acceptodds
Under review as a conference paper at ICLR 2027

GIGA: Grouped Sparse Updates for Large-Batch Recommendation Model Training

Abstract

Scaling the worker count at a fixed local batch size increases throughput but also coalesces sparse updates, which in turn often degrades downstream recommendation quality. In click-through rate (CTR) models, this effect is particularly frequency-sensitive: frequent IDs lose more update opportunities over a fixed sample budget than rare IDs, so a single sparse learning-rate multiplier cannot simultaneously correct head and tail IDs. To this end, we propose **Group-wise ID Gradient Aggregation (GIGA)**, which restores fine-grained control over update coalescing. Specifically, GIGA averages each active ID's gradients within worker groups, sums the group means for the update numerator, and separately accumulates their elementwise squares. Group size interpolates between active-worker averaging and retaining every active-worker contribution, while leaving single-worker IDs unchanged. GIGA requires neither historical frequency statistics, additional collectives, nor persistent per-ID state. We show that, for fixed local batch and group size, the expected number of retained group terms over a given sample budget is independent of worker count, and relate the fused update to sequential optimization. In a controlled 96-worker comparison, GIGA improves AUC and request-level GAUC by and percentage points, respectively. Combined with dense-parameter scaling, GIGA supports a 16-fold larger global batch of 307,200 examples on 512 workers and delivers sample throughput while matching or improving both metrics relative to the 32-worker baseline. Gains extend to a separate search CTR model under warm-start and from-scratch training.

Primary area:
optimization
Keywords:
Lagre-scale Training, Optimization, Recommender Systems, Ranking, Click-Through Rate

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fxMQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GIGA addresses frequency-dependent loss of sparse update opportunities as worker count and global batch grow. It forms active-support means within fixed worker groups, sums those means for the parameter update, and separately accumulates their element wise squares in Adagrad state. Group size controls retained granularity. Experiments cover two industrial CTR workloads, 32–512 worker scaling, matched aggregation baselines, multi-day stability, and warm-start/from-scratch settings.

Critical strengths
  1. The paper identifies a precise and practically important failure mode of large-batch sparse optimization. The method follows directly from the analysis and cleanly couples the update numerator with the Adagrad-state increment.
  2. The controlled experiments are strong. They isolate aggregation, compare plausible alternatives, ablate the state update, show multi-day and month-long behavior, generalize to a second workload, and report full-system scaling and overhead.
Critical weaknesses
  1. All efficacy evidence uses private industrial CTR workloads and Adagrad. It remains unclear whether the phenomenon and update generalize to public benchmarks or other adaptive optimizers.
  2. Group size is consequential and tuned separately by scale; the method lacks an automatic selection rule. The one-step approximation preserves state accumulation but not sequential parameter updates, and no convergence analysis is provided.
Questions for authors
  1. Can you provide a public-data or controlled synthetic reproduction?
  2. Can group size be selected from observed active-group statistics without quality sweeps?
  3. How does the method extend to row-wise Adagrad, Adam-family optimizers, or asynchronous updates?
9 comments

We sincerely appreciate the reviewer’s great efforts and insightful comments to improve our manuscript. Below, we address these concerns point by point and try our best to update the manuscript accordingly.


[Q1 / W1] Public-data or controlled synthetic reproduction

Can you provide a public-data or controlled synthetic reproduction?

Response: We train DeepFM from scratch on public Criteo-1TB with 32 workers and local batch 4096. With$G=1$ throughout, GIGA improves AUC over RBI across all three optimizers:

OptimizerRBI AUCGIGA AUCΔAUC (pt)
Adagrad0.77990.7819+0.20
Row-wise Adagrad0.78050.7822+0.17
AdamW0.78800.7886+0.06

AdamW values are means of five runs. Appendix G reports this public-data evaluation (Table G.1).

These results extend the evidence beyond private industrial workloads and standard Adagrad, with consistent gains across all three tested optimizers.

[Q2 / W2] Selecting G from active-group statistics

Can group size be selected from observed active-group statistics without quality sweeps?

Response: We provide a group-size selection rule based on sampled worker activity. Let$W$ be the number of workers and$G$ the number of workers per group, giving$W/G$ groups. Each record$i$ corresponds to an item ID in a sampled step. A worker is active if it contains the ID, and a group is active if at least one of its workers is active. Let$n_i$ and$a_i(G)$ be the numbers of active workers and active groups, respectively. We replay the same worker hits for every candidateG.

We define hot records as$\mathcal H={i:n_i\geq W/2}$ and cold records as$\mathcal C={i:1\leq n_i\leq3}.Let. Letq_{99}(G)$ be the 99th percentile of$a_i(G)$ over$\mathcal H$. The three costs are:

Phot(G)=q99(G)q99(1)P_{\mathrm{hot}}(G)=\frac{q_{99}(G)}{q_{99}(1)}

Shot(G)=1∣H∣∑i∈H1[ai(G)=W/G]S_{\mathrm{hot}}(G)=\frac{1}{|\mathcal H|}\sum_{i\in\mathcal H}\mathbf{1}[a_i(G)=W/G]

Lcold(G)=1−∑i∈Cai(G)∑i∈CniL_{\mathrm{cold}}(G)=1-\frac{\sum_{i\in\mathcal C}a_i(G)}{\sum_{i\in\mathcal C}n_i}

Each active group supplies one virtual update, contributing its mean gradient to the update numerator and its squared mean to the Adagrad state. The costs capture how grouping balances compensation for lost head-ID updates with preservation of contribution granularity:

  • Lower$P_{\mathrm{hot}}$ moderates head-ID update amplification. For equal, aligned group gradients, retaining more contributions increases the fused Adagrad step even with separate-square state accumulation, so retaining every worker contribution can overcompensate for lost updates.
  • Lower$S_{\mathrm{hot}}$ avoids widespread saturation of head-ID contribution counts. Once an ID occupies all groups, additional worker hits are averaged into existing group means instead of supplying separate virtual updates, reintroducing the coalescing that GIGA seeks to alleviate.
  • Lower$L_{\mathrm{cold}}$ preserves tail-ID contribution granularity. Rare IDs lose few update opportunities when the global batch grows, so merging their already sparse contributions introduces coalescing where little compensation is needed. Single-worker IDs remain unchanged for everyG.

IncreasingG moderates head-ID amplification but merges more contributions within groups. The score balances these effects while preserving tail-ID contributions.

We select the group size by minimizing the worst of the three costs. For each candidate group size GG, we compute

\mathrm{Score}(G) = \max\left\\{P_{\mathrm{hot}}(G),\\ S_{\mathrm{hot}}(G),\\ L_{\mathrm{cold}}(G)\right\\}.

We select the GG in the candidate set G\mathcal{G} with the lowest score, breaking ties in favor of the smaller group size. The tables below use the same notation and report AUC and GAUC differences in percentage points relative to the 32-worker reference.

128 workers:

GPhotP_{\mathrm{hot}}ShotS_{\mathrm{hot}}LcoldL_{\mathrm{cold}}Score ↓ΔAUC / ΔGAUC (pt)
1100%1.79%0.00%100.00%+0.13 / −0.01
250%10.35%0.23%50.00%+0.05 / +0.06
425%56.84%0.69%56.84%+0.04 / +0.05

256 workers:

GPhotP_{\mathrm{hot}}ShotS_{\mathrm{hot}}LcoldL_{\mathrm{cold}}Score ↓ΔAUC / ΔGAUC (pt)
1100%1.34%0.00%100.00%+0.14 / −0.04
250%7.89%0.12%50.00%+0.15 / −0.02
425%41.67%0.37%41.67%+0.05 / +0.03
812.5%97.48%0.86%97.48%+0.04 / +0.03

The rule selectsG=2 at 128 workers andG=4 at 256 workers using activity statistics alone. Both recover positive AUC and GAUC gains, whereasG=1 leaves GAUC below the reference at both scales. Further increasingG raises hot-ID saturation and yields no additional quality gain. These results support a practical, mechanism-based choice ofG: moderate head-ID compensation without excessive coalescing, preserve tail-ID contributions, and recover both ranking metrics without using quality sweeps to selectG.

[Q3 / W1] Row-wise Adagrad, Adam, and asynchronous updates

How does the method extend to row-wise Adagrad, Adam-family optimizers, or asynchronous updates?

Response: The Criteo-1TB experiments in Q1 show AUC gains of 0.17 pt with row-wise Adagrad and 0.06 pt with AdamW. We also evaluate warm-start incremental training on industrial CTR data with 32 workers, local batch 600, and GIGA G=1G=1. Gains over the corresponding RBI runs are reported in percentage points:

OptimizerΔAUC (pt)ΔGAUC (pt)
AdamW+0.002+0.048
Row-wise Adagrad+0.006+0.041

AdamW warm-starts model parameters and initializes optimizer states from scratch, with results measured after three days of incremental training. Row-wise Adagrad initializes its accumulator by averaging the existing Adagrad state within each row.

These results extend the optimizer evidence to warm-start industrial training, with gains in both AUC and GAUC. Appendix F defines the optimizer extensions below.

Row-wise Adagrad. For a mean-square scalar accumulator, GIGA gives

\qquad e_j^{+} = e_j - \eta\frac{\sum_{k\in\mathcal{A}_j}\bar{g}_{j,k}}{\sqrt{h_j^{+}}+\epsilon}.$$ Here $e_j$ is the embedding row, $h_j$ is its scalar accumulator, $d$ is the embedding dimension, and $\mathcal{A}_j$ is the set of groups containing ID $j$. Within group $k$, reduce-by-ID averages the locally aggregated gradients $g_{j,w}$ over the workers $S_{j,k}$ that contain ID $j$, giving $\bar{g}_{j,k}=\frac{1}{|S_{j,k}|}\sum_{w\in S_{j,k}}g_{j,w}$. The resulting scalar accumulator equals the final accumulator obtained by sequentially accumulating the same prescribed group gradients (Appendix F.1). **Adam.** For a given row, let $S$ be its active workers and $\mathcal A$ its active groups. Using the group means defined above, RBI and GIGA construct their moment inputs as $$r=\frac{1}{|S|}\sum_{w\in S}g_w,\qquad M=\sum_{k\in\mathcal A}\bar g_k,\qquad Q=\sum_{k\in\mathcal A}\bar g_k^{\odot 2}.$$ RBI uses $(r,r^{\odot 2})$ as the first- and second-moment inputs. GIGA feeds $M$ and $Q$ to the first moment $m$ and second moment $v$, respectively: $$m^+=\beta_1m+(1-\beta_1)M,\qquad v^+=\beta_2v+(1-\beta_2)Q,$$ $$e^+=e-\eta\frac{m^+/(1-\beta_1^t)}{\sqrt{v^+/(1-\beta_2^t)}+\epsilon}.$$ Both methods advance $t$ once per physical table update and use the same normalization and sparse-row mask, updating present rows even when their gradients cancel and leaving inactive rows unchanged. One active group recovers the corresponding masked Adam update (Appendix F.2). **Asynchronous updates.** We outline a protocol with group-local synchronization and independent progress across groups. For a given row, let $z_b$ be the active-worker mean in completed group record $b$: | Step | Stage | Proposed rule | | --- | --- | --- | | 1 | **Group completion** | Each group submits its active-worker means after all its workers finish the batch. | | 2 | **Staleness check** | Before committing an update, check sparse-row age against $\tau$ local owner updates. Return expired groups for recomputation. | | 3 | **Owner update** | Combine accepted records into $M=\sum_b z_b$ and $Q=\sum_b z_b^{\odot 2}$, then apply one atomic GIGA update. Consume each record once. | Separate-square state accumulation remains exact for admitted gradients. Appendix F.3 describes this asynchronous protocol. # [W2] Sequential approximation and convergence > The one-step approximation preserves state accumulation but not sequential parameter updates, and no convergence analysis is provided. **Response:** **The new analysis establishes convergence of GIGA's actual fused iterates without assuming equivalence to sequential parameter updates.** Appendix A.4 proves the following rate for a smooth, lower-bounded objective, unbiased bounded worker gradients, stationary IID worker sampling, fixed grouping and parameter set, positive initial accumulators, and fixed learning rates: $$\frac{1}{T}\sum_{t=1}^{T}\mathbb E\|\nabla f(x_t)\|_2^2 =O\left(\frac{\log(1+T)}{\sqrt T}\right).$$ Gradients are recomputed at each iterate. The key inequality $M_{t,i}^2\le KQ_{t,i}$, where $K=W/G$, bounds the accumulated squared steps by a logarithmic term. The proof accounts for the activity-dependent mean of $M_t$ and the dependence between the current gradient and adaptive denominator. The sequential comparison establishes a separate property: identical state accumulation for prescribed group gradients.

Changes in the revised manuscript (highlighted in red)

  • Added a convergence theorem and proof in Appendix A.4 (Proposition 1, Equation A.3), summarized in Section 3.4.
  • Added an activity-based group-size selection rule and comparisons at 128 and 256 workers in Appendix C.2 (Tables C.4–C.5).
  • Added row-wise Adagrad updates in Appendix F.1.
  • Defined Adam moment updates, clock, and sparse-row masking in Appendix F.2.
  • Specified an asynchronous group-completion and owner-update protocol in Appendix F.3.
  • Added public Criteo-1TB results for Adagrad, row-wise Adagrad, and AdamW in Appendix G.

Thank you for your constructive feedback. We hope the additional experiments, convergence analysis, and clarifications address your concerns, and we would appreciate your reconsideration of the final score.

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

This paper proposes GIGA, a group-wise sparse gradient aggregation method, which is designed to address frequency-dependent update coalescing in large-batch recommendation training. GIGA aggregates gradients within worker groups while separately accumulating their squared contributions to the Adagrad optimizer state. Experiments on two industrial CTR datasets demonstrate improvements in training quality and scalability.

Critical strengths
  1. The paper identifies a practical limitation of conventional sparse gradient aggregation: frequent and infrequent IDs are affected differently by increasing global batch sizes.

  2. The experiments cover different worker counts, batch sizes, aggregation strategies, and group configurations, with additional multi-day evaluations. The demonstrated scaling to 512 workers with 9.8× throughput and maintained recommendation quality provides useful evidence of practical applicability.

Critical weaknesses
  1. method novelty not clear. Table C.5 shows that at 512 workers, GIGA, Reduce-Sum, and Capped RBI achieve identical reported AUC and GAUC improvements, raising questions about when the group-wise design is necessary.

  2. Not fully ablate GIGA's contribution to scalability and generalization. The reported 9.8× throughput improvement combines increased worker count, dense-parameter scaling, and GIGA, making it difficult to attribute the overall benefit specifically to the proposed aggregation method

Questions for authors

Q1. Necessity of group-wise aggregation. see w1 for details.

Q2. How well does GIGA extend to other adaptive optimizers, such as Adam, and to publicly reproducible recommendation workloads?

3 comments

We sincerely appreciate the reviewer’s great efforts and insightful comments to improve our manuscript. Below, we address these concerns point by point and try our best to update the manuscript accordingly.


[Q1 / W1] Benefit beyond simpler activity-dependent aggregation

Table C.5 shows that at 512 workers, GIGA, Reduce-Sum, and Capped RBI achieve identical reported AUC and GAUC improvements, raising questions about when the group-wise design is necessary.

Response: We repeated the 128/600 and 256/600 comparisons with three matched seeds. Values are mean ± sample standard deviation in pt relative to the 32-worker reference:

WorkersMethodΔAUC (pt)ΔGAUC (pt)
128GIGA, G=2+0.0454 ± 0.0056+0.0510 ± 0.0050
128Capped RBI, c=64+0.0389 ± 0.0049+0.0229 ± 0.0036
128Reduce-Sum+0.0358 ± 0.0059+0.0366 ± 0.0031
128GIGA, Q=M²+0.0381 ± 0.0024+0.0340 ± 0.0047
256GIGA, G=4+0.0523 ± 0.0015+0.0316 ± 0.0080
256Capped RBI, c=64+0.0417 ± 0.0043+0.0028 ± 0.0025
256Reduce-Sum+0.0452 ± 0.0041+0.0130 ± 0.0117
256GIGA, Q=M²+0.0427 ± 0.0027+0.0101 ± 0.0056

GIGA achieves the highest mean AUC and GAUC at both scales. Its separate-square state improves mean GAUC over Q=M² by 0.0170 and 0.0215 pt at 128 and 256 workers, respectively. At 512 workers, GIGA and the simpler controls match at the reported precision (Table C.7 in the revision, originally C.5), possibly because quality gains approach saturation in this setting.

[W2] Isolating GIGA's Contribution to Scaling

The reported 9.8× throughput improvement combines increased worker count, dense-parameter scaling, and GIGA.

Response: We isolate GIGA's quality contribution by replacing RBI while holding the remaining configuration fixed, including worker count, batch size, and dense settings:

Workers / local batchMatched comparisonAUC gain (pt)GAUC gain (pt)Reference
96 / 600RBI → GIGA G=1+0.03+0.13Figure C.2(a)
512 / 600RBI → GIGA G=8+0.12+0.34Table C.7

The 9.8× throughput measures the complete 32-to-512-worker scaling recipe, which achieves +0.05 pt AUC and matches baseline GAUC (Table 3). Thus, the matched comparisons establish GIGA's independent contribution to AUC and GAUC at fixed worker count, batch size, and dense settings, while the full recipe delivers the throughput gain.

[Q2] Other adaptive optimizers and public workloads

How well does GIGA extend to other adaptive optimizers, such as Adam, and to publicly reproducible recommendation workloads?

Response: We train DeepFM from scratch on public Criteo-1TB with 32 workers and local batch 4096. With G=1G=1, GIGA improves AUC over RBI with all three optimizers:

OptimizerRBI AUCGIGA AUCΔAUC (pt)
Adagrad0.77990.7819+0.20
Row-wise Adagrad0.78050.7822+0.17
AdamW0.78800.7886+0.06

AdamW values are means of five runs. Appendix G reports this public-data evaluation (Table G.1), and Appendix F.1–F.2 describes the row-wise Adagrad and Adam extensions.

We also evaluate warm-start incremental training on industrial CTR data with 32 workers, local batch 600, and GIGA G=1G=1. Gains over the corresponding RBI runs are reported in percentage points:

OptimizerΔAUC (pt)ΔGAUC (pt)
AdamW+0.002+0.048
Row-wise Adagrad+0.006+0.041

AdamW warm-starts model parameters and initializes optimizer states from scratch, with results measured after three days of incremental training. Row-wise Adagrad initializes its accumulator by averaging the existing Adagrad state within each row.

Together, the public-data and industrial results support GIGA's applicability beyond standard Adagrad, under both from-scratch and warm-start training.

Changes in the revised manuscript (highlighted in red)

  • Added explicit Adam moment updates, physical-step normalization, and sparse-row masking rules in Appendix F.2.
  • Added public Criteo-1TB optimizer results in Appendix G.

We hope the above discussion will fully address your concerns about our work. We really appreciate your insightful and constructive comments to further help us improve the quality of our manuscript. Thank you!

Dear Reviewer eaLh,

Thank you again for your thoughtful review. As the discussion period closes in approximately two days, we wanted to briefly follow up on our responses and additional experiments. We would greatly appreciate your thoughts and welcome the opportunity to clarify anything that remains unresolved.

Many thanks for your time and help.

Best regards,

The Authors of Submission #29233

Dear Reviewer eaLh,

As the discussion period closes in about one day, we would be grateful for the opportunity to hear your thoughts on our rebuttal and additional experiments. We will be online awaiting your feedback and ready to respond promptly to any questions or points that need further clarification.

Many thanks for your time and help.

Best regards,

The Authors of Submission #29233

Reviewer_CeYpOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GIGA to address frequency-dependent coalescing of sparse embedding updates in large-batch recommendation training. GIGA sums within-group active-worker gradient means for parameter updates while accumulating their elementwise squares separately in Adagrad. The authors provide theoretical analysis and evaluate the method on two industrial CTR workloads. Combined with dense-parameter scaling, GIGA maintains baseline ranking quality while improving training throughput at larger batch sizes.

Critical strengths
  • Well-motivated method. GIGA jointly addresses sparse update coalescing and Adagrad state accumulation. Its formulation has interpretable limiting cases and provides a principled way to control aggregation granularity.
  • Practical integration. GIGA reuses existing sparse communication without additional collectives or persistent per-ID state beyond the existing optimizer state, making it practical for distributed embedding training.
  • Substantial industrial evaluation. Experiments include controlled aggregation comparisons, group-size sweeps, state ablations, and multi-day evaluations on industrial workloads. The complete training recipe demonstrates improved throughput at larger batch sizes while preserving ranking quality.
Critical weaknesses
  • Although GIGA preserves the RBI update for single-worker IDs, the paper does not report frequency-stratified prediction results showing how the overall ranking gains are distributed between head and tail IDs.
  • Reduce-Sum and Capped RBI also recover baseline quality at the largest reported batch size, matching GIGA at the displayed precision. Although the state ablations support benefits at smaller scales, the incremental gains over these simpler alternatives are modest and lack uncertainty estimates.
Questions for authors
  1. What evidence establishes a reliable benefit from GIGA’s grouping and separate-square state construction beyond simpler activity-dependent aggregation, particularly given the matching results in Table C.5?
  2. How do GIGA’s ranking gains vary across ID-frequency groups? Does GIGA improve ranking quality for frequently accessed IDs while preserving performance on rare IDs?
  3. Figure 2 uses H++ϵ\sqrt{H^+ + \epsilon} in the update denominator, whereas Equation (7) and Algorithm 1 use H++ϵ\sqrt{H^+}+\epsilon. Which formulation is used in the implementation and reported experiments?
5 comments

We sincerely appreciate the reviewer’s great efforts and insightful comments to improve our manuscript. Below, we address these concerns point by point and try our best to update the manuscript accordingly.


[Q1 / W2] Benefit beyond simpler activity-dependent aggregation

What evidence establishes a reliable benefit from GIGA’s grouping and separate-square state construction beyond simpler activity-dependent aggregation, particularly given the matching results in Table C.5?

Response: At 512 workers, all three methods recover baseline GAUC and improve AUC by 0.05 pt (Table C.7 in the revision, originally C.5). A possible explanation is that quality gains approach saturation in this setting. Within GIGA, changing G from 1 to 8 closes a 0.06-pt GAUC deficit while retaining a 0.05-pt AUC gain. At 128 and 256 workers, we repeated the comparison with three matched seeds. Results are mean ± standard deviation in pt relative to the 32-worker reference:

WorkersMethodΔAUC (pt)ΔGAUC (pt)
128GIGA, G=2+0.0454 ± 0.0056+0.0510 ± 0.0050
128Capped RBI, c=64+0.0389 ± 0.0049+0.0229 ± 0.0036
128Reduce-Sum+0.0358 ± 0.0059+0.0366 ± 0.0031
128GIGA, Q=M²+0.0381 ± 0.0024+0.0340 ± 0.0047
256GIGA, G=4+0.0523 ± 0.0015+0.0316 ± 0.0080
256Capped RBI, c=64+0.0417 ± 0.0043+0.0028 ± 0.0025
256Reduce-Sum+0.0452 ± 0.0041+0.0130 ± 0.0117
256GIGA, Q=M²+0.0427 ± 0.0027+0.0101 ± 0.0056

GIGA uses G=2 at 128 workers and G=4 at 256 workers. It achieves the highest mean AUC and GAUC at both scales. The separate-square state also improves mean GAUC over Q=M² by 0.0170 and 0.0215 pt, respectively.

[Q2 / W1] Prediction quality across ID-frequency groups

Does GIGA improve ranking quality for frequently accessed IDs while preserving performance on rare IDs?

Response: We split evaluation samples into three equal-sized groups by the scored item's exposures over the preceding 15 days. Starting from the same online baseline model, RBI and GIGA train incrementally for 30 days and are evaluated on day 31. GIGA improves both metrics in every group:

Item frequencyΔAUC (pt)ΔGAUC (pt)
Low+0.021+0.040
Medium+0.044+0.119
High+0.054+0.170

These results align with our motivation: large-batch coalescing disproportionately affects frequent IDs, for which GIGA retains finer-grained group contributions. The largest gains occur in the high-frequency group, while the low-frequency group also improves (Section 4.2, Table 2).

[Q3] Position of the Adagrad stabilizer

Which formulation is used in the implementation and reported experiments?

Response: Thank you for pointing out this typo. Our implementation and all reported experiments use H++ϵ\sqrt{H^+}+\epsilon, with ε outside the square root, as in Equation (7) and Algorithm 1. We have corrected Figure 2 accordingly.

Changes in the revised manuscript (highlighted in red)

  • Corrected the position of ε in Figure 2 to match Equation (7) and Algorithm 1 (page 5).
  • Added frequency-bucket prediction results in Section 4.2 (Table 2).

We hope the above discussion will fully address your concerns about our work. We really appreciate your insightful and constructive comments to further help us improve the quality of our manuscript. Thank you!

Dear Reviewer CeYp,

We are writing to follow up on our rebuttal, which includes additional experiments and clarifications. With about two days left in the discussion period, we would be grateful if you could review our responses and share your feedback. We would be happy to address any remaining concerns.

Thank you very much for your time and consideration.

Best regards,

The Authors of Submission #29233

Dear Reviewer CeYp,

With only about one day remaining in the discussion period, we would like to briefly follow up on our rebuttal. We would greatly appreciate your feedback on our responses and additional experiments. We will stay online and wait for your reply so that we can address any remaining concerns promptly.

Thank you very much for your time and consideration.

Best regards, The Authors of Submission #29233

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.