $G$ in Appendix A.2.3. I do not think this makes the general finite-ablation comparison circular. But it gives a concrete alternative explanation that I would want to rule out in the experiments, especially the text results. Table 16 reports dominant-to-runner-up effect ratios of roughly 195 to 1456, and the separate complement experiment gives median effects of 2.21 versus 0.004 logits. Those numbers make me wonder how much of the dominant component's effect comes from nearly resetting the representation to baseline. A useful check would be to report the residual displacement $$\frac{\lVert(I-P)\Delta z\rVert}{\lVert\Delta z\rVert}$$ in the native basis, together with the difference between the intervened output and the baseline output. I would also find the signed ratio $\operatorname{tr}(PC)/\operatorname{tr}(C)$ informative where the denominator is not close to zero. The comparisons I am missing are a full baseline reset, matched-rank native units selected by conductance, and directions selected by other attribution methods. Reporting intervention magnitudes alongside ranks would help establish how much of the advantage comes from the chosen subspace rather than the size of the perturbation.
Questions for authors
  1. How much of the dominant-component ablation effect comes from approximately returning to baseline? On the existing evaluation paths, could you report ∥(I−P)Δz∥/∥Δz∥\lVert(I-P)\Delta z\rVert/\lVert\Delta z\rVert in the native basis and check whether the utility advantage remains when this residual is substantial? I would find comparisons with a full baseline reset and matched-rank native units selected by conductance particularly helpful here, with intervention magnitudes reported as well.

 

  1. I would like to understand the intended interpretability task more concretely. In what setting is full affine invariance necessary, beyond invariance under the original architecture's symmetries? Could you give an example where the spectral explanation tells us something useful that native-neuron conductance misses?
7 comments

Thank you for your careful review and mathematical analysis. We appreciate your positive comments on the matrix formulation of conductance, the Fillmore orbit characterization, and the experimental design. Below we address each of the weaknesses and questions you raised.


Response to Critical Weakness 1

We agree that the equivalence class must match the interpretability question (concrete empirical comparison in Response to Question 2, Table R3):

  • Native group (ΠD\Pi D) vs. linear group (GL(d)GL(d)): Under permutation and positive scaling (ΠD\Pi D), diag⁡((ΠD)C(ΠD)−1)=Πdiag⁡(C)\operatorname{diag}((\Pi D)C(\Pi D)^{-1}) = \Pi\operatorname{diag}(C), so ΠD\Pi D suffices for studying axis-aligned unit selectivity within a fixed parameterization (Bau et al., 2017; Godfrey et al., 2022; Whittington et al., 2023) or shared Rosetta neurons (Li et al., 2016; Dravid et al., 2023). Full GL(d)GL(d) equivariance is required when comparing or transferring linear subspace interventions across representations linked by a dense invertible linear map (such as cross-checkpoint stitching or shared-channel reparameterizations, where Dravid et al., 2026 also observe that axis-aligned Rosetta fractions decrease with scale relative to distributed subspaces).
  • Why GL(d)GL(d) rather than smooth nonlinear diffeomorphisms (Diff(Rd)\text{Diff}(\mathbb{R}^d)): Our intervention framework presupposes a fixed affine space, baseline z0z_0, straight displacement Δz=z1−z0\Delta z = z_1 - z_0, and linear projection (I−P)Δz(I-P)\Delta z. Pushing a linear subspace projection forward through a nonlinear diffeomorphism ϕ\phi generally yields a curved manifold intervention rather than a linear subspace projection in the new coordinates.

Response to Critical Weakness 2

Your rank-one derivation for an uncontracted straight path (C=Δz gˉ⊤  ⟹  PΔz=ΔzC = \Delta z\,\bar{g}^\top \implies P\Delta z = \Delta z when tr⁡(C)≠0\operatorname{tr}(C)\neq 0) is exact, and from an output-logit perspective our data partially support your concern while clarifying where the spectral subspace adds value beyond a reset (full diagnostics in Response to Question 1, Tables R1–R2):

  • Representation vs. output reset: On the shallow AG News CLS vector cut (z∈R128z \in \mathbb{R}^{128}), gradients rotate by only 2.0%2.0\%, giving effective rank 1.21, native residual ∥(I−P)Δz∥2/∥Δz∥2=0.041\lVert(I-P)\Delta z\rVert_2/\lVert\Delta z\rVert_2 = 0.041, and 0.2% remaining logit gap (confirming an approximate representation and output reset in Table 16). On our spatially or token-contracted cuts (Cch=∑ℓ=1LΔzℓgℓ⊤C_{\text{ch}} = \sum_{\ell=1}^L \Delta z_\ell g_\ell^\top, effective rank 9.72–45.2), the hidden state does not collapse to z0z_0 (192/192 vision and Layer-0 paths have native residual ≥0.75\ge 0.75, medians 0.999 and 1.024), although in output space the rank-1 ablation removes 79.4% of the full-reset logit gap on pooled vision (leaving 1.80 of 10.31 logits, or 20.6%) and 65.6% on NLP Layer-0 (leaving 34.4%).
  • Direction vs. perturbation-norm controls (Table R2): At matched rank and Euclidean norm, the dominant subspace beats the average path-gradient subspace g^g^⊤\hat{g}\hat{g}^\top on 87.5% (vision) and 91.7% (Layer-0) of paths and beats the ℓ2\ell_2-norm-matched runner-up on 93.3% and 91.7%. When scaled down to the single-channel native ℓ2\ell_2 norm (Dom. at Native ℓ2\ell_2 Norm vs. Native Top-rr), it wins 61.7% pooled on vision (66.7%–75.0% on ResNet-18) and 75.0% on Layer-0, but drops to 54.2% on WRN-28-2 and 41.7% on PlainConvNet (showing higher per-unit-norm directional sensitivity on ResNet-18/Layer-0, vs. larger multi-channel displacement norm on WRN/PlainConvNet).

Response to Question 1

Tables R1 and R2 report the requested diagnostics across all 120 vision holdout paths (layer3 on ResNet-18, 8×88\times 8; layer2 on WRN-28-2, 16×1616\times 16; features[:14] on PlainConvNet, 8×88\times 8), 72 deep AG News Layer-0 paths (64 tokens), and 144 shallow CLS paths. Note that because CC is non-normal, I−PI-P is an oblique projector (P2=P,P⊤≠PP^2=P, P^\top \neq P), so ∥(I−P)Δz∥2/∥Δz∥2\lVert(I-P)\Delta z\rVert_2/\lVert\Delta z\rVert_2 can slightly exceed 1.01.0; likewise tr⁡(PC)/tr⁡(C)\operatorname{tr}(PC)/\operatorname{tr}(C) can slightly exceed 1.01.0 via spectral cancellation. Hidden IG integrates along the straight hidden segment (max relative completeness error 0.85%), and Grad×Act\text{Grad}\times\text{Act} uses z1,i∂iG(z1)z_{1,i}\partial_i G(z_1) summed over positions.

Table R1: Native-basis residual displacements, signed trace ratios, perturbation magnitudes, and baseline reset diagnostics

Setting & Cut (dd, median rank rr)Paths (Residual ≥0.75\ge0.75)Native Residual Ratio [IQR]Signed Trace Ratio [IQR]Perturbation Norms (Dom. / RU / Native)Full Baseline ResetIntervened-to-Baseline Gap (Paired %)
Vision Pooled (5 cells)120 (120)0.999 [0.928, 1.103]0.722 [0.570, 0.906]10.16 / 9.38 / 2.4810.31 logits1.80 logits (20.6%)
ResNet-18 CIFAR-10 (d=256,r=1d=256, r=1)24 (24)0.938 [0.885, 1.053]0.711 [0.605, 0.955]10.09 / 7.94 / 2.709.32 logits1.71 logits (20.6%)
WRN-28-2 CIFAR-10 (d=64,r=1d=64, r=1)24 (24)1.037 [0.945, 1.111]0.623 [0.548, 0.691]18.05 / 15.80 / 4.9514.24 logits3.06 logits (22.5%)
PlainConvNet CIFAR-10 (d=128,r=1d=128, r=1)24 (24)1.071 [0.935, 1.166]0.395 [0.207, 0.844]7.05 / 9.48 / 1.6618.13 logits7.55 logits (40.0%)
ResNet-18 CIFAR-100 (d=256,r=1d=256, r=1)24 (24)0.958 [0.928, 1.027]0.796 [0.658, 0.998]11.58 / 10.08 / 2.6211.79 logits0.93 logits (8.6%)
ResNet-18 SVHN (d=256,r=1d=256, r=1)24 (24)1.051 [0.947, 1.121]0.820 [0.709, 1.009]8.24 / 6.16 / 1.977.64 logits1.40 logits (20.4%)
NLP Layer-0 Cut (Deep, d=128,r=1d=128, r=1)72 (72)1.024 [0.997, 1.137]0.910 [0.729, 1.144]32.11 / 45.71 / 7.052.23 logits0.76 logits (34.4%)
NLP CLS Cut (Shallow, d=128,r=1d=128, r=1)144 (0)0.041 [0.022, 0.060]1.003 [0.991, 1.023]7.87 / 0.29 / 1.552.34 logits0.0039 logits (0.2%)

Table R2: Target-logit effects (medians) and paired win rates vs. norm-matched, native, and attribution directions

SettingDominant Spectralvs. Runner-Up (Raw / Norm-Matched)vs. Native Top-rr Cond.Dom. at Native ℓ2\ell_2 Norm vs. Native Top-rrvs. IG Top-rrvs. Grad×Act\text{Grad}\times\text{Act} Top-rrvs. Path-Grad Subspace g^g^⊤\hat{g}\hat{g}^\topvs. Full Complement (I−PI-P, rank d−rd-r)
Vision Pooled (5 cells)8.302.61 (90.8%) / 3.02 (93.3%)1.18 (97.5%)1.45 vs. 1.18 (61.7%)1.48 (95.8%)1.02 (97.5%)5.00 (87.5%)6.80 (70.8%)
ResNet-18 CIFAR-106.781.37 (87.5%) / 1.85 (83.3%)0.28 (95.8%)0.32 vs. 0.28 (66.7%)0.67 (91.7%)0.38 (95.8%)2.04 (87.5%)1.86 (83.3%)
WRN-28-2 CIFAR-109.543.23 (95.8%) / 4.37 (100.0%)1.79 (100.0%)1.80 vs. 1.79 (54.2%)1.94 (100.0%)1.16 (100.0%)4.36 (87.5%)9.78 (50.0%)
PlainConvNet CIFAR-1012.634.32 (70.8%) / 4.20 (83.3%)2.05 (91.7%)1.93 vs. 2.05 (41.7%)3.59 (87.5%)2.80 (91.7%)6.84 (79.2%)13.44 (50.0%)
ResNet-18 CIFAR-10011.253.28 (100.0%) / 3.86 (100.0%)1.12 (100.0%)1.76 vs. 1.12 (75.0%)1.53 (100.0%)1.06 (100.0%)7.17 (91.7%)7.21 (87.5%)
ResNet-18 SVHN6.081.50 (100.0%) / 2.57 (100.0%)0.81 (100.0%)1.23 vs. 0.81 (70.8%)1.03 (100.0%)0.60 (100.0%)4.67 (91.7%)4.74 (83.3%)
NLP Layer-0 Cut1.760.41 (93.1%) / 0.25 (91.7%)0.15 (98.6%)0.27 vs. 0.15 (75.0%)0.15 (98.6%)0.03 (100.0%)1.03 (91.7%)0.37 (91.7%)
NLP CLS Cut2.340.0046 (100.0%) / 0.12 (100.0%)0.14 (100.0%)0.38 vs. 0.14 (94.4%)0.14 (100.0%)0.08 (100.0%)2.20 (93.8%)0.0047 (100.0%)

Response to Question 2

Full GL(d)GL(d) invariance is required when transferring subspace attributions across checkpoints aligned by a dense linear representation map (Mstitch∈GL(d)M_{\text{stitch}} \in GL(d)), while even in a single unmodified checkpoint (M=IM=I) the spectral projector identifies a distributed multi-channel coalition that requires ablating 11.0–32.5 native channels to match (Table R3):

  1. What spectral attribution captures at M=IM=I: Because any dense non-axis-aligned operator in dimension dd naturally has high off-diagonal Frobenius mass (98.4%–99.6%), the stronger test at M=IM=I is how many native channels must be ablated together to match a rank-1 spectral intervention. Across 16 paths per vision model and 24 Layer-0 paths, the leading eigenvector spans PR∈[19.9,45.5]\mathrm{PR} \in [19.9, 45.5] channels, and cumulative native top-kk ablation (k=1,…,dk=1,\dots,d) requires a median of 11.0–32.5 native channels among matched paths to reach the rank-1 spectral effect (with 5/16 PlainConvNet and 7/24 Layer-0 paths never matching at any kk).
  2. Where full GL(d)GL(d) invariance matters in practice (Seed 2026→20272026 \to 2027 transfer): Procrustes-regularized stitching maps z(2027)≈Mstitchz(2026)+cz^{(2027)} \approx M_{\text{stitch}} z^{(2026)} + c fitted on held-out development activations are dense (κ∈[3.21,8.04]\kappa \in [3.21, 8.04], 74.4%–93.7% non-peak row mass). While exact commutation across MstitchM_{\text{stitch}} inside the stitched model FstF_{\text{st}} follows directly from our theorem (100% top-1 flip vs. 0.00 spectral discrepancy), the practical test is transferring source explanations between the two original checkpoints by transporting both the source spectral projector (S=MstitchP(2026)Mstitch−1S = M_{\text{stitch}} P^{(2026)} M_{\text{stitch}}^{-1}) and the source native top-rr projector (A=MstitchQnative(2026)Mstitch−1A = M_{\text{stitch}} Q_{\text{native}}^{(2026)} M_{\text{stitch}}^{-1}) through MstitchM_{\text{stitch}}. Raw transfer SS vs. AA wins 95.8%–97.2% across all four models, and norm-inflated A-up retains 66.7%–100.0% wins on ResNet-18, WRN-28-2, and AG News (50.0% on PlainConvNet); conversely, scaling SS down to the single-channel norm (S-down) retains 75.0% wins on ResNet-18 and 100.0% on AG News, but drops to 41.7% on WRN-28-2 and 33.3% on PlainConvNet (showing directional alignment on ResNet-18/AG News vs. multi-channel norm dominance on WRN/PlainConvNet).

Table R3: Native-basis (M=IM=I) off-diagonal coalition analysis and cross-seed linear stitching (Mstitch∈GL(d)M_{\text{stitch}} \in GL(d), Seed 2026→20272026 \to 2027)

Model & CutDim ddNative Off-Diagonal Frobenius RatioNative Leading-Eigenvector PRDominant vs. Native Top-1 / Top-5 / Top-10 EffectFirst Matching Native kk (Matched Paths)Median Non-Peak Row Mass (κ\kappa)Interface Flip / Jaccard / Spectral Effect DiscrepancyTransfer SS / AA (Raw; A-up; S-down Win %)
ResNet-18 (layer3)25699.6%27.4 ch.6.51 vs. 0.22 / 0.79 / 1.2132.5 ch. (16/16 matched)84.0% (κ=4.48\kappa=4.48)100.0% / 0.000 / 0.007.16 / 0.22 (95.8%; 91.7%; 75.0%)
WRN-28-2 (layer2)6498.4%19.9 ch.10.34 vs. 1.54 / 4.54 / 8.5314.0 ch. (16/16 matched)79.3% (κ=3.21\kappa=3.21)100.0% / 0.000 / 0.006.58 / 1.00 (95.8%; 66.7%; 41.7%)
PlainConvNet (features[:14])12899.5%43.8 ch.12.76 vs. 1.99 / 7.31 / 11.0911.0 ch. (11/16 matched)74.4% (κ=3.90\kappa=3.90)100.0% / 0.000 / 0.009.84 / 3.16 (95.8%; 50.0%; 33.3%)
AG News (CLS stitch / L0 audit)12899.5%45.5 ch.2.44 vs. 0.096 / 0.51 / 0.8631.0 ch. (17/24 matched)93.7% (κ=8.04\kappa=8.04)100.0% / 0.000 / 0.002.00 / 0.057 (97.2%; 100.0%; 100.0%)
Reviewer_dspUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies whether hidden-layer conductance is identifiable under function-preserving changes of coordinates. For a fixed hidden cut, input path, and scalar target, it derives a similarity law under compensated affine transformations. Applying Fillmore’s theorem shows that every coordinate attribution vector with the same trace is attainable when the conductance matrix is nonscalar. The paper then constructs equivariant interventions using Riesz projectors onto isolated spectral clusters and evaluates their numerical behavior and intervention effects in vision and text models.

Critical strengths
  1. The paper gives a complete characterization of coordinate-attribution non-identifiability. It describes the full set of attainable attribution vectors rather than only showing individual changes in neuron rankings, and clarifies the limits of completeness.

  2. Spectral projectors provide an operational equivariant intervention. The vision and text experiments examine intervention effects as well as algebraic identities. The authors clearly report near-tie cases, mixed directions of attack effects, and the scope of their conclusions.

Critical weaknesses
  1. I remain unconvinced that the theoretical advance is substantial enough for acceptance. The similarity law follows from the chain rule, the prescribed-diagonal characterization uses Fillmore’s theorem, and projector equivariance follows from the resolvent identity. Applying these results to conductance identifiability is valuable, but much of the advance lies in formalizing the problem and combining established tools. My concern is the significance of this advance, rather than the correctness of the derivations.

  2. The evidence for the practical benefit of the intervention rule remains centered on local spectral comparisons. The vision utility experiments mainly compare against the same-rank runner-up, while the whole-complement comparison comes from the AG News seed-2029 holdout. These results support the authors’ claims about target-conditioned intervention effects, but practical gains over a broader range of equivariant alternatives remain insufficiently established. Appendix A.3.7 explains the comparator choice and acknowledges the scope, so my concern is the breadth of the contribution rather than an invalid comparator or an omitted explanation.

Overall assessment: I lean toward weak rejection. The complete characterization of conductance coordinate non-identifiability is valuable, and the equivariance argument for spectral interventions is clear. However, the theoretical advance mainly combines established matrix results for this problem, while the practical benefit of the intervention rule is supported primarily by local spectral comparisons. The authors explain these scope limitations, but the depth of the new contribution and its practical value have not yet sufficiently convinced me to support acceptance.

Questions for authors

Beyond the reported runner-up and whole-complement comparisons, are there any existing results comparing the proposed rule with other equivariant selection rules? If so, please briefly summarize them, as they could change my assessment of the methodological contribution. If not, I would regard this as a direction for future work.

6 comments

Thank you for reviewing our paper and for your positive assessment of its soundness, clarity, and characterization of coordinate non-identifiability and spectral interventions. Below we address the two weaknesses and your question point by point.


Response to Critical Weakness 1

We agree that the chain rule, Fillmore's theorem, and the resolvent identity are classical; what is non-trivial is why lifting internal attribution from a vector to an operator forces a non-normal, oblique spectral geometry where standard orthogonal decompositions (SVD, PCA, or symmetric parts) fail GL(d)GL(d) equivariance:

  1. Why operator lifting is non-trivial: tangent–cotangent duality and the failure of orthogonal SVD/PCA. Along a hidden path, velocity z˙(t)\dot{z}(t) transforms contravariantly (Mz˙M\dot{z}) in the tangent space while the downstream gradient ∇G(z(t))\nabla G(z(t)) transforms covariantly (M−⊤∇GM^{-\top}\nabla G) in the cotangent space. Their path integral C=∫01z˙(t)∇G(z(t))⊤dtC = \int_0^1 \dot{z}(t)\nabla G(z(t))^\top dt is therefore a mixed (1,1)(1,1)-tensor transforming by similarity (MCM−1MCM^{-1}) rather than congruence, and is generically non-normal (CC⊤≠C⊤CCC^\top \neq C^\top C) whenever gradients rotate along the path. Because MUΣV⊤M−1M U \Sigma V^\top M^{-1} is not an SVD and 12(C+C⊤)\frac{1}{2}(C+C^\top) mixes dual spaces, standard orthogonal projections (SVD/PCA) violate GL(d)GL(d) equivariance. Formulating the attribution intervention via oblique Riesz contour projectors PΛ(C)P_\Lambda(C) together with non-normal conditioning gates (∥PΛ∥2\lVert P_\Lambda\rVert_2, relative eigengap, and modulus-tie abstention) is therefore mathematically required to achieve equivariance. Empirically, even in the unmodified native basis (M=IM=I), CC has effective rank 9.72–45.2 on our vision and deeper text cuts, and matching a single rank-1 Riesz projection requires ablating 11.0–32.5 top native channels.
  2. Exact impossibility boundary for completeness. For any nonscalar CC at a fixed cut, path, and scalar target, Theorem 2 shows that the GL(d)GL(d) similarity orbit of diag⁡(MCM−1)\operatorname{diag}(MCM^{-1}) is the entire completeness hyperplane {a∈Rd:1⊤a=tr⁡(C)}\lbrace a \in \mathbb{R}^d : \mathbf{1}^\top a = \operatorname{tr}(C)\rbrace (with the scalar exception C=λIC=\lambda I stated explicitly). Thus completeness alone leaves all d−1d-1 transverse degrees of freedom unconstrained unless additional basis-fixing assumptions or off-diagonal operator structure are introduced.

Response to Critical Weakness 2

To evaluate the practical benefit of the dominant-modulus rule (max⁡∣λ∣\max |\lambda|) beyond runner-up and single-seed complement comparisons, we expanded our frozen-checkpoint evaluation across vision development paths, five vision holdout cells, AG News CLS and deeper Layer-0 cuts, and a decoupled Dev-to-Test protocol (full results in Response to Question 1, Table R1):

  1. Five same-rank GL(d)GL(d)-equivariant spectral alternatives: 80.2%–100.0% pooled win rates across runner-up, third-largest, uniform-cluster mean, median non-dominant, and weakest clusters (77.1%–100.0% under a native-basis ℓ2\ell_2-norm-matched runner-up control).
  2. Higher-rank full complement (I−PdomI-P_{\text{dom}}, rank d−rd-r): 80.9% wins on vision dev, 70.8% on pooled vision holdouts (83.3%–87.5% on ResNet-18), 91.7%–100.0% on per-path NLP cuts, and 76.4%–97.2% on decoupled NLP cuts.
  3. Fixed-basis attribution baselines (on the 120 vision holdouts): 95.8%–97.5% wins against native top-rr conductance (97.5%), straight-path hidden IG top-rr (95.8%), and endpoint Grad×Act\text{Grad}\times\text{Act} top-rr (97.5%).

Response to Question 1

Yes. Beyond the runner-up and single-seed complement comparisons in the manuscript, we evaluated the dominant rule (max⁡∣λ∣\max |\lambda|) against six equivariant spectral alternatives and a native-basis perturbation-size control across both per-path and decoupled Dev-to-Test protocols (in Dev-to-Test, each class-conditional Cdev(y)C_{\text{dev}}^{(y)} averages six development paths for y∈{0,1,2,3}y \in \lbrace 0,1,2,3 \rbrace and its frozen projectors are applied to six disjoint test displacements per class without computing CtestC_{\text{test}}):

  1. Runner-up cluster: 2nd-largest modulus among eligible clusters of the dominant channel rank rr.
  2. Third-largest eligible cluster: 3rd-largest modulus in that same-rank ordering.
  3. Uniform-cluster mean: exact uniform mean of individual intervention effects over all eligible rank-rr nonzero clusters (including the dominant cluster), strictly GL(d)GL(d)-equivariant without a Euclidean metric.
  4. Median non-dominant cluster: upper median in the modulus ordering of eligible non-dominant rank-rr clusters.
  5. Weakest eligible cluster: smallest nonzero modulus (>10−8> 10^{-8}) among eligible same-rank clusters (conjugate-closed).
  6. Entire complement (I−PdomI - P_{\text{dom}}): higher-rank (d−rd-r) spectral complement.
  7. Native-basis ℓ2\ell_2-norm-matched runner-up (control): scales the runner-up perturbation to ∥PdomΔz∥2\lVert P_{\text{dom}}\Delta z\rVert_2 in the native Euclidean norm.

Table R1: Spectral alternatives and a separate native-basis norm control (median absolute logit effects and paired win rates)

Setting / ProtocolPathsDominant (max⁡∣λ∣\max\lvert\lambda\rvert)vs. Runner-Upvs. Norm-Matched RU (Native Control)vs. 3rd-Largest Eligiblevs. Uniform-Cluster Meanvs. Median Non-Dominant Clustervs. Weakest Clustervs. Entire Complement (I−PI-P, rank d−rd-r)
Vision Dev Study (Seed 2030)476.2587.2% (1.86)95.7% (1.65)87.2% (1.27)100.0% (0.63)100.0% (0.053)100.0% (0.0044)80.9% (3.33)
Vision Holdouts Pooled (5 cells)1208.3090.8% (2.61)93.3% (3.02)91.7% (1.79)97.5% (1.16)98.3% (0.16)98.3% (0.025)70.8% (6.80)
ResNet-18 CIFAR-10 (Seed 2027)246.7887.5% (1.37)83.3% (1.85)95.8% (0.70)100.0% (0.52)100.0% (0.034)100.0% (0.0088)83.3% (1.86)
WRN-28-2 CIFAR-10 (Seed 2030)249.5495.8% (3.23)100.0% (4.37)75.0% (3.96)91.7% (3.13)91.7% (0.93)91.7% (0.30)50.0% (9.78)
PlainConvNet CIFAR-10 (Seed 2030)2412.6370.8% (4.32)83.3% (4.20)95.8% (1.70)95.8% (1.87)100.0% (0.56)100.0% (0.10)50.0% (13.44)
ResNet-18 CIFAR-100 (Seed 2030)2411.25100.0% (3.28)100.0% (3.86)95.8% (2.86)100.0% (1.29)100.0% (0.15)100.0% (0.016)87.5% (7.21)
ResNet-18 SVHN (Seed 2030)246.08100.0% (1.50)100.0% (2.57)95.8% (1.16)100.0% (0.60)100.0% (0.042)100.0% (0.007)83.3% (4.74)
NLP Holdouts CLS (Seeds 2027–2029)1442.34100.0% (0.0046)100.0% (0.12)100.0% (2.8e-05)100.0% (0.25)100.0% (0.00)100.0% (0.00)100.0% (0.0047)
NLP Holdout Layer-0 Cut (Deep)721.7693.1% (0.41)91.7% (0.25)97.2% (0.22)98.6% (0.11)100.0% (0.00082)100.0% (1.3e-05)91.7% (0.37)
Decoupled Dev-to-Test Vision965.5180.2% (2.60)77.1% (2.44)84.4% (1.28)85.4% (1.75)83.3% (0.35)89.6% (0.28)41.7% (6.71)
Decoupled Dev-to-Test NLP Layer-0721.6490.3% (0.29)90.3% (0.27)87.5% (0.30)97.2% (0.28)95.8% (0.13)98.6% (0.031)76.4% (0.60)
Decoupled Dev-to-Test NLP CLS722.4998.6% (0.034)100.0% (0.16)98.6% (0.026)97.2% (0.19)97.2% (0.021)98.6% (0.0034)97.2% (0.063)

Table R1 also clarifies the natural boundaries of the rule: finite nonlinear effects need not be strictly monotonic across every individual cell (e.g., WRN-28-2's third-largest cluster has median 3.96 vs. 3.23 for the runner-up), and in decoupled vision a single frozen development subspace captures 5.51 logits vs. 6.71 logits for the remaining 63–255 dimensions of the complement (41.7% wins).

Dear Reviewer dspU,

Thank you for your valuable time and constructive feedback. We would like to gently remind you that we have posted our point-by-point responses addressing all your comments and concerns in the discussion forum.

As the discussion period is ongoing, we would greatly appreciate it if you could review our responses at your convenience, and we welcome any further questions or discussions.

Best regards,

Authors

Reviewer_T6tyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks which hidden-neuron attribution claims are determined by the represented network function. It derives a similarity law for the path-integrated conductance matrix under compensated affine changes of hidden coordinates. Combining this with Fillmore's theorem, it shows that any trace-compatible coordinate attribution vector is attainable when the matrix is nonscalar. The authors then construct interventions using isolated spectral clusters and Riesz projectors, proving that these interventions commute with the gauge. Experiments test coordinate reassignment, numerical equivariance, and the target-logit effects of the largest-modulus cluster across vision and text models, with additional adversarial-endpoint tests.

Critical strengths
  1. A precise identifiability result. Theorem 2 characterizes the whole coordinate-attribution orbit, including the scalar exception. This is stronger than demonstrating that a particular ranking changes. For a specified isolated cluster, the Riesz projector also gives a well-defined equivariant object for non-normal matrices and repeated eigenvalues. The paper distinguishes this theoretical coverage from the narrower eigenvector-based numerical implementation.

  2. Several distinct empirical checks. The 27-cell vision matrix and 2,592 gauge trials test numerical equivariance across architectures, datasets, and seeds. The utility holdouts give 109/120 dominant-cluster wins against a matched-rank runner-up, while the fresh text holdout gives 72/72 wins against the entire complement (Sections 4.4–4.5). The manuscript separates these outcomes from exact identity checks and reports the weakest vision cell, mixed signed attack effects, and two attack cells failing the isolation gate. This makes the supported scope reasonably clear.

Critical weaknesses
  1. The utility benchmark leaves the choice of intervention target insufficiently justified. The dominant cluster is constructed from the same target-specific conductance matrix used to motivate the subsequent same-target intervention. This is a legitimate test of the frozen rule: largest spectral modulus does not guarantee the largest finite nonlinear endpoint effect. Nevertheless, beating one runner-up selected from the same matrix leaves open whether the dominant rule is preferable to other admissible intervention choices, or useful when projector construction and evaluation use different information. The next-class control is weaker, with one holdout at 50% wins (Section 4.5). The text complement result is useful additional evidence, but does not resolve cross-target generalization. The authors correctly disclaim semantic and causal interpretation; my concern is the limited positive case for choosing this particular target in practice.

  2. The practical significance of the full gauge needs a more concrete demonstration. Section 3.2 explicitly allows an inserted affine interface that need not remain within the original architecture, and the paper repeatedly preserves parameterization-relative neuron analyses. I agree with that distinction. What remains unclear is where a practitioner encounters equivalent implementations in different dense gauges and would draw a consequentially different conclusion from coordinate conductance. The relocation experiment demonstrates the mathematical possibility, but does not establish such a use case. The constructive method is also evaluated on modest vision models and a small text classifier with a 128-dimensional cut. The complete-orbit theorem is a worthwhile formal contribution, but the practical reach of its proposed replacement remains limited by these evaluations.

Overall assessment: The theory is precise and the reported experiments establish numerical equivariance and conditional output relevance. My reservation is significance: the paper does not yet demonstrate a compelling practical setting for the full gauge or establish why the dominant spectral rule should be preferred beyond the current target-conditioned comparisons. The text replication strengthens the evidence, but the application case remains insufficient for acceptance.

Questions for authors
  1. Could you separate projector construction from evaluation, for example by selecting a projector from an aggregate conductance matrix on development paths at a fixed cut and target, then evaluating fresh path displacements or another target? The existing holdouts freeze the rule but still construct a new target-conditioned matrix for each test path. An additional rank-controlled comparator that respects the same full GL(d) equivariance would help assess the selection rule itself.

  2. Can you give a concrete application where dense gauge equivalence matters to an attribution decision, while acknowledging that neuron analysis within a fixed trained implementation remains legitimate? A worked example involving equivalent checkpoints or an explicit representation interface would make the motivation more persuasive.

  3. Which downstream cuts are used for the vision utility comparisons, and how nonlinear are their downstream maps? Appendix A.2.3 describes the special rank-one linear-head boundary. Please also report matrix-construction and projector-computation costs, and discuss whether the isolation diagnostics and any low-rank approximation remain usable at substantially larger hidden widths.

6 comments

Thank you for your detailed review and constructive questions. We appreciate your positive assessment of Theorem 2's coordinate-orbit characterization, the Riesz-projector formulation for non-normal matrices, and the empirical protocol. Below we respond to each weakness and question in order.


Response to Critical Weakness 1

To test the dominant spectral rule beyond single-path, same-target runner-up comparisons, we ran three additional evaluations on our frozen checkpoints (full protocol and results in Response to Question 1, Table R1):

  • Decoupled construction and evaluation (Dev-to-Test): Freezing spectral projectors from class-conditional development averages Cdev(y)C_{\text{dev}}^{(y)} and applying them directly to unseen test displacements Δztest\Delta z_{\text{test}} (without computing CtestC_{\text{test}}) yields 80.2%–98.6% same-target win rates across vision and text cuts.
  • Broader equivariant and norm-matched comparators: Across per-path and decoupled protocols, the dominant rule wins 80.2%–100.0% against five same-rank GL(d)GL(d)-equivariant alternatives, beats the full (d−r)(d-r)-dimensional complement I−PdomI-P_{\text{dom}} on 70.8% (vision) and 91.7% (deep NLP), and maintains 77.1%–100.0% wins under a native-basis ℓ2\ell_2-norm-matched runner-up control.
  • Decision-level functional checks: Evaluating GyG_y-constructed projectors on the endpoint's highest competing-class logit, decision margin, cross-entropy loss, and predictive KL divergence shows consistent transfer on ResNet-18, WRN-28-2, and AG News, while exposing weaker decoupled transfer on PlainConvNet (41.7%–58.3%).

Response to Critical Weakness 2

We address the practical role of dense GL(d)GL(d) transformations and width scalability through three results (detailed in Responses to Questions 2 and 3, Tables R2–R3):

  • Cross-checkpoint attribution transfer (Mstitch∈GL(d)M_{\text{stitch}} \in GL(d)): Linear stitching maps between independently trained seeds (Seed 2026→20272026 \to 2027) are dense (74.4%–93.7% non-peak row mass). Transporting source spectral projectors (SS) vs. source native top-rr projectors (AA) through the same map MstitchM_{\text{stitch}} yields 95.8%–97.2% raw transfer wins, while norm-scaled controls (A-up, S-down) show that transfer gains reflect directional alignment on ResNet-18 and AG News (75.0%–100.0% S-down wins) but mainly larger multi-channel perturbation norms on WRN-28-2 and PlainConvNet (33.3%–41.7% S-down wins).
  • Multi-channel coalitions in the unmodified native basis (M=IM=I): Even within a single fixed checkpoint, the leading spectral direction spans PR∈[19.9,45.5]\mathrm{PR} \in [19.9, 45.5] channels, and matching a single rank-1 spectral intervention requires ablating 11.0–32.5 top native channels.
  • Scalability up to d=4096d=4096: On synthetic rank-64 factorized operators (C=UV⊤C=UV^\top), dual Gram reduction and matrix-free Arnoldi compute the d=4096d=4096 Riesz projection in 4.82 ms and 22.43 ms (vs. 34,678.2 ms for dense eig), with spatially contracted cuts (S=L⋅T≥dS = L\cdot T \ge d) using the matrix-free Arnoldi path.

Response to Question 1

For each class y∈{0,1,2,3}y \in \lbrace 0,1,2,3 \rbrace, averaging the six development conductance matrices gives Cdev(y)C_{\text{dev}}^{(y)}, which transforms under M∈GL(d)M \in GL(d) as MCdev(y)M−1M C_{\text{dev}}^{(y)} M^{-1} by linearity; thus any isolated cluster of Cdev(y)C_{\text{dev}}^{(y)} defines a GL(d)GL(d)-equivariant frozen projector (max pullback error ≤1.49×10−13\le 1.49\times 10^{-13}). We apply these frozen projectors directly to six disjoint test displacements per class without forming CtestC_{\text{test}} (96 vision test paths across four cells; 72 text paths across three seeds at both CLS and Layer-0).

Table R1 compares the dominant rule against five same-rank GL(d)GL(d)-equivariant rules (runner-up, 3rd-largest, uniform-cluster mean over all eligible rank-rr nonzero clusters including dominant, median non-dominant, and weakest nonzero cluster with ∣λ∣>10−8|\lambda| > 10^{-8} and conjugate pairing), a separate native-basis ℓ2\ell_2-norm-matched runner-up control, and four decision functionals (highest competing-class logit, margin, cross-entropy, and KL divergence):

Table R1: Per-path and decoupled (Dev-to-Test) evaluation across spectral comparators, native-norm controls, and functional checks

Protocol & SettingEval PathsDom. Logit EffectWin vs. Runner-Up (Raw / Native-Norm Control)Win vs. 3rd-LargestWin vs. Uniform-Cluster MeanWin vs. Median ClusterWin vs. Weakest ClusterWin vs. RU (Competing Logit / Margin / CE / KL)
Per-Path Vision Pooled (5 cells)1208.3090.8% / 93.3%91.7%97.5%98.3%98.3%65.0% / 90.0% / 90.0% / 90.0%
Per-Path NLP Layer-0 Cut (Deep)721.7693.1% / 91.7%97.2%98.6%100.0%100.0%62.5% / 91.7% / 93.1% / 94.4%
Decoupled Dev-to-Test Vision Pooled965.5180.2% / 77.1%84.4%85.4%83.3%89.6%72.9% / 75.0% / 74.0% / 75.0%
ResNet-18 CIFAR-10 (Dev-to-Test)247.7691.7% / 83.3%100.0%100.0%100.0%100.0%87.5% / 87.5% / 91.7% / 91.7%
WRN-28-2 CIFAR-10 (Dev-to-Test)244.1975.0% / 62.5%75.0%70.8%70.8%70.8%83.3% / 62.5% / 58.3% / 66.7%
PlainConvNet CIFAR-10 (Dev-to-Test)243.3454.2% / 62.5%62.5%70.8%62.5%87.5%58.3% / 50.0% / 45.8% / 41.7%
ResNet-18 SVHN (Dev-to-Test)245.94100.0% / 100.0%100.0%100.0%100.0%100.0%62.5% / 100.0% / 100.0% / 100.0%
Decoupled Dev-to-Test NLP Layer-0721.6490.3% / 90.3%87.5%97.2%95.8%98.6%77.8% / 87.5% / 93.1% / 91.7%
Decoupled Dev-to-Test NLP CLS722.4998.6% / 100.0%98.6%97.2%97.2%98.6%63.9% / 95.8% / 98.6% / 94.4%

95% Wilson lower bounds for decoupled same-target wins are 0.711 (vision), 0.813 (NLP Layer-0), and 0.925 (NLP CLS). Against the full (d−r)(d-r)-dimensional complement I−PdomI-P_{\text{dom}}, the dominant projector wins 70.8% (per-path vision), 91.7% (per-path NLP Layer-0), 76.4% (decoupled NLP Layer-0), and 41.7% (decoupled vision, where a single frozen development subspace is compared against all remaining d−rd-r dimensions simultaneously). Table R1 also shows where transfer weakens: PlainConvNet achieves 54.2% raw (62.5% norm-matched) same-target wins and 41.7%–58.3% cross-functional wins in Dev-to-Test, and competing-logit transfer (63.9%–77.8% decoupled) is lower than same-target intervention.


Response to Question 2

Table R2 evaluates cross-checkpoint attribution transfer (Seed 2026→20272026 \to 2027) and native-basis (M=IM=I) coalitions across all four architectures:

  1. Cross-checkpoint transfer (SS vs. AA). We fit z(2027)≈Mstitchz(2026)+cz^{(2027)} \approx M_{\text{stitch}} z^{(2026)} + c via Procrustes-regularized regression on held-out development activations (24 vision paths per model; 36 CLS paths). While comparing coordinates before vs. after MstitchM_{\text{stitch}} inside the stitched network Fst=G(2027)(MstitchH(2026)+c)F_{\text{st}} = G^{(2027)}(M_{\text{stitch}} H^{(2026)} + c) is a direct numerical check of our theorem (100.0% top-1 flip vs. 0.00 spectral discrepancy), the practical test is transferring source explanations between the two original checkpoints: we transport both the source spectral projector (S=MstitchP(2026)Mstitch−1S = M_{\text{stitch}} P^{(2026)} M_{\text{stitch}}^{-1}) and the source native top-rr projector (A=MstitchQnative(2026)Mstitch−1A = M_{\text{stitch}} Q_{\text{native}}^{(2026)} M_{\text{stitch}}^{-1}) through the same map to intervene on Seed 2027, reporting raw effects alongside AA scaled up to SS's Euclidean norm (A-up) and SS scaled down to AA's single-channel norm (S-down).
  2. Native-basis (M=IM=I) coalition audit. Because any dense non-axis-aligned operator in dimension dd naturally has high off-diagonal Frobenius mass (98.4%–99.6%), the stronger test at M=IM=I is how many native channels must be ablated together to match a rank-1 spectral intervention (scanned over k=1,…,dk=1,\dots,d on 16 paths per vision model and 24 Layer-0 paths).

Table R2: Cross-checkpoint linear stitching (Mstitch∈GL(d)M_{\text{stitch}} \in GL(d), Seed 2026→20272026 \to 2027) and native-basis (M=IM=I) coalition audit

Architecture & CutDim ddκ(Mstitch)\kappa(M_{\text{stitch}})Median Non-Peak Row MassStitched Test Acc. (Agreement)Interface Top-1 Flip / Top-5 JaccardSpectral Effect DiscrepancyTransfer SS / AA (Raw; A-up; S-down Win %)Native Off-Diagonal Ratio & PRFirst Matching Native kk (Matched Paths)
ResNet-18 (layer3)2564.4884.0%92.0% (93.2%)100.0% / 0.0000.007.16 / 0.22 (95.8%; 91.7%; 75.0%)99.6% (PR: 27.4 ch.)32.5 ch. (16/16 matched)
WRN-28-2 (layer2)643.2179.3%90.8% (92.2%)100.0% / 0.0000.006.58 / 1.00 (95.8%; 66.7%; 41.7%)98.4% (PR: 19.9 ch.)14.0 ch. (16/16 matched)
PlainConvNet (features[:14])1283.9074.4%86.8% (89.4%)100.0% / 0.0000.009.84 / 3.16 (95.8%; 50.0%; 33.3%)99.5% (PR: 43.8 ch.)11.0 ch. (11/16 matched)
AG News (CLS stitch / L0 audit)1288.0493.7%76.4% (75.6%)100.0% / 0.0000.002.00 / 0.057 (97.2%; 100.0%; 100.0%)99.5% (PR: 45.5 ch.)31.0 ch. (17/24 matched)

Raw transfer SS wins 95.8%–97.2% across all four architectures, and norm-inflated A-up retains 66.7%–100.0% wins on ResNet-18, WRN-28-2, and AG News (50.0% on PlainConvNet). Scaling SS down to the single-channel norm (S-down) retains 75.0% wins on ResNet-18 and 100.0% on AG News, but drops to 41.7% on WRN-28-2 and 33.3% on PlainConvNet, showing that cross-checkpoint transfer reflects directional alignment on ResNet-18 and AG News, whereas on WRN-28-2 and PlainConvNet it is largely driven by the larger multi-channel displacement norm captured by SS.


Comments on the paper

2 comments

Dear Reviewers,

Thank you for your valuable time and constructive feedback. We would like to gently remind you that we have posted our point-by-point responses addressing all your comments and concerns in the discussion forum.

As the discussion period is ongoing, we would greatly appreciate it if you could review our responses at your convenience, and we welcome any further questions or discussions.

Best regards,

Authors

Dear ACs and SACs,

As the discussion period closes, we briefly summarize the review status and what our rebuttal resolved.

All three reviewers (T6ty, dspU, ojhE) recognized our exact coordinate-orbit characterization (Theorem 2), Riesz-projector formulation, and careful experimental protocol, raising no major soundness flaws (dspU: Soundness 3/3, Presentation 3/3; ojhE: Presentation 3/3). Their initial Weak rejection scores focused on practical scope rather than errors, and our rebuttal experiments resolved all shared concerns:

  1. Broader comparators and unseen test transfer (T6ty, dspU): Our dominant subspace outperforms five equivariant rules, the full complement, norm-matched controls, and standard baselines (native conductance, IG, Grad ×\times Act). More importantly, subspaces built only on dev inputs transfer directly to unseen test inputs with 80%–99% win rates across four metrics.
  2. Not a baseline reset (ojhE): While we confirmed that the shallow 1-token CLS cut behaves near a reset, across all 192 paths on vision cuts and a deeper text cut, the remaining representation keeps almost its full distance from baseline (median ratio ≈1.0\approx 1.0) while capturing most of the predictive effect.
  3. Practical need for GL(d)GL(d) equivariance and scalability (T6ty, dspU, ojhE): Under fixed cross-seed alignment (stitching / CCA), non-GL(d)GL(d) rules (neuron coalitions and orthogonal/SVD subspaces) flip necessity/sufficiency conclusions on up to 36%–83% of inputs across coordinate frames, whereas Riesz subspaces give 0% flips. We also clarified why dual velocity-gradient scaling breaks SVD/PCA equivariance, and cut width-4096 runtime from 34.7s to 5–22ms.

Discussion status and manuscript: During the discussion, Reviewer ojhE raised their score after our rebuttal and follow-up experiments addressed their concerns, and Reviewer dspU also responded positively and affirmed our additional experiments (while keeping the initial rating). All clarifications and new results are already integrated into the revised 10-page paper and Appendix.

Thank you very much for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.