acceptodds
Under review as a conference paper at ICLR 2027

When More Capacity Is Not Enough: Rethinking Potential-Based Attention

Abstract

Potential-based attention models organize computation through a scalar function, providing mathematical structure for analyzing model behavior. A natural goal is to use this structure to reproduce an existing attention model, preserving its outputs on the same inputs while making its behavior easier to analyze. Contrary to the intuition that scaling model capacity can overcome such limitations, we find that an unavoidable output error can persist even with arbitrarily complex potentials. Generating multiple outputs through the gradient of a single convex function requires them to obey shared rules of variation, which the original attention response may not satisfy. For a class of fixed attention maps, we first prove a capacity-independent error lower bound that quantifies this structural restriction. We then give a criterion computable directly from the original model that determines the minimum number of independently corrected output directions needed to recover its outputs exactly. We also find that fewer corrections can sometimes approximate the original outputs arbitrarily well, but only as the relative scaling of output directions becomes increasingly uneven. This reveals the numerical cost of reducing the required structural change. Finally, numerical experiments with rigorous error guarantees establish that independent corrections can cross an accu- racy limit that increasing potential capacity alone cannot overcome. These results provide a theoretical basis for deciding when the architecture must change instead of continuing to scale the model. We believe these insights can guide attention designs that combine mathematical structure with expressive power.

Primary area:
learning theory
Keywords:
Potential-Based Attention, Representation Learning, Attention Mechanisms

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_8g84OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the expressive limitations of potential-based attention models. The authors show that increasing the complexity of a convex potential cannot always eliminate approximation errors due to structural constraints imposed by gradient representations.

The main contribution is a characterization of the minimum number of independent nonlinear correction directions required for exact representation and arbitrarily accurate approximation, based on the rank of a residual matrix determined by the attention keys and values. The authors also establish quantitative error bounds and show that reducing the correction dimension may require ill-conditioned output transformations.

Numerical experiments, including certified error bounds and finite-model comparisons, support the theoretical results. The experiments also demonstrate that fewer correction directions do not necessarily reduce computational costs.

Critical strengths
  1. Rigorous theoretical characterization. The paper provides a mathematical characterization of the expressive limitations of potential-based attention models. In particular, the authors establish quantitative error bounds and identify the minimum number of nonlinear correction directions required for exact representation and arbitrarily accurate approximation. These results clarify the structural limitations that cannot be overcome simply by increasing model capacity.

  2. Connection between theoretical results and numerical validation. The paper supports its theoretical findings with numerical experiments, including certified approximation error bounds. The analysis of output conditioning and computational costs further clarifies the practical implications of the theoretical results. This combination of mathematical analysis and numerical validation strengthens the contribution of the paper.

Critical weaknesses

Weakness 1: Restrictive assumptions on the attention model and input distribution.

The theoretical results rely on several restrictive assumptions, including fixed attention keys and values, genericity conditions on the keys, and a uniform probability measure on a compact domain. It is unclear to what extent the results extend to more general attention models or realistic input distributions. These restrictions limit the applicability of the theoretical findings to practical attention architectures.

Weakness 2: Limited experimental validation on practical attention models.

Although the paper provides rigorous theoretical results and certified numerical experiments, the empirical evaluation relies primarily on synthetic examples and a relatively simple regression task. Experiments involving larger attention models or more realistic learning tasks would help establish the practical significance of the proposed framework.

Questions for authors

Question 1: Explicit characterization of the lower-bound constant.

In Theorem 1, the constant c_k=c(K,\eta,D,r-k)>0 plays an important role in quantifying the approximation error. Although the proof provides a construction of this constant, could the authors give a more explicit or computable lower bound in terms of the key geometry, attention bias, and domain D? In particular, it would be helpful to clarify how the constant behaves when the key configuration approaches degeneracy.

Question 2: Sharpness of the conditioning-dependent error bound.

Theorem 2 establishes a lower bound involving the factor \chi^{-2}, and the paper demonstrates its sharpness for a particular rank-two example. Is this dependence optimal for more general attention configurations? In particular, under what additional assumptions can a matching upper bound be established?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 8g84

We sincerely appreciate your insightful comments and your recognition of the rigorous representation results and their numerical validation. We have added explicit quantitative answers, clarified the assumptions, and evaluated the framework in pretrained models.

1. Fixed banks, genericity, and input distributions (W1)

The revision treats each assumption separately. A nonuniform mean law ν\nu with locally positive continuous density admits an interior-ball uniform law μ0\mu_0 and ν≥αμ0\nu\ge\alpha\mu_0. Hence

Ek,χ∗,ν≥αc~k,0(τk2+χ−2σk+1(R)2),τk2=∑j=k+2rσj(R)2.\mathcal E_{k,\chi}^{*,\nu}\ge\alpha\widetilde c_{k,0}\left(\tau_k^2+\chi^{-2}\sigma_{k+1}(R)^2\right),\qquad \tau_k^2=\sum_{j=k+2}^{r}\sigma_j(R)^2.

The upper construction is pointwise and applies to every probability law on the compact domain. For context-dependent keys and values, the lower bound applies conditionally and is then averaged over contexts, including for shared-parameter readouts, when the conditional banks and query laws satisfy the assumptions.

The upper construction requires affine spanning alone. Matching lower bounds use the stronger requirements that every (d+1)(d+1)-key subset is affinely independent and distinct such subsets have distinct vector sums. Appendix Q.3 tests affine spanning; Appendix O.2–O.3 specifies the assumptions for each extension.

2. Practical model validation (W2)

We evaluate Qwen2.5-0.5B and 1.5B on WikiText-2 and Penn Treebank, replacing prefix attention in interleaved layers and recomputing subsequent network states. Final prediction preservation is:

ModelCorpusPPL change vs. original (%)KL reduction vs. fitted correction (%)Top-prediction agreement (%)
Qwen2.5-1.5BWikiText-2+0.0696.699.5
Qwen2.5-1.5BPenn Treebank+0.0195.099.4
Qwen2.5-0.5BWikiText-2+0.0697.899.3
Qwen2.5-0.5BPenn Treebank−0.0296.499.1

The learned-bank comparison also separates capacity from direction choice. Entries below are median error reductions relative to an affine readout; all corrected models use the same correction dimension.

Potential widthNo correction (%)Random directions (%)Value-residual directions (%)
160.00.51.0
640.02.24.2
2560.26.612.3

Table 2 and Table 4 contain these results. Appendix Q gives direction controls and paired uncertainty; Appendix R reports computation and numerical sensitivity.

3. Explicit constant and behavior near degeneracy (Q1)

For n>d+1n>d+1, an explicit choice is

c∗=λmin⁡(G)2λmin⁡(S)2Cψ∥Hinc∥op2,ck=c∗,c~k=c∗r−k−12(r−k).c_* = \frac{\lambda_{\min}(G)^2\lambda_{\min}(S)}{2C_\psi\lVert H_{\mathrm{inc}}\rVert_{\mathrm{op}}^2},\qquad c_k=c_*,\qquad \widetilde c_k=c_*\frac{r-k-1}{2(r-k)}.

Here GG is the Gram matrix of simplex weight functions, SS sums outer products of barycentric gradients, CψC_\psi sums test-function gradient energies, and HincH_{\mathrm{inc}} records simplex vertex replacements. All depend only on keys, bias, domain, and the chosen tests. Appendix O.1 gives their definitions and query-coordinate integrals for evaluating the constant independently of the values.

For compact generic key families with bounded biases and uniformly controlled interior test support, continuity gives a uniform positive bound. The formula identifies geometric factors that can weaken the certificate near degeneracy. Figure 2 holds the residual spectrum fixed while changing key-and-domain geometry and measures the resulting certificate.

4. Optimal conditioning dependence beyond the rank-two example (Q2)

The revision establishes matching bounds for general residual spectra. Under the lower-bound assumptions, for 0≤k≤r−20\le k\le r-2,

c~k(τk2+χ−2σk+1(R)2)≤Ek,χ∗≤2τk2+2ρ2Mp2χ−2σk+1(R)2.\widetilde c_k\left(\tau_k^2+\chi^{-2}\sigma_{k+1}(R)^2\right)\le\mathcal E_{k,\chi}^*\le2\tau_k^2+2\rho^2M_p^2\chi^{-2}\sigma_{k+1}(R)^2.

Here ρ\rho is the domain radius and MpM_p bounds the derivative of attention probabilities in mean coordinates. The spectral-tail and conditioning terms match up to constants fixed by geometry, bias, and domain. In particular,

rank⁡(R)=k+1<r⟹Ek,χ∗=Θ(χ−2).\operatorname{rank}(R)=k+1<r\quad\Longrightarrow\quad\mathcal E_{k,\chi}^*=\Theta(\chi^{-2}).

This proves optimal inverse-square decay throughout this class. The upper construction itself requires only affine spanning. Theorem 2 states the bounds, and Appendix O.2 gives the construction and proof.

The common response also describes the theoretical advances and new experiments. We hope these additions address the central weaknesses you identified and respectfully ask whether you would consider revisiting your rating in light of the revised evidence. Thank you again for your careful review; we welcome further discussion to improve the final manuscript!

Thank you for the clarification and for stating the upper bound in Theorem 2. It remains unclear to me whether the dependence on MpM_p is necessary. Could you clarify whether the upper estimate can be obtained without the MpM_p factor under the same assumptions?

Thank you for this question. Your feedback has helped us recognize an important blind spot in our thinking and make our contributions more solid. Yes, under the same assumptions, the upper estimate can be obtained without an explicit MpM_p factor. We have strengthened Theorem 2 to

Ek,χ⋆≤τk2+ρ2GK2d+2χ−2σk+1(R)2\mathcal E_{k,\chi}^{\star}\le \tau_k^2+\frac{\rho^2G_K^2}{d+2}\chi^{-2}\sigma_{k+1}(R)^2,

where GKG_K depends only on the fixed keys. This bound is uniform over all finite attention biases.

The key observation is that the derivative of the attention weights in mean coordinates is a convex combination of barycentric-coordinate derivatives of key simplices. This provides a bound determined entirely by the keys, without separately estimating MpM_p. The complete proof is now included in Appendix O.2. This improvement allows us to refine the theoretical proof without affecting the rest of the paper.

Reviewer_CPYiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the minimum number of nonlinear correction directions needed to approximate or exactly reproduce standard attention using a potential-based readout theoretically. It constructs a readout with affine terms, a full-row-rank transformation of a convex-potential gradient, and a low-dimensional nonlinear correction term. It uses mainly synthetic attention maps to verify that potential-only readouts can have irreducible approximation error, while nonlinear corrections can overcome this limitation, though they do not necessarily reduce computational cost.

Critical strengths
  • The paper theoretically characterizes how many nonlinear correction directions are needed to approximate or exactly reproduce standard attention, based on the rank of the value residual. This insight provides a useful criterion for determining whether a compact potential-based readout can represent a given attention operation.
  • The paper shows that potential-based readouts can have an irreducible approximation error regardless of model capacity, and that reducing the number of correction directions can require increasingly ill-conditioned output transformations. This helps distinguish structural limitations from insufficient model capacity.
Critical weaknesses
  • The motivation and potential for this potential-based model is not sufficiently highlighted. For example, if this potential-based model can reproduce an attention model losslessly, how can it benefit model training or inference in an application perspective? Even if it can't fully reproduce an attention model, can a hybrid model interleaving potential-based and typical attention preserve model capability while achieving some sort of efficiency our stability benefit?
  • The proposed nonlinear correction branch is relatively unrestricted, without explicitly exploiting the structural constraints imposed by the potential gradient. It would be interesting to investigate more structured correction mechanisms that specifically address violations of the Jacobian compatibility condition mentioned in the paper, rather than relying on a general smooth nonlinear function.
  • The experiments are conducted on synthetic attention maps, with limited evaluation on learned models. Their relevance to real attention mechanisms remains unclear. In addition, whether attention keys in pretrained Transformers satisfy the affine-spanning assumption is not empirically examined.
Questions for authors
  • Is there a more structured way to design the correction branch that specifically addresses the incompatibility left by the potential-based term?
  • Could the proposed method also be evaluated using real LLM attention keys and values rather than synthetic constructions? Could the authors empirically verify whether attention keys in pretrained Transformers affinely span the key space?
  • Could the authors investigate head-wise or layer-wise variability in the observed value residuals of pretrained Transformers? For example, measuring the effective ranks of the residual matrices across attention heads and layers could reveal whether if the potential-based readouts work in a fine-grained manner in a transformer model.
7 comments
AuthorsOpenReview ↗

Response to Reviewer CPYi ( Part 1 / 2 )

Thank you for your detailed and constructive feedback. We appreciate your recognition that our insight provides a useful criterion for assessing potential-based readouts. We have strengthened the connection between this criterion, explicit corrections, and pretrained-model behavior.

1. Application value, hybrid capability, and computational cost (W1)

Our criterion makes the architectural decision explicit: can a potential-gradient readout preserve the required attention response, and how many output directions need independent nonlinear computation? If a certified error floor exceeds the desired tolerance, increasing potential capacity cannot attain that accuracy. The residual criterion determines the minimum correction dimension for exact recovery; calibration response errors guide direction selection at finite accuracy. This quantifies the representation cost of adopting a potential-based structure.

We implement this choice by replacing attention to a fixed 256-token prefix in six or seven interleaved layers of Qwen2.5. Later hidden states, queries, and suffix keys and values are recomputed through the modified network. Prediction changes are measured against the original model:

ModelCorpusPPL change (%)Top-prediction agreement (%)
Qwen2.5-1.5BWikiText-2+0.0699.5
Qwen2.5-1.5BPenn Treebank+0.0199.4
Qwen2.5-0.5BWikiText-2+0.0699.3
Qwen2.5-0.5BPenn Treebank−0.0299.1

Forward-time ratios are paired medians relative to the original model. CPU calibration is measured separately per prefix; GPU memory is the additional peak allocation.

ModelCorpusForward-time ratioCPU calibration (s)Extra GPU memory (MiB)
Qwen2.5-1.5BWikiText-21.000.862.65
Qwen2.5-1.5BPenn Treebank1.010.852.65
Qwen2.5-0.5BWikiText-21.010.351.15
Qwen2.5-0.5BPenn Treebank0.950.361.15

Calibration perturbation gains remain below one, with a maximum of 0.39 for direct response corrections.

This demonstrates prediction-preserving hybrid replacement at comparable forward cost and quantifies implementation sensitivity. The convex branch retains mathematical structure; the complete hybrid's stability depends on its full response. The revised PDF includes the construction in Sec. 3.6, prediction results in Table 4, and cost and sensitivity measurements in Appendix R.

2. Structured correction and Jacobian compatibility (W2, Q1)

The correction is explicit in the original attention map. After affine value regression, the remaining nonlinear response and its Jacobian are

c(s)=R⊤p~(s),Dc(s)=R⊤diag⁡(p~)(K−1s⊤)C(s)−1.c(s)=R^\top\widetilde p(s),\qquad Dc(s)=R^\top\operatorname{diag}(\widetilde p)(K-\mathbf1s^\top)C(s)^{-1}.

This derivative contains the nonaffine variation identified by the compatibility tests. For a fitted readout gg, let e=a−ge=a-g. An orthogonal output projector gives

a^k=g+Pke,∥a−a^k∥2=∥e∥2−∥Pke∥2.\widehat a_k=g+P_ke,\qquad \lVert a-\widehat a_k\rVert^2=\lVert e\rVert^2-\lVert P_ke\rVert^2.

Thus the selected response error is removed exactly at every input. Its derivative contribution is PkDeP_kDe. In the full-dimensional implementation with T=IT=I, the remaining error satisfies

D[(I−Pk)e]−D[(I−Pk)e]⊤=(I−Pk)De−De⊤(I−Pk)⊤.D[(I-P_k)e]-D[(I-P_k)e]^\top=(I-P_k)De-De^\top(I-P_k)^\top.

The full residual removes the defect completely. At smaller dimensions, calibration PCA minimizes response error, and the derivative formula specifies the remaining components. Evaluation uses current-query probabilities and cached projected values to compute the selected response components. Sec. 3.6 and Appendix P give the general-dimensional construction, weighted identity, and implementation.

AuthorsOpenReview ↗

Response to Reviewer CPYi ( Part 2 / 2 )

3. Real attention keys and values, and affine spanning (W3, Q2)

The revised evaluation uses learned Qwen2.5 banks and complete-model predictions. The following results include propagation through the modified network:

ModelCorpusKL reduction vs. fitted correction (%)PPL change vs. original (%)
Qwen2.5-1.5BWikiText-296.6+0.06
Qwen2.5-1.5BPenn Treebank95.0+0.01
Qwen2.5-0.5BWikiText-297.8+0.06
Qwen2.5-0.5BPenn Treebank96.4−0.02

All evaluated windows improve in KL. Paired confidence intervals for mean KL and negative-log-likelihood differences lie below zero. For affine spanning, the longer sampled banks give:

Prefix lengthKey-space dimensionBanks with full affine dimension
256 keys128100%
512 keys128100%

This result is consistent across the tested singular-value thresholds and covers every layer and KV head sampled from Qwen2.5-1.5B. Secs. 4.3 and 4.5 and Appendix Q give the protocol and results.

4. Variation across heads and layers, and its use in correction (Q3)

The observed residual spectra vary substantially across heads and layers:

Diagnostic or controlled comparisonResult
Effective rank across head/layer means, 256 keys26–64
Median residual energy / centered value energy, 256 keys16%
Median residual energy / centered value energy, 512 keys29%
Response-error vs. value-residual directions, same direct correction and budget31–62% lower predictive KL

Even query heads sharing a KV bank have different spectra after output projection. Effective rank describes spectral concentration; the exact representation theorem uses algebraic rank. Direction choice improves prediction preservation in every evaluated model/corpus setting. Group-budget allocation helps the 0.5B model relative to randomly permuting the same budgets, while the 1.5B model favors uniform response budgets. Appendix Q.2–Q.3 gives these comparisons and the observation protocol.

The common response also presents the strengthened theory and experimental findings. We hope these additions address your concerns and would appreciate any remaining feedback. In light of the revised evidence and explanations, we respectfully ask whether you would consider revisiting your rating. Thank you again for your time and careful assessment.

Reviewer_CPYiOpenReview ↗

Response to the authors

Thank you for the substantial additional experiments and clarifications. I consider my original concerns regarding the Jacobian compatibility, empirical validation of affine spanning, and head-/layer-wise variability addressed. I also appreciate the additional analysis of correction directions across different attention heads and layers.

That said, I still have follow-up questions regarding (1) practical benefits and (2) the revised hybrid readout scheme.

(1) Practical Benefits

While the revised experiments provide evidence addressing "whether the approach can preserve model capability", "why potential-based architecture is computationally or practically better in the first place" remains unanswered. In my view, this missing piece is critical to justify the contribution of the paper.

Ultimately, the practical value depends on whether incorporating a potential-based readout brings measurable benefits. In the current experiments, the hybrid approach largely preserves model predictions but provides little to no inference speedup. The actual memory/bandwidth savings or improvements in model training dynamics also remain unclear.

It would be nice to show a Pareto curve by sweeping the ratio of inserted potential-based modules, with an accuracy metric on the y-axis and an efficiency metric (e.g., latency, memory usage, or any metric that this method meaningfully improves) on the x-axis. This would make the practical benefits and tradeoffs of the proposed method much clearer.

(2) The revised hybrid readout scheme

We find the proposed hybrid readout, ak(s)=g(s)+Pk(a(s)−g(s))a_k(s)=g(s)+P_k(a(s)-g(s)), interesting, as it combines a potential-based approximation with a low-rank correction from the original attention response.

However, this is fundamentally different from the original potential-only readout, y^(s)=Ls+b+T∇ϕ(s)\hat y(s)=Ls+b+T\nabla\phi(s), which does not explicitly use the original value bank at inference. Therefore, conclusions about the representational capability and limitations of these two formulations should not be conflated. Particularly, the inference-time behavior differs significantly regarding the role of the value bank.

The newly proposed hybrid approach, to me, is more naturally viewed as a potential-assisted value-cache compression technique, which might be useful for fixed-prefix caching if the practical benefits discussed in (1) can be demonstrated.

Reviewer_N2doOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Important Note to AC, and SAC: I have absolutely no knowledge in the subject of the paper. I tried to understand the paper with AI tools, but my evaluation is an educated guess at best.

To the authors: I apologize for not being able to provide a helpful review. I asked to be unassigned from the paper but it has not been completed yet.

Summary

The paper shows that potential-based attention models have fundamental expressivity limitations that cannot be overcome simply by increasing network capacity. These limitations arise from the mathematical structure of generating multiple outputs through the gradient of a single scalar potential.

The authors establish a theoretical characterization of these limitations, determine the minimum number of additional nonlinear output directions needed to overcome them, and quantify the tradeoff between expressive power and numerical conditioning.

Critical strengths

Theorems 1 and 2 seem to establish some important results. Theorem 1 characterizes the minimum number of independent nonlinear correction directions needed to overcome the representational limitations of potential-based attention, whereas Theorem 2 shows that reducing the number of nonlinear correction directions comes at the cost of numerical ill-conditioning. I am certainly ignorant of the literature, but based on my reading of the paper, I believe these results are novel and technically significant.

Critical weaknesses

Not able to identify weaknesses due to my lack of expertise.

Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer N2do

Thank you for your positive feedback and for recognizing the representation and conditioning results. Your perspective helps us improve how the paper communicates its contribution to the broader ML community.

All outputs generated from one potential must satisfy shared derivative constraints. Our results identify when a learned attention response violates these constraints and determine how many output directions need independent computation to recover it. The revision adds an explicit correction construction and pretrained-model validation, with tables organized around the modeling decisions they inform.

The common response presents the strengthened theory and new experiments, which are incorporated into the revised PDF. We would particularly welcome your perspective on whether the revision makes the contribution easier to understand. Thank you again for your time and consideration.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Common Response

We sincerely thank the reviewers for their thoughtful and constructive feedback and for recognizing the paper's theoretical contributions. We also thank the Area Chair for organizing the review and discussion. We appreciate the time and care devoted to our submission.

We have carefully considered every comment and incorporated additional theory, experiments, and discussion into the revised manuscript. Changed prose and table captions are highlighted in yellow. The paper characterizes the representation cost of potential-based attention: increasing potential capacity can leave unavoidable output error, independent corrections can remove it, and fewer correction directions can require increasingly uneven output scaling.

Stronger quantitative theory (8g84: W1, Q1, Q2). Appendix O.1 gives an explicit lower-bound constant determined by keys, bias, and domain and explains its geometric dependence. Theorem 2 now gives matching upper and lower bounds for general residual spectra. In particular, rank⁡(R)=k+1<r\operatorname{rank}(R)=k+1<r implies optimal error Ek,χ∗=Θ(χ−2)\mathcal E_{k,\chi}^*=\Theta(\chi^{-2}). Appendix O.3 extends the bounds to nonuniform distributions with local continuous probability mass and extends the lower bounds to context-dependent banks by conditioning and averaging.

A concrete architectural decision (CPYi: W1, W2, Q1). If a certified error floor exceeds the desired tolerance, increasing potential capacity cannot reach that accuracy. The residual criterion determines the minimum correction dimension for exact recovery. For finite accuracy, Sec. 3.6 and Appendix P recover selected components of the remaining response error while retaining the convex-potential branch. Calibration errors select the directions, and the Jacobian formula identifies their derivative contributions.

Pretrained-model validation (CPYi: W1, W3, Q2, Q3; 8g84: W2). We replace attention to a fixed prefix in interleaved layers of Qwen2.5, recomputing later hidden states and queries through the modified model. The table reports final prediction preservation. PPL change and agreement compare with the original model; KL reduction compares with fitted corrections at the same direction budget.

ModelCorpusPPL change (%)KL reduction (%)Top-prediction agreement (%)
Qwen2.5-1.5BWikiText-2+0.0696.699.5
Qwen2.5-1.5BPenn Treebank+0.0195.099.4
Qwen2.5-0.5BWikiText-2+0.0697.899.3
Qwen2.5-0.5BPenn Treebank−0.0296.499.1

The additional controls distinguish structural choices and their costs:

ComparisonResultRevised manuscript
Increase potential width without correctionsAt most 0.2% median error reduction relative to an affine readoutTable 2
Add value-residual corrections at the largest width12.3% median error reduction, versus 6.6% with random directionsTable 2
Choose response-error directions instead of value-residual directions, keeping direct correction fixed31–62% lower predictive KLAppendix Q.2
Calibrated hybrid forward time / original forward timeSetting medians range from 0.95 to 1.01Appendix R

Appendix R also reports calibration, memory, and numerical sensitivity. These experiments demonstrate prediction-preserving replacement and quantify its cost. As the paper's main contributions are theoretical, and given the time and resource constraints, we have not conducted experiments on larger-scale models. Nevertheless, we believe that the additional experiments presented above demonstrate how our theoretical results perform in pretrained models. We will include more comprehensive experiments in the final version.

Dear Reviewers,

During the rebuttal period, in addition to the updates described in our common response, we have conducted further experiments and refined parts of the theoretical analysis in the paper.

As the discussion period is coming to an end, we would greatly appreciate hearing whether these additions have addressed your concerns and improved your overall assessment of our work.

Thank you all for your time, effort, and thoughtful feedback!

Best regards,

The Authors

AuthorsOpenReview ↗

Remarks for the Reviewers and Area Chair

Remarks for the Reviewers and Area Chair

We thank the reviewers and Area Chair for the discussion. N2do recommended weak acceptance. CPYi confirmed that the revisions addressed their original concerns about compatibility, affine spanning, and variation across heads and layers. 8g84 recognized the stronger theorem and maintained their score, citing the need to clarify its significance relative to prior work. We have also provided detailed responses to further questions and hope they will consider these clarifications and additional results in their final assessment. We summarize our core contribution and the key responses below.

Core contribution. When researchers redesign an attention layer to make it easier to interpret, they want the new layer to match the original layer's behavior. If it falls short, the usual fix is simply to scale it up. However, we show that the redesign itself can impose a hard accuracy limit that no amount of scaling can overcome. Rather than wasting compute training larger models, researchers can use the original layer upfront to detect this ceiling—and determine how much structural constraint to relax to recover the original behavior.

  • Relation to prior work (8g84, N2do). Hopfield–Fenchel–Young Networks use a potential to define a retrieval update with an energy guarantee. For an existing attention map, Theorem 1 gives the minimum number of independent correction directions needed to preserve its outputs. Table 1 compares this number with the classical reduced-rank baseline.

  • Cost of fewer corrections (8g84). When saving one direction requires increasingly uneven output scaling, Theorem 2 gives the optimal squared-error rate Θ(χ−2)\Theta(\chi^{-2}) in the allowed condition number χ\chi. The new upper bound is uniform over finite biases and requires no separate estimate of MpM_p (Appendix O).

  • Contribution of the potential (CPYi). With one correction direction on the Figure 1 target, the best affine readout had a mean squared error of 1.49×10−31.49\times10^{-3} on the evaluation grid. A fitted potential reached 8.36×10−88.36\times10^{-8} at χ=100\chi=100, meeting a precision target the affine readout could not attain.

  • Pretrained attention (CPYi, 8g84). All sampled 256- and 512-key Qwen2.5-1.5B banks had full affine dimension, with residual spectra varying across heads and layers (Appendix Q). At the largest tested width, value-residual corrections reduced median error by 12.3% relative to an affine readout, compared with 6.6% for random directions (Table 2).

  • Cache savings and attribution (CPYi). At about 20% prefix-storage savings, perplexity rose by 0.04–0.68% with the hybrid and 12.84–20.08% with direct value SVD. An affine predictor with the same correction rank matched the hybrid at the reported precision (Table 9).

  • Deployment cost (CPYi). Figure 3 traces the quality–storage trade-off as more layers are replaced. In separate timing runs under CUDA Graphs, full decoding took 7–9% longer than uncompressed grouped attention (Appendix S).

  • Stability after correction (CPYi). Outside the correction subspace, outputs still share one potential after removing the affine term. Our follow-up example shows that even one corrected direction can increase the retrieval energy and produce a two-cycle.

We thank the reviewers and Area Chair for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.