acceptodds
Under review as a conference paper at ICLR 2027

Residual-Aware Neural Spectral Capacity: Allocating Capacity across the Network from Architecture Specification Alone

13 views

Abstract

Predicting relative trained performance from architectural specifications can guide design when training costs limit experimentation. Neural Spectral Capacity (NSC) provides such a heuristic, but scores projections independently and assigns identical scores to compatible reorderings of the same layer configurations. We introduce Residual-Aware Neural Spectral Capacity (ResNSC) for Transformers, using a specification-derived statistical reference for accumulated residual updates to determine the input and residual-state statistics in each projection's capacity calculation, while retaining additive evaluation without data or instantiated weights. When residual widths vary, earlier choices affect later scores, so we augment NSC's allocator (NSC-DP) with residual history beyond its layer-and-budget state. ResNSC-DP jointly allocates residual and feed-forward widths under a resource budget, with certified bounds on the best numerically evaluated ResNSC score within a specified finite candidate set. At the 121M and 374M pretraining scales, selected architectures achieve lower mean perplexity than uniform, variable-width, and tapered allocations on FineWeb-Edu, WikiText-2, and LAMBADA at comparable parameter budgets. In inherited-weight LLaMA-7B pruning using the LoNAS supernet, ResNSC-DP improves eight-task average accuracy over NSC by 3.28–5.46 percentage points across eight matched parameter budgets. Paired reversals preserve the projection shapes and NSC scores yet worsen performance in both settings, exposing performance differences hidden by position-independent scoring.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Transformers, spectral capacity, residual connections, training-free architecture evaluation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_LSnSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Residual-Aware Neural Spectral Capacity (ResNSC), an extension of NSC for scoring neural architectures from their specifications without training or instantiated weights. ResNSC maintains a coordinate-wise reference for accumulated residual variance and uses it to construct projection-level input and noise covariances. A history-aware allocator then selects residual and FFN widths under a resource budget. The paper evaluates the method through from-scratch pretraining, inherited-weight pruning, and architecture-ranking benchmarks.

Critical strengths
  1. The paper demonstrates a meaningful limitation of NSC. The reversal experiments show that rearranging the same projection configurations can change trained performance despite preserving the NSC score.

  2. The mathematical formulation is clearly scoped. The local identities and optimization proofs appear sound under the stated reference-model assumptions. The history-aware state representation provides a principled way to optimize the proposed numerical objective, with guarantees restricted to the specified finite candidate set.

  3. The architecture-selection results are promising. Table 2 shows improved pretraining performance over NSC-selected architectures at comparable parameter budgets. Table 4 reports substantial pruning improvements across matched budgets using the same supernet and candidate menus.

Critical weaknesses
  1. The architectural setting needs clearer explanation and separation across experiments. At pretraining, the full residual state has a fixed width within each model with only a prefix of its coordinates update at each block; at pruning and ranking experiments, constant active residual widths are used. This distinction is essential for understanding both the architecture and the residual ledger, but is easy to miss.

  2. The pretraining comparisons do not establish matched training compute. Different residual-width profiles can incur different attention FLOPs. The paper does not report total training FLOPs for each pretraining configuration. It is therefore unclear whether the improvements persist under a matched training-FLOP budget.

  3. The statement - "that architectural context can help guide resource allocation." - is loosely supported by the provided benchmarks in the ranking experiment. Table 5 shows nearly unchanged correlations on FlexiBERT and GPT-2, with an improvement on AutoFormer-Small. Increasing the number of distinct scores demonstrates greater sensitivity to architectural differences, but does not by itself establish that those distinctions are predictive.

Questions for authors

See the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer LSnS (1/3): W1

We thank the reviewer for recognizing the limitation of NSC demonstrated by the reversal experiments, the clearly scoped mathematical guarantees, and the promising architecture-selection results. In response, we clarify the architectural settings across experiments (W1), report training-FLOP estimates and add comparisons with VWT at matched training FLOPs at both scales (W2), and refine our prediction claims with direct pairwise evidence for architectural distinctions that ResNSC enables us to resolve (W3).

1. Architectural setting and residual history of each experiment (W1)

Thank you for highlighting this distinction. We clarify below how the architectural settings and corresponding residual histories differ across experiments.

In the pretraining architectures, the full residual-state width is fixed within each model, while the active residual width may vary across blocks. Each block's attention and FFN branches read and update only the prefix of residual coordinates within that block's active width; the remaining coordinates are preserved. The FFN hidden width controls the expansion inside the branch, whose output returns to the active residual coordinates (Sec. 3.2; Appendix B.1).

ExperimentActive residual widthArchitectural choices evaluatedReference history used for scoring
121M/374M pretrainingMay vary across blocksJoint allocation of residual width and FFN hidden widthCoordinate-wise history determined by preceding active-width choices
LoNAS-LLaMA-7B pruningFixed at 4,096 across all 32 blocksFFN widths and Q/K/V adapter ranksShared across active coordinates; varies with branch position
FlexiBERT, GPT-2, AutoFormer rankingConstant within each candidate; may differ between candidatesExisting candidates with benchmark-specific depth, attention, and FFN configurationsShared across active coordinates; varies with branch position

The fixed-width pruning and ranking settings therefore retain position-dependent scoring, while variable-width pretraining also accounts for differences in residual history across coordinates.

Changes in the revised paper. Sec. 3.2 will distinguish the full residual-state width, active residual width, and FFN hidden width, and explicitly state that inactive coordinates are preserved. The comparison table above will be added at the beginning of Sec. 4. The setup paragraphs of Secs. 4.1–4.3 will identify each experiment's active-width setting and corresponding residual history.

AuthorsOpenReview ↗

Response to Reviewer LSnS (2/3): W2

2. Comparison at matched training FLOPs (W2)

Thank you for highlighting the fact that comparable parameter counts and equal training tokens do not necessarily imply equal training FLOPs, because different residual widths can change attention cost. The two tables below show that ResNSC's performance benefits are not solely explained by higher training FLOPs. The first examines the comparisons reported in the submitted paper; the second adds matched-compute comparisons with VWT at both scales.

Existing comparisons. The following table pairs estimated training FLOPs with the mean WikiText-2 and LAMBADA perplexities reported in Table 2 of the submitted paper. At both scales, ResNSC achieves lower perplexities than NSC with 3.1% and 4.1% fewer estimated FLOPs, and lower perplexities than Reverse at equal estimated FLOPs. At 374M, it also achieves lower perplexities than Uniform and TLM with fewer estimated training FLOPs.

Method121M EFLOPs121M WikiText-2121M LAMBADA374M EFLOPs374M WikiText-2374M LAMBADA
ResNSC3.13751.13202.826.46331.2053.0
NSC3.23753.40221.427.60432.4066.9
Reverse3.13753.22234.626.46333.9377.7
Uniform (a)2.95055.47262.926.77733.8276.2
TLM (a)2.95055.43252.126.77734.1978.1
VWT2.71152.33252.625.01731.9970.6

(a) Projection-parameter totals are nearly equal within each scale, but ResNSC's mean active width is larger at 121M (730 versus 640) and smaller at 374M (992 versus 1,024). Attention-product cost follows the sum of active widths, explaining the higher/lower FLOPs relative to Uniform/TLM. Each scale is searched independently under a parameter ceiling.

Reduced-budget comparison with VWT. In the submitted paper, the architectures found by ResNSC at both scales use more training FLOPs than VWT. To test whether the gains persist without this additional compute, we reduce ResNSC's training-token budgets to approximately 86.4% and 94.5% of the submitted-paper budgets at 121M and 374M, respectively, matching VWT's 2.711 and 25.017 EFLOPs. Token accounting includes both main training and final decay. The table compares the reduced-budget runs with VWT's submitted-paper means.

Method121M EFLOPs121M WikiText-2121M LAMBADA374M EFLOPs374M WikiText-2374M LAMBADA
ResNSC (reduced training-token)2.71152.53212.025.01631.5552.9
VWT2.71152.33252.625.01731.9970.6

At matched training FLOPs, ResNSC retains lower LAMBADA perplexity at both scales and lower WikiText-2 perplexity at 374M. At 121M, WikiText-2 perplexity is 0.20 higher than VWT's (52.53 versus 52.33). The same reduced-budget 121M run also uses fewer estimated FLOPs than Uniform and TLM (2.711 versus 2.950 EFLOPs) and retains lower WikiText-2 and LAMBADA perplexities than both (52.53 and 212.0, versus 55.47 and 262.9 for Uniform and 55.43 and 252.1 for TLM).

Training FLOP estimation. We use the same dense matrix-multiplication estimate for every method:

Cdense=6N[∑ℓ=1L(4dℓ2+3dℓfℓ+2Sdℓ)+VE].C_{\mathrm{dense}}=6N\left[\sum_{\ell=1}^{L}\left(4d_\ell^2+3d_\ell f_\ell+2S d_\ell\right)+VE\right].

Here dℓd_\ell and fℓf_\ell are the active residual and FFN hidden widths, S=4096S=4096 is sequence length, V=32000V=32000 is vocabulary size, and EE is the embedding/readout width (640 at 121M; 1024 at 374M). NN includes main-training and final-decay tokens. The estimate covers attention/FFN projections, attention products, and vocabulary readout, counting two FLOPs per multiply-add and the backward pass as twice the forward pass. One EFLOP is 101810^{18} operations.

Changes in the revised paper. Sec. 4.1 will clarify the training-FLOP comparisons at both scales and report the reduced-budget comparisons with VWT: lower LAMBADA perplexity at both scales, lower WikiText-2 perplexity at 374M, and 0.20 higher WikiText-2 perplexity at 121M. The training-and-evaluation appendix will include both comparison tables, the dense FLOP-estimation formula, and the token-budget adjustments to approximately 86.4% and 94.5% of the submitted-paper budgets at 121M and 374M, counting both main training and final decay.

AuthorsOpenReview ↗

Response to Reviewer LSnS (3/3): W3

3. What the ranking experiments show (W3)

We agree with the reviewer that the greater architectural sensitivity obtained by incorporating residual context in this work does not establish better prediction in all cases. We distinguish two regimes:

  • Global ranking of diverse architectures with large variations in model size. Parameter count alone can already be a strong performance predictor in this regime (Kaplan et al., 2020; Serianni and Kalita, 2023; Wang, 2025; see also Appendix I). Residual context does not consistently improve prediction over NSC across the evaluated benchmarks: correlations are nearly unchanged on FlexiBERT and GPT-2 (Table 23), with mixed results on AutoFormer (Appendix I).

  • Substantial benefits for subtle variations that NSC cannot distinguish. At similar model sizes, NSC supplements parameter count by capturing how parameters are organized into projections (Zhu et al., 2026, Sec. 4.1). However, its position-independent aggregation cannot distinguish permutations of the same layer configurations. Resolving these differences is important for designing Transformers with heterogeneous layer settings and for pruning Transformers; our residual-context model enables ResNSC to distinguish such arrangements.

The submitted paper supports this indirectly through the improved performance of architectures selected by ResNSC-DP in variable-width Transformer design and pruning (Tables 2 and 4). To directly test the predictive value of the additional distinctions, the table below compares pairs with identical or similar NSC scores from 1,000 evaluated LoNAS subnetworks, using seven-task mean accuracy under inherited weights.

NSC score differencePairsPairwise accuracy: ResNSCPairwise accuracy: NSCPairwise Kendall τPW\tau_{\mathrm{PW}}: ResNSCPairwise Kendall τPW\tau_{\mathrm{PW}}: NSC
Identical15571.0%--0.42--
Within 1% of range17,70073.8%51.3%0.480.03
Within 5% of range88,04674.2%52.7%0.480.05
Within 10% of range171,90074.0%55.3%0.480.11

The identical-score pairs differ only in the ordering of their FFN widths across layers and have equal parameter counts; ResNSC orders 110 of these 155 pairs correctly. For pairs with similar NSC scores, its ranking accuracy is 73.8–74.2%, versus 51.3–55.3% for NSC. These results directly demonstrate useful performance distinctions among architectures that NSC cannot separate or scores similarly.

Changes in the revised paper. Sec. 4.3 will clarify that the additional architectural sensitivity from residual context does not establish better prediction in all cases. It will distinguish global ranking across diverse model sizes from resolving layer arrangements that NSC cannot distinguish, and cite prior work supporting parameter count as a strong predictor in the former regime. Appendix I will add the LoNAS pairwise results and evaluation protocol. The resource-allocation conclusion will be grounded in the architecture-selection results in Secs. 4.1–4.2 (Tables 2 and 4), with the new pairwise analysis providing direct prediction evidence.

Reviewer_pP6DOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes ResNSC to address NSC's inability to distinguish different layer orderings. ResNSC constructs a statistical reference for the accumulated residual state from architectural specifications, incorporating residual context and within-branch propagation into per-projection capacity scores. It represents history through residual-width counts and introduces a search algorithm that jointly allocates residual and FFN widths. Experiments cover language model pretraining at two scales, LoNAS-LLaMA-7B subnetwork selection, and multiple architecture-ranking benchmarks.

Critical strengths
  1. The reversal experiment directly exposes a blind spot in NSC. In Table 2, reversing the 374M configuration increases FineWeb-Edu perplexity from 13.60 to 14.48, while the collection of projection shapes and the NSC score remain unchanged. This provides a more direct motivation for accounting for layer position than performance correlations alone.
  2. The main comparison with NSC is well controlled. The pretraining experiments use the same candidate menu and parameter cap, and the main 121M experiment includes three independent training runs. Across eight matched parameter budgets, the LoNAS experiments improve average accuracy by 3.28–5.46 percentage points (Table 16).
  3. The method is clearly connected to the search objective. Under the specified reference model, width counts are sufficient to determine future scores. Appendix E provides arguments for dominance pruning and completion upper bounds. The optimality guarantee is explicitly limited to numerical scores over the finite candidate set.
Critical weaknesses
  1. Sensitivity of the residual reference to its specific conventions is insufficiently validated. Section 3.2 introduces additional inverse-background weighting after normalization, yielding weights that decay with depth when the residual width is fixed at full width. Unit-variance updates are also assumed. The depth-only and mean-history ablations in Table 3 are helpful, but do not establish whether the selected architectures are robust to the update magnitude and discounting form. In particular, the LoNAS results with fixed residual width do not independently establish the necessity of coordinate-level history modeling. It remains unclear how much the gains depend on these specific conventions.
  2. Ranking results under approximately matched budgets do not show a consistent advantage. Table 19 in Appendix I already provides pairwise comparisons with relative parameter differences below 10%, and ResNSC performs worse than NSC on three of the four benchmarks. For AutoFormer-Small, pairwise accuracy drops from 67.42% to 66.16%. Although the global correlation improves, these results limit the extent to which ResNSC can be regarded as a reliable rule for architecture selection within a budget.
  3. Evidence on variability in the larger-scale results is limited. The 374M experiment and the 121M ablations each use only one full training run and two late-stage continuation runs. These error bars cannot capture variability across independent full training runs. This is particularly relevant because, relative to VWT, the 374M model improves perplexity only from 13.65 to 13.60 and Avg8 from 44.26% to 44.60% (Table 2).
  4. The search overhead introduced by the additional history state is not sufficiently reported. Table 9 in Appendix E.6 reports a zero numerical gap, but does not provide runtime, peak memory, or the actual number of states for the joint width search. The 0.28 seconds reported in Appendix H is for the LoNAS search with fixed residual width and does not represent the cost of the joint search.
Questions for authors
  1. For which types of candidates do the declines in Table 19 primarily occur? How should these results be interpreted alongside the positive results for architecture allocation in pretraining, and how should the scope of applicability be defined?
  2. Are architecture selections stable when the reference update magnitude or the form of inverse-background discounting changes? Can the existing evidence further distinguish the roles of the depth prior and coordinate-level residual history?
  3. How much time and memory did each of the two joint width searches use, and how many states did each visit? What set of resource prices was used?
4 comments
AuthorsOpenReview ↗

Response to Reviewer pP6D (1/3): W1 & Q2

We thank the reviewer for the careful review and for recognizing the reversal experiments, the controlled comparison with NSC, and the link between the history representation and the search objective. All four comments are specific, and we answer each with data: W1 and W3 with new independent training runs, W4 with new measurements, and W2 with new pairwise ranking evidence and its connection to the allocation results.

1. Sensitivity of the residual reference; depth dependence and coordinate-level history (W1 / Q2)

Sensitivity to the update magnitude cc. The update magnitude cc sets the reference variance of each independent residual update per active coordinate, relative to unit embedding-coordinate variance. The paper's c=1c=1 therefore assigns equal per-coordinate reference variance to the embedding and residual updates (Sec. 3.2, Eq. 2; Appendix B.1). At c=0c=0, the reference ignores residual updates. Increasing cc makes reference power accumulate faster and, in the full-width specialization, discounts later branches more strongly relative to earlier ones (Appendix B.3). We evaluate its effect on architectural selection and trained-model perplexity in the 121M search using the same width menus and parameter budget as in the paper. The reported values are means ± standard deviations over three independent full training runs:

MethodBlocks changed relative to c=1c=1 (out of 16)WikiText-2 PPL ↓
ResNSC (c=0c=0) (same selection as NSC)1553.40 ± 0.31
ResNSC (c=0.5c=0.5)051.13 ± 0.21
ResNSC (c=0.75c=0.75)051.13 ± 0.21
ResNSC (c=1c=1) (paper's unit update)051.13 ± 0.21
ResNSC (c=1.5c=1.5)752.18 ± 0.49
ResNSC (c=2c=2)752.18 ± 0.49
ResNSC (c=3.0c=3.0)852.02 ± 0.41
NSC–53.40 ± 0.31
Uniform–55.47 ± 0.07
TLM–55.43 ± 0.53
VWT–52.33 ± 0.15

The results support the paper's unit-update choice and show stable selection around c=1c=1. The settings c=0.5,0.75,1c=0.5, 0.75, 1 all select the paper's architecture, whose mean WikiText-2 PPL (51.13 ± 0.21) is lower than that of every architecture selected at c=0c=0 or at larger cc. Larger values c=1.5,2,3c=1.5, 2, 3 change 7 to 8 blocks and select alternatives with higher WikiText-2 PPL (52.18 ± 0.49 and 52.02 ± 0.41). Every setting with c>0c > 0 has lower mean WikiText-2 PPL than NSC, Uniform, TLM, and VWT.

What coordinate-level history adds beyond depth dependence. We agree that the fixed-width LoNAS pruning experiments cannot distinguish coordinate-level history from depth dependence: all active coordinates share the same reference history. We therefore assess the added value of coordinate-level history in the variable-width Transformer experiments (Table 3), where preceding layer widths can produce different histories across residual coordinates. All scoring variants use the same 121M width menus and parameter budget. Depth-only scoring retains branch position but removes dependence on preceding residual widths; mean-history scoring retains average reference power over the active coordinates but removes coordinate-wise differences (scoring-ablation protocol). The table reports updated means and standard deviations over three independent full training runs per selected architecture, replacing the original statistics from one full run and two final-decay continuations:

Residual information retainedWikiText-2 PPL ↓LAMBADA PPL ↓
No residual context (same selection as NSC)53.40 ± 0.31221.4 ± 10.1
Depth only53.34 ± 0.24214.5 ± 10.6
Mean history only54.51 ± 0.70235.3 ± 7.3
Full coordinate-level history51.13 ± 0.21202.8 ± 13.5

Depth-only scoring yields only a small reduction in mean PPL relative to no residual context, whereas full coordinate-level history yields substantially larger reductions in mean PPL on both benchmarks. Keeping only the mean history instead yields higher mean PPL than no residual context. This highlights the challenge of incorporating residual context effectively: a simple averaging strategy can adversely affect performance, while our coordinate-level representation supports better width allocation in the tested variable-width setting.

AuthorsOpenReview ↗

Response to Reviewer pP6D (2/3): W2 & Q1

2. Ranking scope and allocation benefits of ResNSC (W2 / Q1)

Scope of applicability of ResNSC in ranking architectures. ResNSC incorporates residual context so that a layer configuration's score depends on its position and preceding layer configurations.

  • Architectures with comparable parameter counts. We agree with the reviewer that incorporating residual context does not necessarily ensure a consistent ranking advantage over NSC, as reflected in Table 25.
  • Permutations of the same layer configurations. NSC's position-independent aggregation assigns identical scores to such permutations, while ResNSC can distinguish them through their different residual contexts. We acknowledge that the original submission did not make the predictive value of this distinction sufficiently clear. We therefore conducted additional analysis, with results summarized in the table below.

We compare pairs with identical or similar NSC scores from 1,000 LoNAS subnetworks, using seven-task mean accuracy under inherited weights. The score-difference thresholds are percentages of the NSC-score range across these subnetworks; Table 25 instead restricts differences in parameter count.

NSC score differencePairsPairwise accuracy: ResNSCPairwise accuracy: NSCPairwise Kendall τPW\tau_{\mathrm{PW}}: ResNSCPairwise Kendall τPW\tau_{\mathrm{PW}}: NSC
Identical15571.0%--0.42--
Within 1% of range17,70073.8%51.3%0.480.03
Within 5% of range88,04674.2%52.7%0.480.05
Within 10% of range171,90074.0%55.3%0.480.11

The identical-score pairs have equal parameter counts and differ only in the ordering of their FFN widths across layers; ResNSC orders 110 of these 155 pairs correctly (71.0%), while NSC ties them. For pairs with similar NSC scores, ResNSC achieves 73.8–74.2% pairwise accuracy, compared with 51.3–55.3% for NSC. These results show that residual context provides useful predictions of performance differences among the evaluated LoNAS subnetworks, both for layer permutations that NSC cannot distinguish and for architectures it scores similarly.

Interpretation of the variable-width Transformer design and pruning results. Both tasks require deciding how resources should be distributed at each layer. Even the same collection of layer configurations can yield different performance when reordered. NSC assigns these permutations the same score, whereas ResNSC provides contextual scores that ResNSC-DP can use when choosing layer configurations and their placement. Thus, distinguishing layer permutations is the key capability critical to these allocation tasks.

The reversal experiments further demonstrate why this capability matters. Reversing the selected configurations preserves the collection of projection shapes and the NSC score, yet raises WikiText-2 perplexity from 51.13 to 53.22 at 121M and from 31.20 to 33.93 at 374M (Table 2). In pruning, reversing the FFN-width assignments reduces Avg8 from 68.87% to 59.49% (Table 4). These results, together with the new pairwise analysis, support the value of residual context for layer allocation. The positive results on these two tasks therefore do not contradict Table 25: ResNSC can improve placement decisions without consistently improving ranking across all architectures with comparable parameter counts.

AuthorsOpenReview ↗

Response to Reviewer pP6D (3/3): W3, W4 & Q3

3. Reliability across independent training runs (W3)

We agree that variability in the 374M comparison and the 121M ablations should be assessed using independent full training runs. The updated 121M ablation table reports three independent full training runs per selected architecture, each starting from a fresh initialization with a different seed.

121M ablations. All scoring variants use the same search menu and parameter budget. The table reports means and standard deviations across the independent runs:

Scoring ruleFineWeb-Edu PPL ↓WikiText-2 PPL ↓LAMBADA PPL ↓
ResNSC (full)19.39 ± 0.0151.13 ± 0.21202.8 ± 13.5
Without propagated spectrum19.64 ± 0.0752.25 ± 0.61187.3 ± 6.0
Depth only19.92 ± 0.0753.34 ± 0.24214.5 ± 10.6
No residual context (selects the same structure as NSC)20.00 ± 0.0353.40 ± 0.31221.4 ± 10.1
Mean history only20.14 ± 0.0854.51 ± 0.70235.3 ± 7.3

The ordering by mean perplexity is identical on FineWeb-Edu and WikiText-2, with full ResNSC lowest on both, supporting the ablation conclusions on these benchmarks. The variant without propagated spectrum has the lowest mean LAMBADA perplexity.

374M comparison. The table reports three independent full training runs for ResNSC and VWT at 374M scale:

MethodFineWeb-Edu PPL ↓WikiText-2 PPL ↓LAMBADA PPL ↓Avg$_8$ (%) ↑
ResNSC13.61 ± 0.0131.22 ± 0.1452.5 ± 5.544.61 ± 0.12
VWT13.65 ± 0.0231.96 ± 0.2171.6 ± 6.444.28 ± 0.09

The independent-run results for 374M confirm the reliability of the 121M ablation conclusions: ResNSC achieves lower mean PPL than VWT on all three benchmarks and higher Avg$_8$ accuracy. Together with the 121M ablation results, these findings support the stability of the residual-context allocation across scales.

4. Complete cost of the joint search (W4 / Q3)

Short answer: the complete joint search takes 2.69 minutes at 121M and 19.34 minutes at 374M (single CPU thread, starting from an empty cache), with peak memory of 0.93 GiB and 6.72 GiB and 1.45 million and 13.71 million history states.

Both measurements were taken on the same Intel Xeon Gold 6133, on a single thread, running ResNSC-DP from an empty search cache. Timing covers history-graph construction, spectral-score precomputation, resource-price and completion-bound construction, feasible-solution preparation, and the final search; peak memory covers the whole pipeline. Each scale was measured once.

Measurement121M374M
Complete first-run time161.13 s (2.69 min)1160.30 s (19.34 min)
History-graph construction6.88 s134.14 s
Spectral-score precomputation and saving137.10 s971.74 s
Resource-price and completion-bound precomputation9.87 s46.88 s
Feasible-solution preparation and final search5.49 s5.70 s
Peak memory of the full pipeline (RSS)0.93 GiB6.72 GiB
Distinct precomputed history states1,447,47013,705,367
Final numerical gap00

Total time includes program initialization and first-time compilation. "History states" counts the history graph that is actually built and evaluated.

The bulk of the cost is the spectral-score precomputation (84%–85%), which grows with the number of history states; the final search itself takes only a few seconds. The resource price is determined by a budget-guided bisection (2.581637×10−42.581637\times10^{-4} at 121M and 1.189476×10−41.189476\times10^{-4} at 374M), with 45 and 36 bisection steps, respectively, all of which are included in the bound precomputation above. The price is used only to construct upper bounds on the score of the remaining part; the final search optimizes the original ResNSC objective. Both runs recover the per-layer configurations used in the paper, with a final numerical gap of zero.

AuthorsOpenReview ↗

Follow-up to Reviewer pP6D

Dear Reviewer pP6D,

Thank you again for your thorough and insightful review. Your questions on the update magnitude, the ranking comparison, training variability, and the cost of the joint search led to new experiments, including a sensitivity study on the update magnitude and three independent full runs for the ablations and the 374M comparison, and the paper is stronger for them. The corresponding changes are in the revised paper, marked in blue.

As the discussion period is drawing to a close, we would like to kindly ask whether our responses have addressed your concerns.

Thank you very much for your time and effort.

Best regards,

The Authors

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

This paper proposes a new quantity for Neural Spectral Capacity (NSC) which is called ResNSC (Residual-Aware NSC). The existing approach has a drawback such that it is determined independent of the layer orders because it assumes each layer takes unit variance input. The proposed method traces the change of input variance through layers by calculating the accumulated residual variance recursively. Based on this quantity, the authors propose to optimize the network architecture so that ResNSC is maximized under a budget constraint. The numerical experiments show a clear improvement compared with NSC and other related approaches.

Critical strengths
  • Usefulness: It is quite natural to take the change of input variance across layers. This paper proposes a simple method to resolve this problem. The proposed ResNSC is easy to calculate and is useful for architecture search. Actually, the method does not use information of trained matrix and the internal state's distribution for reference input. It makes architecture search extremely fast.
  • Numerical experiments: The numerical experiments show a good improvement of performance compared with some baseline methods. This is an encouraging result.
Critical weaknesses
  • Theoretical justification: While the motivation of the proposed method is convincing, its derivation is still heuristic. Actually, theoretical justification of the proposed method (and NSC) heavily depends on a lot of approximations such as Gaussian assumption, no-correlation between channels, ignorance of nonlinearity of each computational block. Although the simplicity of methods and its technical correctness would be in a relation of trade-off, validity of these approximation are not well justified.
  • Numerical experiments: (i) The examined models are rather small as up-to 374M for pre-training and 7B for pruning LLaMA-7B. (ii) And, the performance improvement in Table 2 from some baseline methods such as VWT looks small. It is a bit difficult to say significant improvement by the proposed method in this experiment. (iii) Finally, the compared baseline methods are not the state of the art architecture search methods. There are several NAS and pruning methods proposed in the literature. To justify usefulness of the approach, it is better to compare it with stronger baselines.
Questions for authors
  • Is architecture search only the application of ResNSC (and NSC)? If there are more situations where ResNSC is useful, it is informative to list them up.
  • A good point of ResNSC is that it does not require any information of weight and input, but it could deteriorate the potential performance. I am wondering how much information and outcome performance this simplification looses.
6 comments
AuthorsOpenReview ↗

Response to Reviewer eQBD (1/4): W1

We thank the reviewer for recognizing that tracing the input variance across layers is natural, that ResNSC is easy to compute, and that the experimental results are encouraging. Below we respond to the two weaknesses (W1, W2) and the two questions (Q1, Q2).

W1: Basis of the modeling assumptions

We agree with the reviewer that simplicity and general applicability must be balanced against modeling fidelity. ResNSC uses a statistical reference to incorporate residual context into NSC while retaining evaluation from architectural specifications alone. Below, we clarify the three assumptions raised by the reviewer, distinguishing modeling choices from their mathematical consequences and examining the supporting evidence.

  • Gaussian reference. We choose Gaussian input and independent Gaussian noise with prescribed covariances. This standard information-theoretic reference provides a closed-form information measure for a linear projection, allowing architectural context to enter through those covariances (Cover & Thomas, 2006; Telatar, 1999). Conditional on the projection matrix, Eq. (1) follows exactly under this reference (Sec. 3.1; Appendix A).

  • Correlation between channels. The uncorrelated form is used for the reference of the shared residual stream: for each coordinate, the ledger records the reference variance accumulated from the updates of preceding branches (Sec. 3.2, Eq. (2)), so the residual context follows directly from the preceding residual widths. Correlations enter the score within each branch. The attention output and FFN down projections receive features transformed inside the branch, and their input covariance (Sec. 3.3, Eq. (4)) is a full matrix with off-diagonal entries, which ResNSC retains inside the log-determinant. Removing this propagated spectrum changes the selected 121M architecture and raises FineWeb-Edu perplexity from 19.39 to 19.64 (revised Table 3).

  • Nonlinearity. ResNSC scores the linear projections, and the nonlinearities enter through their references. The FFN propagation reference combines a Gaussian matrix with an independent SiLU gate, whose statistics enter the input spectrum of the down projection; normalization is modeled in mean-field form and preserves the relative variances of the residual coordinates (Secs. 3.2–3.3; Appendix C). Appendix D validates the deterministic spectral formulas against finite-matrix sampling within the prescribed reference, including the SiLU-gated case: the largest relative differences are 0.34% for read projections and 0.69% for write projections.

Empirical validity of the above modeling assumptions as a whole. We assess the combined reference through the quality of the architectures it selects. Under the same search menus, parameter ceilings, and training protocol as NSC, ResNSC lowers FineWeb-Edu perplexity from 20.00 to 19.39 at 121M and from 13.82 to 13.60 at 374M (Table 2). On LoNAS, it raises Avg$_8$ by 3.28–5.46 points across eight matched budgets (Fig. 3). These results support the combined reference's usefulness for allocation in the tested settings; they do not establish that each assumption reproduces trained-network statistics.

AuthorsOpenReview ↗

Response to Reviewer eQBD (2/4): W2

W2: Baselines, experimental scale, and practical benefits

We emphasize that the practical value of ResNSC should be assessed through both the performance of the selected architectures and the cost of identifying them. ResNSC identifies layer-wise resource allocations directly from architectural specifications, without training candidate models or evaluating them on data during selection.

Choice of baselines. For variable-width Transformer design, we use strong recent baselines VWT (arXiv, June 2026) and TLM (arXiv, June 2026), which study whole-block and FFN-width allocation, respectively. Both report improvements over parameter-matched uniform models across scales. For LLaMA-7B pruning, we also use strong and recently published baselines NSC (NeurIPS 2026) and W-PCA (ICLR 2025); we also include the classic methods GradNorm, SNIP, and SynFlow. Following the reviewer's suggestion, we also include published results from the original LoNAS validation-guided NSGA-II search (Muñoz et al., LREC-COLING 2024) as an additional baseline that directly evaluates candidate subnetworks using validation accuracy.

Experimental scale. For Transformer architecture design, our 374M pretraining setting is comparable in scale to VWT's initial 500M shape comparison and TLM's 440M schedule-selection study (Wu et al., 2026, Sec. 3.2; Bayat et al., 2026, Sec. 3.2). For pruning, we use the same LLaMA-7B setting as the comparisons with NSC and W-PCA in Sec. 4.2. We are committed to extending the heterogeneous-layer Transformer design experiments to the 1B pretraining scale before the final version, as completing this experiment is not feasible within the rebuttal period.

Practical benefits: performance and architecture-identification cost.

Variable width Transformer design. We believe ResNSC offers meaningful improvements over the baselines, with virtually no cost for architecture identification compared with pretraining. At 374M, it reduces WikiText-2 and LAMBADA perplexity by 2.5% and 24.9% relative to VWT, respectively, and raises Avg$_8$ from 44.26% to 44.60%, under the same training-token budget and comparable parameter counts. Identifying the architecture takes only 19.34 minutes on one CPU thread, without training candidate models.

MethodFineWeb-Edu PPL ↓WikiText-2 PPL ↓LAMBADA PPL ↓Avg$_8$ (%) ↑Identification time
ResNSC13.6031.2053.044.6019.34 min (CPU)
NSC13.8232.4066.944.265.65 s (CPU)
VWT13.6531.9970.644.26$\sim$40 days (a) (8 A100s)
TLM14.4334.1978.143.73$\sim$20 days (a) (8 A100s)

(a) Estimated time for the original profile-selection sweeps: VWT trains 24 candidates at 500M for 25B tokens each; TLM trains 15 at 440M for 30B tokens each. We extrapolate from our 374M implementations on eight A100-40GB GPUs: 2.376 s per 524,288-token step for VWT and 1.922 s for TLM, scaling linearly with parameters, training tokens, and candidate count. These are rough sequential-training estimates on one eight-GPU node; profiles and implementation efficiency differ. They exclude controls and later experiments.

AuthorsOpenReview ↗

Response to Reviewer eQBD (3/4): W2 (continued) & Q1

W2 (continued): Practical benefits in pruning

Pruning. At the common 5.7B parameter ceiling, ResNSC achieves 68.87% Avg$_8$, compared with 65.47% for NSC and 63.83% for the strongest proxy baseline, W-PCA (Table 4). Its scoring and selection across all eight budgets take under one second on one CPU thread, compared with the reported 22–45 minutes for the proxy-based searches. Thus the selected subnetwork improves mean task accuracy while requiring little computation to identify. The table reports each method's selected model size; search cost excludes supernet training and final subnetwork evaluation.

MethodParameters (B)Avg$_8$ (%) ↑Architecture-identification time
ResNSC5.6968.87< 1 s
NSC5.6965.47< 1 s
W-PCA5.6963.8345 min
GradNorm5.5062.4522 min
SNIP5.4460.7322 min
SynFlow5.2360.6823 min

Additional comparison against NSGA-II on LLaMA-7B pruning with multiple FLOPs-targets. For architectures A–F shown below, ResNSC's reported test Avg$_8$ is consistently 1.09–1.85 percentage points higher than NSGA-II results from directly searching on validation accuracy. ResNSC scores and selects architectures across all eight budgets in under one second on one CPU thread, whereas NSGA-II evaluates 200 subnetworks over 1–2 days on one GPU.

ArchitectureTFLOPsNSGA-II on validation acc. Avg$_8$ (%) ↑ResNSC Avg$_8$ (%) ↑Improvement (pp)
A1.4167.168.69+1.59
B1.4067.168.70+1.60
C1.3966.968.34+1.44
D1.3867.068.09+1.09
E1.3766.667.80+1.20
F1.3665.967.75+1.85

Q1: Uses beyond architecture search

NSC and ResNSC aim to quantify how parameter organization affects learning capability from network specifications alone. Architecture search is one application of this measure. Beyond the architectural assessment and resource allocation demonstrated in this paper, the information-channel perspective motivates the following broader uses and research directions.

  • Architecture understanding. ResNSC provides a quantitative way to study how a projection's position, preceding structure, and read/write role affect its assessed contribution. The ranking experiments (Sec. 4.3) support direct architecture assessment, while the reversal controls (Tables 2 and 4) show that rearranging the same projection shapes can change performance despite an unchanged NSC score. This offers a tractable tool for investigating the relationship between parameter organization and learned performance, complementing theoretical studies of expressivity and trainability.

  • Studying Transformer training. Extending the reference analysis to controlled learning models could connect the spectra underlying architectural capacity to how quickly task-relevant directions are learned. Accounting for alignment with the learning task, available data, and training duration could clarify when greater architectural capacity translates into lower held-out loss within a finite training budget, helping explain how architecture affects learning dynamics and training requirements.

  • Architecture-aware scaling laws. Architectural capacity could augment scaling laws based on model size and training tokens by accounting for how parameters are organized. The gains observed at 121M and 374M (Table 2) motivate studying how allocation benefits change with scale. With empirical calibration, capacity-augmented laws could forecast performance across sizes and allocations, estimate training resources needed for a target loss, and support transfer of scaling predictions across architecture families.

  • Broader architectural design. Extending the contextual capacity formulation could provide design principles for parameter activation, reuse, and sharing. For MoE, distinguishing the capacity available across the expert pool from that activated per token could guide expert width, count, and selection. For looped Transformers, accounting for shared weights across passes could clarify when additional computation is beneficial and when it becomes redundant. For KV-cache sharing in grouped-query attention, modeling dependencies among heads could identify when sharing preserves complementary information and when it creates a bottleneck. These extensions would support design under parameter, computation, and cache-memory constraints.

Reviewer_LGePOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a Residual-Aware Neural Spectral Capacity for transformers. It uses a specification-derived statistical reference for accumulated residual updates. The method is an extension of Neural Spectral Capacity (NSC). This paper introduces a Residual History to NSC. The method is tested on 121M and 374M pretraining scales with several datasets. The method outperforms several similar methods and significantly improves NSC.

Critical strengths
  1. The experiments are solid and robust. The method was tested in different settings on different datasets with proper Ablation. It also compares with and outperforms recent similar methods. All of these make the empirical part of the paper strong.
Critical weaknesses
  1. The method seems to be an incremental extension of NSC by adding residual context. This raises concerns about the paper's true novelty, which weakens it.
Questions for authors
  1. Can the author(s) clearly explain the differences and novelty of the method and the core contribution of the paper over NSC?
  2. In Table 4, the method gains over 3% in avg accuracy over NSC. Was NSC run completely on similar settings or a budget for fair comparison or is it totally the impact of Residual Context? If it is the impact of Residual Context, please explain why Residual Context matters and how it gives a significant performance boost on this task.
  3. Introducing Residual Context/History should bring an overhead to the method. Can the author(s) confirm whether there is any overhead or not? If there is an overhead, how was it engineered to solve it?
  4. The author(s) showed the same cost for the method as NSC in Table 4. Can the author(s) explain and verify this?
9 comments
AuthorsOpenReview ↗

Response to Reviewer LGeP (1/3): W1 & Q1

W1 & Q1: Relation to NSC, key differences, and core contribution

Thank you for asking us to clarify the relationship between NSC and ResNSC. We explain their common motivation, the technical advances over NSC, and the practical benefits below.

1. Common ground of ResNSC and NSC

NSC and ResNSC share the same scientific motivation: how does the architectural organization of parameters affect learning capability, and can we quantify this from network specifications alone, before training, without instantiating the network or running it on data?

Our overall approach is to model linear projections in a Transformer as noisy information channels and characterize their capacity using information theory. Architectural dimensions and initialization conventions specify reference distributions for the projection matrices, allowing their spectral capacities to be evaluated without sampling weights or processing data. Both NSC and ResNSC follow this information-channel approach and use the resulting capacity measure to guide architectural design before training.

2. Technical advances of ResNSC over NSC

NSC provides an initial realization of this idea by treating linear projections in a Transformer as independent, parallel channels for capacity assessment: each projection is evaluated separately, and their capacities are summed. Its local score depends on the projection's dimensions and initialization variance, independently of its position in the network. Consequently, compatible reorderings of the same projection configurations receive identical NSC scores.

ResNSC advances this approach toward a more faithful representation of architecture by considering interactions among linear projections mediated by residual connections. Its technical contributions are twofold.

  • Residual-mediated interactions from specifications alone. ResNSC derives a coordinate-wise statistical reference for the shared residual stream from preceding branch updates, using it to define read-input and write-noise covariances (Sec. 3.2, Eqs. (2)–(3)). A prescribed within-branch propagation model also determines write-input covariances (Sec. 3.3, Eq. (4)). These contextual references enter the common capacity functional in Eq. (1); deterministic spectral formulas approximate the expected capacities directly from architectural specifications, without sampling weights or processing data (Sec. 3.3; Appendix D).

  • Capacity-guided allocation with residual history and optimality bounds. Residual-mediated interactions make a layer's capacity contribution depend on the context established by earlier layers. When allocation choices alter this context, as in joint residual-width allocation, NSC-DP's layer-and-budget state no longer suffices to determine future scores. ResNSC-DP augments its state with counts of preceding blocks at each residual width, sufficient under our reference to reconstruct future scoring contexts (Sec. 3.4, Eq. (7)). Feasible complete designs supply lower bounds, while history-indexed backward DP supplies upper bounds; dominance and bound pruning guide the search (Algorithm 1). Under the stated assumptions, closing the gap certifies optimality for the numerical ResNSC objective within the specified finite design space (Eq. (8); Appendix E).

3. Practical benefits of ResNSC over NSC

Residual context allows the same width increase to yield different gains in the capacity score at different layers. ResNSC uses these differences to guide how a resource budget is distributed across layers.

  • Developing Transformers with heterogeneous layer widths (Sec. 4.1, Tables 2–3). Residual context lets ResNSC assess a layer's width together with how its residual-width choice changes later capacity contributions, guiding the joint allocation of residual and FFN widths. Under the same search menus, parameter ceilings, and training protocols as NSC, its allocations reduce mean FineWeb-Edu perplexity from 20.00 to 19.39 at 121M and 13.82 to 13.60 at 374M. Directly testing the role of residual context, removing it while retaining within-branch propagation recovers NSC's allocation and raises 121M perplexity from 19.39 to 20.00 (Table 3; Appendix F.1).

  • Pruning Transformers (Sec. 4.2, Table 4). Even with fixed residual width, depth changes the residual context, so the capacity gain from retaining FFN width can differ across layers. These contextual scores guide where to retain parameters using the same layer-and-budget DP as NSC. With weights inherited from the same LoNAS supernet, identical channel slicing and search menus, and matched parameter budgets, ResNSC raises mean accuracy over eight tasks from 65.47% to 68.87% at 5.69B parameters, with gains of 3.28–5.46 percentage points across eight budgets (Fig. 3). This comparison supports the benefit of residual context for allocation with the search procedure held fixed.

AuthorsOpenReview ↗

Response to Reviewer LGeP (2/3): Q2

Q2: Fairness of the comparison in Table 4 and the role of residual context

Fairness. Yes. The LoNAS experiments of ResNSC and NSC are evaluated in exactly the same setting. Both methods select subnetworks from the same LoNAS supernet with the same channel slicing, the same FFN-width and rank menus, and the same eight parameter ceilings. The selected subnetworks have matching parameter counts at each ceiling, inherit their weights from the same supernet, and are evaluated with the same protocol (Sec. 4.2; Appendix H).

Under this identical setting, the gain of ResNSC over NSC arises from the different structures selected by the two scores. Our results show that at the highest budget (5.69B parameters), the selected FFN-width profiles differ markedly (Appendix H, Table 17):

MethodFFN widths from the first to the last layer (width ×\times number of consecutive layers)Avg$_8$
NSC9632×2;→;8256×309632\times2\\;\to\\;8256\times3065.47%
ResNSC11008×9;→;9632×4;→;8256×5;→;6880×8;→;5504×611008\times9\\;\to\\;9632\times4\\;\to\\;8256\times5\\;\to\\;6880\times8\\;\to\\;5504\times668.87%
ReverseThe ResNSC widths in reverse layer order59.49%

NSC allocates width almost uniformly, whereas ResNSC retains more FFN width in early layers and less in later ones. With matching parameter counts, Avg$_8$ differs by 3.40 points; the selected structures differ at all eight budgets, with gains of 3.28–5.46 points (Fig. 3).

The difference between the structures selected by the two methods originates from the residual context introduced by ResNSC. With fixed residual width, NSC assigns an FFN of a given width the same score at every layer, so any reordering of the same widths leaves the NSC score unchanged. In ResNSC, depth changes the residual context even at fixed residual width: a later branch reads from and writes into a residual stream that already carries the updates of all preceding branches (Sec. 3.2, Eqs. (2)–(3)). The capacity gain from retaining an FFN channel therefore varies across layers, and ResNSC-DP allocates width accordingly, yielding the profile above.

The Reverse control in Table 4 confirms that this placement is the more effective one. Assigning the ResNSC widths in reverse layer order preserves the parameter count, FLOPs, and NSC score, yet lowers Avg$_8$ from 68.87% to 59.49%: the two placements receive the same NSC score and differ by 9.38 points in accuracy.

AuthorsOpenReview ↗

Response to Reviewer LGeP (3/3): Q3 & Q4

Q3: Computational overhead of residual context

Yes, residual context adds computation, in the two places described in our response to Q1:

  • Contextual scoring. NSC scores a projection from its dimensions and initialization variance alone; ResNSC scores it under a reference that also depends on its residual context. These scores are still given by deterministic spectral formulas evaluated from architectural specifications, without sampling weights or processing data (Sec. 3.3; Appendix D).
  • Residual history in the search. When the residual width varies across layers, the search state must also record the residual history. ResNSC-DP summarizes this history as counts of preceding blocks at each residual width (Sec. 3.4, Eq. (7)), reuses scores across identical contexts, and prunes with dominance and bounds (Algorithm 1).

We measured the complete joint search on the 121M and 374M search spaces of Sec. 4.1, on a single thread of an Intel Xeon Gold 6133, with one run per scale:

Stage121M374M
Complete search161.13 s (2.69 min)1160.30 s (19.34 min)
Enumerating residual histories6.88 s134.14 s
Score precomputation137.10 s971.74 s
Upper-bound precomputation9.87 s46.88 s
Final search5.49 s5.70 s

The total also includes program start-up. Score precomputation accounts for 84%–85% of the time, and the final search takes under 6 s at both scales. Both runs return the configurations used in the paper with zero optimality gap (Appendix E.6, Table 9). This is a one-time CPU cost.

Q4: Search cost of the two methods in Table 4 (LoNAS pruning experiment)

Table 4 reports the LoNAS search, in which the residual width is fixed at 4,096 in all 32 blocks. The residual context of a layer then depends only on its depth, and not on the FFN widths or ranks selected elsewhere. ResNSC therefore uses the same layer-and-budget DP as NSC (Sec. 4.2), and the only additional work is evaluating the contextual scores. Both searches finish within 1 s, which is why both entries in Table 4 read "< 1 s". To show the difference between the two search times in finer detail, we report the timing of each method below.

We timed the two methods under one implementation: the same machine, a single CPU core, the same menus and eight budgets, and the median over seven fresh processes per method. "First call" is timed from the first use of the spectral evaluator in a fresh process; "after initialization" recomputes all scores and repeats the search within the same process.

StageNSC-DPResNSC-DP
First call: evaluator initialization and score-table construction269.02 ms292.82 ms
First call, total272.42 ms295.96 ms
After initialization: score-table construction1.37 ms14.67 ms
After initialization, total4.41 ms17.80 ms

The first call is dominated by loading and initializing the spectral evaluator, which is common to both methods. Once it is initialized, the difference lies entirely in scoring: building the score table takes 1.37 ms for NSC and 14.67 ms for ResNSC, while the allocation across the eight budgets takes about 3 ms for both. The additional cost of residual context in this setting is about 13 ms. A difference at the 0.01 s level leaves the search cost of the two methods essentially the same, so Table 4 lists both as "< 1 s". We thank the reviewer for raising this point: in the revision, Table 4 will report the measured search time of each method in place of "< 1 s", so that the difference is directly visible.

Comments on the paper

1 comment
Area_Chair_hPUVOpenReview ↗

Reminder to reviwers to read author's responses

Dear Reviewers,

Thank you for your efforts in providing thorough reviews.

As the discussion period is underway, I would like to kindly remind you to read the authors' responses to your reviews and provide follow-up comments where appropriate or leave a comment. In particular, please consider whether the responses address your questions and concerns, and update your assessment if the clarifications or additional information affect your initial evaluation.

Your engagement in this discussion is important for ensuring a fair and thorough review process.

Thank you again for your time and contributions!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.