acceptodds
Under review as a conference paper at ICLR 2027

When Sparse Routing Looks Additive: Hidden Pairwise Structure Changes Route Choices

Abstract

An additive fit to mean sparse-route loss need not preserve token-local effects or valid route choices. We audit temporal preservation and choice validity separately using complete four-band intervention surfaces and exact functional ANOVA. Beyond guaranteed energy attenuation under averaging, pairwise effects retain less normalized energy than unary effects on average. In a fresh 40-source, two-model study, the mean within-surface pairwise share of nonadditive gold-NLL energy is 86.72%; retaining pairwise terms reduces pooled mean hindsight regret versus unary by 77.5%, equally weighting separately evaluated NLL/KL endpoints. On 160 further documents, degree-two projection fitted to the same 70 feasible native Quest routes as unary reduces mean native-reference KL hindsight regret by 59.9% under fixed Llama/H100 execution. A prespecified support-construction intervention on 80 new documents narrows the retention gap at fixed K and mean overlap. With per-document calibration, two separate 48-document eight-band studies confirm held-configuration prediction at cohort-mean and document levels, including 28.69% lower mean document NLL RMSE versus unary in the prespecified Llama-interleaved condition. Better prediction can still yield a worse route. These findings motivate evaluating approximation-induced choices directly and reporting support construction when comparing preservation gaps.

Primary area:
foundation or frontier models, including LLMs
Keywords:
sparse attention, large language models, pairwise interactions, functional ANOVA, temporal aggregation, route selection, surrogate evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RNxKOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies whether additive models capture interactions among sparse routing choices. Across several settings, pairwise models improve hindsight route selection over unary models, though the results do not establish better deployment efficiency or generation quality.

Critical strengths

● The paper addresses a relevant question for sparse attention and separates temporal aggregation, interaction structure, choice regret, and held-configuration prediction rather than relying on a single fit metric. ● The estimands are clearly defined. The paper distinguishes functional ANOVA on the full Cartesian space from fitting on a constrained feasible set, and explains the scope of its uniform measure and hindsight regret. ● The experiments cover multiple models, datasets, support constructions, and target horizons. Some studies use separate cohorts, and the paper reports both errors corrected and errors introduced by pairwise terms.

Critical weaknesses

● Functional ANOVA is an established tool, and the main contribution appears to be its application to specific routing surfaces and the empirical value of pairwise terms. The paper should clarify what new insight it adds beyond prior interaction analyses and routing interventions, and which findings may generalize beyond the tested settings. ● The main hindsight results fit and select routes after observing a complete or large candidate set, so they do not show how to find better routes with less computation. The prospective calibration study improves KL fidelity relative to an anchor, but does not establish superiority over direct search or confirm gains in NLL or generation quality.

Questions for authors

● When the full routing surface is unavailable at deployment time, which results in this paper can concretely guide candidate-route acquisition or screening? Does the prospective study support a specific candidate-sampling strategy? ● Could the main text include a summary table listing each confirmatory claim, its estimand, independent statistical unit, cohort size, absolute effect, and confidence interval?

1 comment
AuthorsOpenReview ↗

Response to Reviewer RNxK

Contribution and generalization

We use established fANOVA as a measurement tool. The new 128-document held-choice test shows that removing a predicted interaction raises regret while preserving the local mean and average main effects. The temporal/support studies measure order-dependent retention and support-recipe effects; Appendix V separates retention, Unary-relative gain and absolute selected loss.

Domain/checkpoint tests cover NLP-paper and Python-repository text and Qwen3-32B (Section 4.6, Table 6), alongside the native Llama/Qwen panels (Section 4.2). These extensions concern continuation losses and route-choice regret in the tested settings.

Candidate acquisition or screening when the full surface is unavailable

Your question calls for guidance that specifies what information a selector actually has. The concrete guidance is to fit on paid observations, validate the actual recommendations on withheld candidates, compare strong Direct and Static controls, and count calibration and controller costs. A policy should retain its best measured candidate under the same objective.

The held-choice result uses 36 fitting labels. The separate 1,280-document study tests a fixed 8+5 acquisition rule; its gain intervals against Direct and matched batch Unary include zero. Appendix Q shows an anchor-relative KL benefit for its fixed one-guide-token acquisition and four-token selection policy; it fits no P2 model and leaves the Direct comparison unresolved. These studies provide evidence for evaluating screening decisions; a generally superior, inexpensive candidate-sampling rule remains open.

Confirmatory claim summary

Table 1 maps the principal claims to their estimands, independent units, cohort sizes, effects and individual 95% intervals. We have now added the three absolute NLL RMSE reductions below to its C1/C2 rows; the original relative intervals and confirmation decisions remain unchanged.

Prediction estimand / controlIndependent unit; NAbsolute reduction [95% CI]
C1 mean-surface RMSE: Unary-P248 documents0.024013 [0.009527, 0.035686]
C2 mean document RMSE: Unary-P248 documents0.031323 [0.022288, 0.041624]
C2 mean document RMSE: shared prior-P2Same 48 C2 documents0.030372 [0.022296, 0.039695]

These are added post hoc descriptive intervals from the original 20,000 paired, corpus-stratified crossed document/action bootstrap draws, with frozen fits and prior. C1 and C2 use separate cohorts (24 documents per corpus) in the prespecified Llama-interleaved NLL condition. C1 averages document residuals before RMSE; C2 calculates RMSE within each document before averaging. We reproduced the original relative intervals and independently checked the absolute intervals by direct index resampling (Apps. K/L).

Table 1 remains a principal-claim map. The study appendices retain the remaining prespecified comparisons and their original decisions, including unmet criteria. Appendix Y, Tables 74-76, gives all five held-quartet contrasts and document-paired inference; Table 6 gives domain/checkpoint scope. Appendix B records the study populations. Shared documents, repeated checkpoints, quartets and controls are not counted as additional independent observations; opened readbacks and secondary analyses retain their stated status.

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

This paper studies whether an additive approximation to average sparse-route loss preserves token-level interactions and route-selection decisions. It constructs complete intervention surfaces over layer bands and uses functional ANOVA to separate unary and higher-order effects. The experiments distinguish the attenuation guaranteed by averaging from empirical differences in normalized retention across interaction orders. They also compare unary and degree-two route choices, examine support-construction effects, and evaluate prediction of unseen eight-band configurations with per-document calibration. Results across several checkpoints and text domains show substantial pairwise structure and lower average hindsight regret, while also demonstrating that improved prediction does not necessarily produce a better selected route.

Critical strengths
  • S1. The paper makes a useful distinction between approximation quality, temporal preservation, and decision validity. In particular, it does not present Jensen-guaranteed energy attenuation as an empirical discovery, or infer an internal mechanism from endpoint interactions. The complete-surface audit provides concrete evidence that high unary fit can coexist with incorrect route choices. The separate 160-document native study also addresses the fitting-set mismatch identified in the earlier Quest comparison by fitting both degrees on the same 70 feasible actions.

  • S2. The empirical analysis is unusually careful about the scope of its conclusions. It uses source-level uncertainty, separates exploratory analyses from newly specified tests, and retains unfavorable outcomes and numerical sensitivities. The fresh native study reports a 59.9% reduction in mean hindsight regret, while explicitly showing that choices worsen on 41 of 160 documents. The second prediction cohort tests document-level error rather than relying only on an averaged surface, and the support-construction study controls per-layer K and mean overlap. These analyses make the central evaluation lesson more convincing than a simple improvement in aggregate fit would.

Critical weaknesses
  • W1. The main reservation is the significance and scope of the new empirical knowledge. Interaction-aware analyses already exist, including [ProxySPEX (NeurIPS 2025)], which studies interactions within and across attention layers. This does not subsume the present paper's order-dependent temporal retention, support-construction controls, or route-choice audit. Those are useful additions. Nevertheless, the strongest decision comparisons observe the candidate losses before fitting and choosing, and the eight-band predictions require 129 fitting observations per new document, with 32 further observations for the full tuning comparison. Direct search wins at the tested budgets in Appendices D.3 and D.4. Appendix Q adds a prospective test but fits no pairwise surrogate and does not establish superiority to direct search. These findings support a diagnostic lesson more strongly than a general account of when pairwise modeling changes routing decisions beneficially under limited information. A deployable algorithm is not required, but I find the demonstrated incremental significance borderline for acceptance.

  • W2. The strongest native result remains conditional in both magnitude and execution scope. The fresh SAME70 study reports a 59.9% reduction in mean hindsight regret, but exact optimal choices increase from 32 to 35 of 160 documents; 93 choices improve, 41 worsen, and 26 tie. This is compatible with meaningful loss reductions, yet the headline percentage should not be interpreted as a comparable improvement in task quality or optimal-route frequency. The study fixes one Llama checkpoint and H100 profile. In the eight-document execution pilot, 7 of 8 unary choices, 6 of 8 pairwise choices, and 5 of 8 true optima change across GB10 and H100. The paired interval does not establish a hardware difference in mean gain, and the broader model/domain studies remain valid supporting evidence. However, those endpoint extensions do not establish the same breadth for the native decision conclusion; fixed-support replay would help distinguish intervention changes from loss-surface sensitivity.

Questions for authors
  • Q1. Can the authors separate support-selection sensitivity from loss-surface sensitivity by replaying fixed physical supports across the two qualified execution profiles on the existing pilot documents? The key issue is whether the relative unary/pairwise choice advantage persists once the intervention itself is held fixed.

  • Q2. For the fresh SAME70 native study, can the authors provide the distribution of absolute and range-normalized regret gains, including paired intervals for near-optimal selection at several explicitly descriptive tolerances? The 59.9% mean reduction, only three additional exact hits, and the concentration of positive gain in the largest contributor make this useful for assessing the effect’s practical magnitude.

  • Q3. Can the authors clarify which conclusions are intended to motivate explicit pairwise modeling, as opposed to directly testing the choices produced by any approximation? In particular, is there a measured acquisition regime in which a pairwise procedure is competitive with direct search, after charging all calibration observations? The existing negative comparisons should remain part of that assessment.

3 comments
AuthorsOpenReview ↗

Response to Reviewer geFw

W1: What the additional findings change

ProxySPEX already uses interactions for subset optimization and head pruning. Our new test asks whether predicted interactions improve held-out choices.

We collected 40 routes per document on 128 new Llama-3.1-8B-Instruct documents (64 per corpus); all change 15 layers. Each of 270 quartets withholds all four candidate losses and fits on the other 36 routes. Comparators are Unary and 256 fixed rank-matched controls retaining the Unary space.

Unary-minus-P2 mean within-quartet regret is 40.135 [6.281, 76.902] micro-KL. Removing only P2's predicted local interaction, preserving its mean and average main effects, raises regret by 22.260 [13.885, 31.572]. All five prespecified individual 95% lower bounds are positive. Rank-control contrasts average separately evaluated control metrics; quartets aggregate within document before paired, corpus-stratified inference (N=128). Against Unary, 70 documents improve and 58 worsen.

This identifies component value after 36-label calibration, not cheaper acquisition or a neural mechanism (Sec. 4.3, App. Y). The support study separately shows increased pairwise retention can coexist with worse selected loss (Sec. 4.5, App. V).

Q1: Physical-support replay across execution profiles

We crossed support origin, execution profile and KL reference on all eight original documents and 70 routes. Foreign supports transfer indices with coherent destination-state recomputation; own-support replays match dynamic execution exactly.

Changing one factor while fixing the other two gives mean source Spearman ranges of 0.266-0.379 (support), 0.903-0.950 (execution), and 0.367-0.459 (reference); optimal routes change on 7-8/8, 2-5/8 and 7-8/8 documents, respectively. With support/reference fixed, retaining the previous optimum costs only 0.41-1.26% of destination loss range on average. Normalization occurs within document before averaging both transfer directions.

P2's mean regret is lower in all eight cells, but three nominal paired intervals include zero; five are positive. Cells reuse eight sources, and historical numerical settings prevent attributing the contrast to hardware alone (Sec. 4.2, App. T).

Q2: Absolute and normalized gain distributions

All 160 documents and original choices remain. Mean/median raw gains are 208.211/36.938 micro-KL; 93/41/26 improve/worsen/tie. Mean document-wise range-normalized regret falls from 0.14240 to 0.07215, a paired gain of 0.07025 [0.04655, 0.09509].

At prelisted exact/1%/5%/10% loss-range tolerances, near-optimal rate gains are 1.88/6.25/16.88/15.63 percentage points. Exact and 1% intervals include zero; 5%/10% intervals are positive. These opened intervals are nominal, document-paired and corpus-stratified, not task-accuracy estimates.

The largest contributor remains included: 28.33% of positive gains, not net gain. All equal-corpus leave-one-document-out point means remain positive; they are not confidence bounds. Figure 2/App. S retain raw percentiles, all absolute tolerances and every paired interval.

W2: Native choice beyond one checkpoint

A disjoint 256-document Llama panel and Qwen3-8B evaluation use the same documents, with model-specific native references. Full-information Unary-minus-P2 gains are 150.320 [89.504, 217.572] micro-KL for Llama (prespecified secondary) and 539.130 [383.680, 694.581] for Qwen (primary).

Fixed route 00001111 remains strong: P2's Llama advantage is unresolved. Qwen Static-minus-P2 regret is -103.611 [-204.238, -2.984], favoring Static. Thus these extend Unary-relative evidence, not best-policy superiority.

The models share 256 documents, not 512. Windows/interventions differ; architecture is not isolated. Qwen's fixed-sample scope amendment preceded inspection of main outcomes; its planning limitation remains disclosed (Sec. 4.2, Apps. U/X).

Q3: Explicit pairwise modeling and acquisition cost

On 1,280 new documents, P2 and Unary see the same eight paid Direct labels, fit once and freeze five nominations before further labels are revealed. All methods pay 13 labels and return their best actual observation. These are teacher-forced KL searches, not input-only routing or deployment-latency tests.

Direct-minus-P2 returned loss is 2.562 [-0.239, 5.363] micro-KL; matched batch Unary-minus-P2 is 0.068 [-6.088, 6.224]. Neither establishes superiority; zero-spanning intervals do not establish equivalence. App. W reports all controls and measured costs.

Direct can exploit interactions through joint observations without a quadratic model. Its earlier wins and the calibration study's failed quality criteria remain reported. App. Q fits no P2 model. The held-out component test establishes calibrated information value; the fixed-budget study evaluates one acquisition policy.

Thank you for including these additional experiments, and they are valuable additions.

A primary concern remains regarding the practical significance of these findings for routing design. Specifically, it is still unsure when explicit pairwise modeling is worth the calibration effort, or what general design guidance follows from the observed interactions. In addition, it is recommended to improve the overall readability of the paper, especially the abstract and introduction.

AuthorsOpenReview ↗

Practical Implications for Sparse-Routing Design

Thank you for your follow-up and for recognizing the value of the additional experiments.

We agree that identifying precise low-budget regimes in which pairwise surrogates outperform direct search or strong static routes remains an important open question for online router design. Our results do not establish such a break-even regime, particularly because the held-out choice and budgeted-search experiments evaluate different candidate sets.

Nevertheless, we believe our findings provide concrete guidance for sparse-routing design and evaluation.

First, average approximation quality (e.g., R2R^2 or RMSE) should not be treated as a proxy for route-selection validity. Even when additive fits explain most of the observed loss variation, local pairwise interactions can still affect which route is selected. Our held-out choice experiment further isolates a contribution from the predicted interaction terms by removing the local interaction while preserving the average main effects (Section 4.3).

Second, the appropriate use of pairwise modeling depends on whether measurements are already available or must be acquired for the routing decision. Thus, our results suggest a conditional design principle: use pairwise terms to evaluate and potentially improve route choices when calibration measurements are available, but require a matched-budget comparison against direct search before adopting pairwise calibration as an online acquisition strategy (Section 5).

To address your recommendation on readability, we have substantially revised the abstract and introduction to distinguish these two questions from the outset: the decision value of interaction information given existing measurements, and the acquisition cost of obtaining that information. This framing makes the resulting evaluation and design guidance explicit without presenting pairwise calibration as a universally superior online routing strategy.

Reviewer_d7K7OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors study what happens when several routing decisions change together and whether their effects are additive in nature. They show that on average, unary effects have less normalized energy than pairwise effects with the latter still carrying useful signal. They also highlight the effect of token-averaging. Because the paper is very very poorly written (see my first point in the weaknesses), I am not exactly sure this summary captures what the authors intended for the paper to convey and they should work more on the clarity of the presentation of this work.

Critical strengths
  • The paper covers several hypotheses and scenarios, and presents what seems to be a thorough study.
  • The four-band intervention surfaces allow for more rigor and isolation of the effects of the temporal preservation and the choice separately.
  • The simplicity of the sparse routing model is an advantage here.
  • The analysis tools, especially functional ANOVA are interesting.
Critical weaknesses
  • My main issue with this paper is that I'm lost from the very beginning: the abstract. It does not introduce the context properly, the sentences are almost disconnected and the jargon is too heavy; whereas paper writing is mainly about story telling of one's findings. This issue persists throughout the paper that has many sentences that seemingly try to convey something but are so indirect and hard to comprehend that they only make me understand even less the message I thought I understood. I read sentences 4-5 times without understanding their meaning. This paper needs substantial clarity work done!
  • The pairwise effects are shown to carry useful information but we don't have a mechanistic understanding of how they arise.
  • It is unclear to me how the findings translate to sparse attention in large language models or the implications of this work.
  • I think the paper should do a better job in the analysis at changing one hypothesis at a time, right now the analysis is too complicated to follow and despite my best effort, very hard to comprehend and separate the useful findings from complicated language describing them.
Questions for authors
  • Can you summarize for me the findings of this paper in your own words and breakdown the different tested hypotheses and their outcomes? that will help me better assist the paper during the rebuttal. Please do not use an LLM to do so, I would like a simplified version of the abstract and conclusion in your own words.
  • What are the implications of your work on sparse attention LLMs?
  • "Neither the development cohort nor the earlier 160-document KL study’s failed primary is relabeled or rescued." -> which failed primary are you referring to?
1 comment
AuthorsOpenReview ↗

Response to Reviewer d7K7

Thank you for explaining where the paper became difficult to follow.

W1/W4: Presentation and separation of hypotheses

We rewrote the abstract and introduction around the experimental questions and defined routing variables before decomposition (Section 2). Table 1 maps principal claims to estimands, independent units and intervals. Section 4 separates preservation, choice and held-out studies; Section 5 covers the fixed-budget test.

QFA1: Findings and hypotheses

This paper grew out of my work on reducing how much context attention has to read. Early experiments suggested that a relatively simple model could describe route loss averaged over a continuation. But using that description to predict useful routing decisions was still difficult. That led me to test what averaging was hiding, and whether a good average fit preserved the choices we wanted to make.

An interaction can help at one token and hurt at another, so it can look small after averaging even when it matters locally. In our tested surfaces, pairwise effects lost more of their own energy under averaging than unary effects did, on average. Retention compares each component with its own token-level energy. The pairwise share is a different comparison: pairwise terms accounted for most of the nonadditive NLL energy in the natural-text study. Changing the support construction changed the retention gap even at the same reading budget and mean overlap.

The choice question is different. If two routes have similar losses, a small fitting error can reverse their ranking. Pairwise terms can represent how the value of one layer group's choice depends on another's. The native 160-document comparison used all 70 measured route losses. I consider the new 128-document test particularly useful because we fitted on 36 other routes and withheld all four candidate losses. Removing only the predicted local interaction increased mean within-quartet regret, with the mean and average main effects held fixed.

I also wanted to know whether we could use that structure cheaply. Direct search observes joint outcomes too. A pairwise predictor fitted from only a few observations has to suggest better candidates than Direct at the same total budget. Our batch policy did not establish that advantage. The test does not isolate whether the available observations, the model or the policy design limited it. I therefore see efficient acquisition as an open question.

I also genuinely appreciate that you asked this question. I took it as a sign that you were interested not only in the paper's results, but also in the thinking behind them and the person who carried out the work. That means a lot to me as an author.

It was surprisingly difficult to explain the work without leaning on its terminology and equations. Perhaps that is a useful exercise in itself: if I cannot explain in plain language why I pursued these experiments, perhaps I still have more thinking to do. I hope this gives you a better sense of how I see the work.

W3 / QFA2: Implications for sparse-attention LLMs

Evaluate the selected route's actual loss on the intended feasible set, alongside fit error and cases where adding pairwise terms worsens a choice. Token-local effects cannot be inferred from sequence averages alone. Matched reading budgets and mean overlap do not make support recipes equivalent for retention. Report support construction and charge all calibration observations. These findings establish neither deployment speedup nor a generally superior acquisition policy (Sections 2-5).

W2: What the analysis does and does not explain

The new test identifies a predicted interaction component's contribution to choices, not its internal neural origin. Section 4.1 explains how small errors can change a choice. Appendix C.2.1 tests additive logits followed by nonlinear loss. Directly optimizing this control nearly recovers Qwen's mean KL regret improvement over Unary on the earlier 16-source panel; outcomes differ for other models/endpoints (Section 5).

QFA3: Which earlier failed primary?

This referred to a separate 160-document calibration-policy experiment. Its joint primary required positive 95% lower bounds for P2's future 32-token mean dense-reference KL improvement over three controls: forward Direct at budget 11, reverse Direct at budget 11, and a prompt-mass anchor. It also required no higher measured calibration cost than either Direct control. Both cost conditions passed, but all three quality intervals included zero, so quality superiority was not established. This was distinct from the successful native Quest SAME70-160 study. Appendix Q (p. 61) also describes the later 80-document fixed-policy study, which fits no P2 model.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Revision Summary

Thank you for the constructive feedback. We reorganized the manuscript around the routing question, introduced routing variables before the decomposition, and added Table 1 to map the principal claims to their estimands, independent units and intervals. Table 1 now also reports three absolute C1/C2 RMSE reductions with post hoc descriptive 95% intervals; the original relative intervals and confirmation decisions are unchanged.

  • Predicted interactions contribute to held-out choices (Section 4.3, Appendix Y; geFw W1, RNxK). On 128 new documents, each quartet's four candidate losses are withheld; predictors fit on 36 other routes and are compared with rank-matched controls. Removing only the predicted local interaction, preserving the mean and average main effects, increases mean within-quartet regret by 22.260 micro-KL (individual 95% interval: [13.885, 31.572]).

  • Native comparisons extend to new documents and a second checkpoint (Section 4.2; geFw Q1/Q2/W2). We add gain distributions, fixed physical-support replay, and Llama/Qwen evaluations on a shared 256-document panel. The replay separates support, execution and KL-reference effects.

  • Acquisition is tested at a fixed observation budget (Section 5, Appendix W; geFw Q3, RNxK). On 1,280 new documents, all methods pay for 13 observations. The pairwise policy's gain intervals against Direct and matched batch Unary include zero. These studies assess calibrated interaction information's choice contribution and the tested acquisition policy's performance separately.

The revised main text is 10 pages. Companion saved-data CPU replay packages cover the added analyses; our individual responses provide the detailed results and evidence locations.

Guide to the added evidence

The table locates the post-submission additions. It distinguishes new data collection from opened analyses of existing data; detailed results remain in the cited sections.

EvidenceData and statusLocation; related review questions
Interaction controls and held-out choicesR: opened equal-rank controls on existing surfaces. Y: 128 new documents; five prespecified paired comparisons, including local-interaction removal.Section 4.3, Appendices R/Y; geFw W1, RNxK.
Native gain distributions and physical-support replayS: opened readback of all original 160 documents. T: new crossed replay of eight original pilot sources and all 70 routes.Section 4.2, Figure 2, Appendices S/T; geFw Q1/Q2.
Native Llama and Qwen evaluations256 new documents shared across checkpoints. Llama replication is prespecified secondary; Qwen has its own primary and secondary contrasts.Section 4.2, Appendices U/X; geFw W2, RNxK.
Preservation and selected lossOpened readback of the existing 80-document support intervention, retaining both checkpoints and endpoints.Section 4.5, Appendix V; geFw W1.
Fixed-batch acquisition1,280 new documents; frozen policy, 13 paid observations per method.Section 5, Appendix W; geFw Q3, RNxK.

The revised exposition and Table 1 address clarity and claim mapping (d7K7, RNxK). Original confirmation decisions are retained. Companion R-X and Y packages replay the added analyses from saved data; they do not rerun model collection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.