acceptodds
Under review as a conference paper at ICLR 2027

DAPG: Distance-Aware Pruning for Dynamic Graph-Based Approximate Nearest Neighbor Search

Abstract

Approximate nearest neighbor (ANN) search over high-dimensional vector data is a core component of vector databases, recommendation systems, and retrieval-augmented generation. Graph-based indexes provide strong search efficiency, but most are designed for largely static datasets. This implies that under frequent insertions and deletions: (i) graph connectivity may degrade, (ii) maintenance cost can increase, and (iii) redundant edges can further increase query latency and memory usage. To address these challenges, we propose the Distance-Aware Pruned Graph (DAPG), a dynamic graph-based ANN index designed to maintain high-recall search while supporting efficient updates. DAPG constructs an LSH-seeded single-layer proximity graph and applies a two-stage pruning strategy that balances sparsity and navigability. First, a node-specific percentile threshold adapts edge retention to the local distance distribution. Second, adaptive global sparsification constrains high-degree nodes and removes redundant connectivity. For dynamic workloads, DAPG performs localized maintenance by reapplying pruning only to affected hash buckets, candidate neighborhoods, and adjacency lists, thereby avoiding full index reconstruction. We implement DAPG and evaluate it on seven datasets against eleven representative ANN baselines, including state-of-the-art dynamic methods. In static search, DAPG improves the recall-latency trade-off, reducing query latency by up to . For dynamic workloads, DAPG reduces maintenance time by up to relative to rebuild-based baselines. On MS-MARCO1M, specifically, DAPG achieves average Recall@100 of compared to for Wolverine++, while providing higher deletion throughput.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Approximate Nearest Neighbor Search, Dynamic Indexing, Graph-Based ANN, Adaptive Graph Pruning, Vector Search, Locality-Sensitive Hashing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Fqs7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DAPG, a single-layer graph index for dynamic ANN search, combining LSH-based candidate generation, node-specific percentile pruning, global degree control, and localized updates. It is evaluated on seven datasets against eleven ANN baselines.

Critical strengths

S1. Efficient ANN search under frequent updates is important, and the proposed maintenance framework is relatively simple.

S2. The experiments cover multiple datasets, static search, and dynamic workloads.

Critical weaknesses

W1. Insufficient evidence for the pruning mechanism. With fixed candidates and degree cap, Algorithms 3–4 effectively retain the closest min(r_i, T') neighbors. Table 5 shows only a 0.0004 improvement in maximum recall over Global-only at nearly identical average degrees, without uncertainty estimates for this difference.

W2. Insufficient support for dynamic maintenance. Algorithm 1 does not clearly explain incoming-edge tracking or replacement-edge generation after deletion. There is a lack of throughput comparisons at the same recall level.

Questions for authors

Q1. When does percentile pruning produce a different final graph from nearest-neighbor truncation under identical candidates and degree budgets?

Q2. How does Algorithm 1 identify incoming neighbors and restore lost connections after deletion?

Q3. How does distance-based pruning preserve navigational paths, and how much of the search improvement comes from LSH entry points rather than pruning?

Q4. Can dynamic methods be compared at the same Recall@100, including auxiliary-index, incoming-edge maintenance, and cleanup costs?

Q5. The paper proposes an adaptive degree cap (T'=\alpha\bar d_{\mathrm{local}}), but the main static experiments fix (T'=48). What is the independent benefit of the adaptive formulation?

Q6. Could the authors include comparisons with relevant dynamic ANN methods such as IP-DiskANN and SPFresh?

Q7. Why are the reported index size and build time for SIFT100M nearly identical to those for SIFT1M despite the 100-fold difference in scale? Could the authors also clarify the query parameters?

2 comments
AuthorsOpenReview ↗

We thank Reviewer Fqs7 for the detailed feedback and suggestions for additional

Q1.
We agree that, for a fixed candidate set and degree budget, percentile pruning is equivalent to retaining the closest min⁡(⌈pmi⌉,T′)\min(\lceil p m_i\rceil,T') candidates, apart from ties. In DAPG, percentile pruning is applied to LSH-derived candidates before SearchInBuilding. The retained candidates seed graph refinement, which may discover more candidates before global degree control is applied. Thus, local pruning can change the candidate set reaching the final pruning stage. We revised Algorithm 2 to make this order explicit.

Across five paired SIFT1M builds, Full DAPG and Global-only have the same average degree (36.55336.553) and nearly identical maximum Recall@100 (0.9800±0.00090.9800\pm0.0009). At Recall@100 =0.95=0.95, Full DAPG reduces latency from 0.881±0.0350.881\pm0.035 ms to 0.848±0.0380.848\pm0.038 ms (3.76%3.76\%). We therefore narrow our claim to a modest construction-stage efficiency benefit rather than a recall improvement.

Q2.
On deletion, the implementation marks the vertex inactive, removes it from the LSH tables, collects and clears its adjacency list, removes the deleted identifier from those neighbors, and reapplies local and global degree control. Search skips inactive identifiers. The implementation does not maintain a complete reverse-adjacency index or create replacement edges during deletion. Thus, some incoming references may remain inactive and are excluded from traversal. New insertions can create new connections through the normal LSH-seeded SearchInBuilding procedure, but deletion alone does not restore every lost path. We revised Algorithm 1 and the discussion.

Q3.
We evaluate pruning and LSH entry-point selection separately. With LSH-seeded search fixed, Full DAPG and Global-only have nearly identical maximum Recall@100 (0.9800±0.00090.9800\pm0.0009), while Full DAPG reduces latency at Recall@100 =0.95=0.95 from 0.881±0.0350.881\pm0.035 ms to 0.848±0.0380.848\pm0.038 ms (3.76%3.76\%). Thus, local pruning provides a modest efficiency benefit rather than a recall gain. To study entry-point selection, we hold the pruned graph fixed and vary the traversal initialization. We therefore attribute search behavior to both construction-stage refinement and query-time entry-point selection rather than to pruning alone.

Q4.
We added a matched-recall comparison using the same MS-MARCO-1M vectors, queries, ten-round FIFO states, exact ground truth, and one measured worker thread. At Recall@100 ≈0.9594\approx0.9594, DAPG achieves 238/71,120/2,393 search/delete/insert OPS, compared with 804/66,963/2,464 for IP-DiskANN. Wolverine++ achieves approximately 1,550/777/860 OPS at Recall@100 =0.9590=0.9590. IP-DiskANN provides 3.37×3.37\times higher search throughput and 1.03×1.03\times higher insertion throughput, while DAPG provides 1.06×1.06\times higher deletion throughput. Wolverine++ provides higher search throughput, while DAPG provides substantially higher deletion and insertion throughput.

Q5.
The adaptive formulation and fixed-cap evaluation serve different purposes. The rule T′=⌈αdˉlocal⌉T'=\lceil\alpha\bar d_{\mathrm{local}}\rceil automatically calibrates the global degree budget to the candidate-neighborhood scale of a dataset, reducing the need to retune an absolute cap across datasets. In the main experiments, we fix T′=48T'=48 to control final sparsity and isolate local pruning under a common degree budget. Once T′T' is fixed, the adaptive and manually selected caps produce the same global pruning operation. Thus, the intended independent benefit of the adaptive formulation is parameter transferability rather than a different graph at an identical cap.

Q6.
We added a direct matched-recall comparison with IP-DiskANN [1] using the same MS-MARCO-1M vectors, queries, ten-round FIFO workload, exact ground truth, and one measured worker thread. At Recall@100 ≈0.9594\approx0.9594, DAPG achieves 1.06×1.06\times the deletion throughput of IP-DiskANN. IP-DiskANN achieves 1.03×1.03\times higher insertion throughput and 3.37×3.37\times higher search throughput. We also added the experimental settings, uncertainty across FIFO states, and a DAPG degree sweep. We added SPFresh [2] to Related Work but do not report a direct throughput comparison because SPFresh targets a substantially different billion-scale, disk-resident, partition-based architecture, whereas DAPG is evaluated as an in-memory graph index. We now state this scope difference explicitly.

Q7.
The index-size and build-time results in Fig. 4(b) correspond to SIFT1M. We updated Fig. 4(b) to correctly label the dataset as SIFT1M and revised the corresponding text and caption. For the SIFT100M search experiment in Fig. 4(a), we evaluate k∈{10,20,…,100}k\in\{10,20,\ldots,100\} with ef=k+150\mathit{ef}=k+150. DAPG uses L=2L=2, K=18K=18, T=24T=24, T′=48T'=48, W=1.0W=1.0, pC=0.95p_C=0.95, pQ=0.9p_Q=0.9, and efC=80\mathit{ef}_C=80.

References
[1] Xu et al. “In-Place Updates of a Graph Index for Streaming Approximate Nearest Neighbor Search,” 2025.
[2] Xu et al. “SPFresh: Incremental In-Place Update for Billion-Scale Vector Search,” SOSP 2023.

We hope these revisions address your concerns. If they do, we would be grateful if you would consider updating your assessment. We are happy to answer any further questions.

Reviewer_wZWoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DAPG, a dynamic graph-based approximate nearest neighbor search method that aims to maintain high search accuracy while supporting efficient insertions and deletions. DAPG uses LSH for candidate generation and employs a two-stage pruning strategy consisting of node-specific percentile-based pruning and global degree constraints to construct a sparse proximity graph. It further supports dynamic updates through localized graph maintenance without full index reconstruction. Experiments on seven datasets against eleven baselines demonstrate improvements in recall-latency trade-offs and update efficiency.

Critical strengths

S1. This paper addresses an important problem in dynamic graph-based ANNS: maintaining high search quality while supporting efficient insertions and deletions. The proposed DAPG combines sparse graph construction with localized index maintenance, thereby avoiding full index reconstruction. The results on MS-MARCO1M are promising: DAPG achieves a higher average Recall@100 than Wolverine++ and approximately 11× higher deletion throughput. These results suggest that DAPG may be well suited to workloads involving frequent updates.

Critical weaknesses

W1. The novelty and rationale of proposed "distance-aware pruning" are limited. The proposed node-specific percentile pruning is equivalent to retaining a fixed proportion of the nearest neighbors, followed by the global degree cap T′ . Thus, the final neighbor selection depends on distance rankings, while the number of retained neighbors is primarily determined by the candidate set size m_i, percentile p, and degree cap T', rather than the shape of the local distance distribution. It appears to offer little beyond standard kNN graph construction. This is consistent with Table 5: Full DAPG and Global-only pruning have nearly identical average degrees, maximum recall, and query latency. Since local pruning is the paper’s main contribution, its limited demonstrated benefit weakens the paper’s novelty and technical significance.

W2. It does not adequately address the potential degradation of graph connectivity resulting from deletion operations. The deletion procedure in Algorithm 1 removes the deleted node and its incident edges, then reapplies local and global pruning to affected neighborhoods. However, these pruning operations only remove existing edges and do not explicitly introduce replacement connections. Therefore, it is unclear how the algorithm repairs navigation paths disrupted by deletions or prevents connectivity degradation under repeated updates. Although the empirical results show relatively stable recall, the underlying maintenance mechanism and its limitations are not sufficiently explained.

W3. The theoretical guarantees for the proposed pruning strategy are very limited. Proposition 1 follows directly from the global degree cap, while Proposition 2 decomposes traversal cost without establishing a bound on the number of expansions. More importantly, Proposition 3 derives a conditional recall bound using the probability chain rule, but provides no explicit guarantees on \rho_{\mathrm{loc}} or \rho_{\mathrm{reach}} in terms of the pruning parameters. Consequently, the analysis does not establish under what conditions the proposed pruning preserves graph navigability or search recall.

Questions for authors

Q1. In Algorithm 2, each node first selects its top-k candidates. Since local percentile pruning retains approximately a fixed proportion p of these candidates, how does this differ from directly selecting the closest \lceil pk\rceil neighbors?

Q2. Figure 6 shows substantially lower Search OPS for DAPG than Wolverine++, despite its higher recall and deletion throughput. Could the authors report search throughput and update throughput at matched Recall@100 levels? This comparison is important for determining whether DAPG provides a better overall trade-off between search accuracy, query efficiency, and maintenance cost.

1 comment
AuthorsOpenReview ↗

Clarifications and Additional Experiments

We thank Reviewer wZWo for the constructive feedback and careful evaluation.

Response to Q1.

We agree that, for a fixed candidate list with distinct distances, percentile pruning is equivalent to selecting the closest ⌈pmi⌉\lceil p m_i\rceil candidates. In DAPG, the distinction is not the ranking rule itself, but where pruning occurs in the construction pipeline. Local pruning filters the LSH-derived candidates before graph refinement. The retained candidates seed SearchInBuilding, which may discover additional graph candidates before the global degree cap is applied. Thus, the final neighborhood is not generally the closest min⁡(⌈pmi⌉,T′)\min(\lceil p m_i\rceil,T') elements of a single fixed candidate list.

We revised Algorithm 2 to make this execution order explicit and narrowed our claim accordingly. Across five paired SIFT1M builds, Full DAPG and Global-only have nearly identical maximum Recall@100 (0.9800±0.00090.9800\pm0.0009). At Recall@100 =0.95=0.95, Full DAPG reduces latency from 0.881±0.0350.881\pm0.035 ms to 0.848±0.0380.848\pm0.038 ms (3.76%3.76\%). These results support a modest construction-stage efficiency benefit rather than a robust recall improvement.

Response to Q2.

We added a matched-recall comparison using the same MS-MARCO-1M vectors, queries, exact per-round ground truth, ten-round FIFO workload, and one measured worker thread. We swept DAPG ef∈{1200,1400,2100}\mathit{ef}\in\{1200,1400,2100\} and Wolverine++ efSearch∈{100,120,150}\mathit{efSearch}\in\{100,120,150\}.

At mean Recall@100 of 0.95940.9594 for DAPG (ef=1200\mathit{ef}=1200) and 0.95900.9590 for Wolverine++ (efSearch=120\mathit{efSearch}=120), DAPG achieves 238 search OPS, 71,120 deletion OPS, and 2,393 insertion OPS, compared with 1,550, 777, and 860 OPS for Wolverine++, respectively. Wolverine++ therefore provides 6.51×6.51\times higher search throughput, while DAPG provides 91.5×91.5\times higher deletion throughput and 2.78×2.78\times higher insertion throughput. We revised the paper to present these results as a search and update trade-off rather than uniform dominance.

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

DAPG builds a graph index using candidates found by LSH. For each point, it keeps nearby candidates using a local distance threshold and then limits the number of remaining edges. Queries start from points selected through LSH and search the graph for neighbors. Insertions and deletions update the affected parts of the index. Experiments on seven datasets compare recall, query time, index size, and update costs with existing static and dynamic ANN methods.

Critical strengths

(1) The experiments include both FIFO and random updates and measure recall using the points still present in the index. They also follow the index over many update rounds. This gives useful evidence about how search behaves as points are added and removed, beyond a comparison on a fixed dataset.

(2) The paper reports query time, build time, index size, and update time alongside recall. It also tests the local and global pruning steps separately. These results make it easier to see where the savings come from and whether lower cost is accompanied by lower search accuracy.

Critical weaknesses

(1) Both pruning stages keep the closest candidates. When all distances are different, the final neighbors are simply the first few points in the same list sorted by distance. The two stages change how many neighbors are kept, but not how points are chosen once that number is fixed. Table 5 shows only a 0.0004 recall gain over the Global only variant. Full DAPG is also slightly slower at the same recall. This gives limited support for the added local rule.

(2) It is not clear which operations are performed after deletion. Algorithm 1 prunes the candidate neighbors N(u)N(u), but does not say where this list comes from. It could contain the remaining neighbors, candidates saved during construction, or points found by a new LSH search. These choices require different amounts of work and memory and can produce different graphs. The long update experiments are useful, but do not explain which procedure produced the reported update times.

Questions for authors

(1) What does the local percentile rule add beyond choosing how many of the closest neighbors to keep? Please show whether the 0.0004 recall gain in Table 5 remains across repeated index builds. Does it produce a useful improvement in query speed when recall is the same?

(2) Please confirm the search and deletion procedures used in the experiments. In Algorithm 5, k=1k = 1 with one starting point causes search to stop before checking any neighbors. Does the code use this condition? For deletion, please explain how N(u)N(u) is obtained and which operations are included in the measured update time.

(3) At the same Recall@100 after updates, what are the search, insertion, and deletion times for DAPG and Wolverine++? Please state which settings were tried for both methods. Their reported recalls differ, so the current results do not fully show which method is cheaper at the same search accuracy.

3 comments
AuthorsOpenReview ↗

Clarifications and Additional Experiments

We thank Reviewer qmKd for the detailed comments and helpful suggestions.

Response to Weakness (1) and Question (1).

We agree that, for a fixed candidate set with distinct distances and the same degree budget, percentile pruning is equivalent to selecting the closest ⌈pmi⌉\lceil p m_i\rceil candidates. Our original presentation overstated the selection rule itself. In DAPG, its effect comes from its position in the construction pipeline. It filters LSH-derived candidates before SearchInBuilding. The retained candidates seed graph refinement, which may discover additional candidates, and global degree control is applied afterward. Thus, the final neighborhood is not generally obtained by truncating a single fixed candidate list. We revised Algorithm 2 to expose this order and describe local pruning as a rank-percentile construction-stage filter.

MethodAvg. DegreeMax Recall@100[email protected] (ms)Build (s)Dist. Comps./Node
Full DAPG36.5528 ± 0.00190.9800 ± 0.00090.848 ± 0.03890.67 ± 5.351228.49 ± 41.16
Global-only36.5529 ± 0.00200.9800 ± 0.00090.881 ± 0.035119.93 ± 43.361228.57 ± 41.12

Values are mean ± standard deviation over five paired SIFT1M builds using identical seeds and degree budgets. [email protected] denotes latency at Recall@100 = 0.95.

Across the five paired builds, Full DAPG and Global-only have nearly identical average degree (36.553) and the same mean maximum Recall@100 (0.9800 ± 0.0009), confirming that the 0.0004 difference in the single-build ablation is not a robust recall improvement. At Recall@100 = 0.95, Full DAPG reduces mean latency from 0.881 ± 0.035 ms to 0.848 ± 0.038 ms, a 3.76% reduction. We therefore narrow our claim to a modest efficiency benefit from construction-stage filtering, while global pruning primarily determines final graph sparsity.

Response to Weakness (2) and Question (2).

The implementation uses LSH-derived seeds followed by best-first traversal of the pruned graph. The stopping condition shown in Algorithm 5 did not reflect the implementation. With a single initial result, the written condition could terminate before expanding its neighbors. The implementation allows termination only after the result beam contains ef\mathit{ef} candidates and the closest unexpanded candidate is farther than the worst candidate in the beam. We revised Algorithm 5 accordingly.

For deletion of vertex vdv_d, the implementation (i) marks vdv_d inactive, (ii) removes vdv_d from all LSH buckets, (iii) treats the vertices in vdv_d's current adjacency list as the affected set, (iv) clears vdv_d's adjacency list, (v) removes vdv_d from those affected adjacency lists, and (vi) reapplies local and global pruning to their remaining stored neighbors. Thus, the candidate list in Algorithm 1 is the remaining current adjacency list after removing vdv_d, not saved construction candidates or a new LSH search.

The implementation does not maintain a complete reverse-adjacency index or generate replacement edges during deletion. Stale incoming references not represented in vdv_d's adjacency list may remain, but inactive identifiers are excluded from traversal. We revised Algorithm 1 to state this behavior and limitation explicitly. Reported deletion time includes inactive marking, LSH-bucket removal, locking, adjacency cleanup, and repruning. Initial construction, data loading, later insertions, and ground-truth generation are excluded.

Response to Question (3).

We added a matched-recall comparison using identical MS-MARCO-1M vectors, queries, exact per-round ground truth, one measured worker thread, 900K initially active passages, and ten FIFO rounds replacing 10K passages per round. We swept DAPG ef∈{1200,1400,2100}\mathit{ef}\in\{1200,1400,2100\} and Wolverine++ efSearch∈{100,120,150}\mathit{efSearch}\in\{100,120,150\}.

At mean Recall@100 of 0.9594 for DAPG (ef=1200\mathit{ef}=1200) and 0.9590 for Wolverine++ (efSearch=120\mathit{efSearch}=120), the results are:

MethodSearchDeletionInsertion
DAPG238 OPS (4.198 ms)71,120 OPS (0.0141 ms)2,393 OPS (0.418 ms)
Wolverine++1,550 OPS (0.645 ms)777 OPS (1.287 ms)860 OPS (1.163 ms)

Wolverine++ provides 6.51× higher search throughput, while DAPG provides 91.5× higher deletion throughput and 2.78× higher insertion throughput. We therefore present this result as a search and update trade-off rather than uniform superiority by either method.

Thank you for your response. Based on the additional results, the proposed local pruning rule provides no robust recall improvement and only a small latency reduction at the target recall. Thus, my original concern about the limited benefit remains, and I maintain my rating.

AuthorsOpenReview ↗

Clarification of DAPG Novelty and Dynamic Evaluation

Thank you for the assessment. We respectfully clarify that DAPG's novelty is not the local percentile operator in isolation. Our earlier presentation may have overemphasized that component. We design DAPG as a new dynamic proximity graph index with an integrated design spanning construction, querying, and dynamic maintenance.

During construction, DAPG obtains LSH-derived candidates, applies the local filter before graph refinement, uses the retained candidates to seed SearchInBuilding, and applies global degree control only after refinement. Thus, the final neighborhood is not produced by truncating one fixed sorted list. At query time, multiple LSH-derived entry points reduce dependence on a single navigation path through the sparse graph. During updates, DAPG modifies the LSH buckets and currently affected adjacency lists without reconstructing the complete index. These stages work together. Early filtering controls the refinement input, global pruning bounds graph degree, and the resulting low-degree graph reduces localized maintenance cost.

Our new SIFT1M ablation clarifies the role of local pruning. Across five paired static builds, Full DAPG and Global-only have the same average degree (36.55336.553) and nearly identical maximum Recall@100 (0.9800±0.00090.9800\pm0.0009). At matched Recall@100 =0.95=0.95, however, Full DAPG reduces mean query latency from 0.881±0.0350.881\pm0.035 ms to 0.848±0.0380.848\pm0.038 ms, a 3.76%3.76\% reduction. We therefore do not claim a robust recall gain from the local rule. Its measured contribution is a modest construction-stage efficiency benefit, while global pruning primarily controls final sparsity.

The broader system-level benefit is demonstrated by the MS-MARCO-1M dynamic FIFO experiment. At matched Recall@100 (0.95940.9594 for DAPG and 0.95900.9590 for Wolverine++), DAPG provides 91.5×91.5\times higher deletion throughput and 2.78×2.78\times higher insertion throughput, while Wolverine++ [3] provides 6.51×6.51\times higher search throughput. On the same MS-MARCO-1M workload, DAPG and IP-DiskANN achieve mean Recall@100 of approximately 0.95940.9594. DAPG provides deletion throughput comparable to IP-DiskANN [1] and near-equal insertion throughput, whereas IP-DiskANN provides 3.37×3.37\times higher search throughput.

MethodRecall@100 ↑Search OPS ↑Delete OPS ↑Insert OPS ↑
DAPG0.959423871,1202,393
Wolverine++0.95901,550777860
IP-DiskANN0.959480466,9632,464

These results do not establish that local pruning alone is a major source of recall improvement. Rather, they show that the complete DAPG graph design, including LSH-seeded construction, graph refinement, global sparsity control, query traversal, and localized updates, demonstrates a clear trade-off between search and maintenance across static and dynamic workloads.

Reference

[3] Liu et al., Wolverine: Highly Efficient Monotonic Search Path Repair for Graph-Based ANN Index Updates, PVLDB, 2025.

Comments on the paper

1 comment

Dear Reviewer,

Thank you again for your careful reading and constructive feedback. We have uploaded a revised manuscript that addresses the concerns raised, clarifies the algorithms and dynamic-update procedures, and includes new matched-recall experiments and additional baseline comparisons. All revisions and newly added results are highlighted in blue.

With the author response deadline approaching, we wanted to ask whether you have any remaining questions or concerns. We would be happy to clarify any points while the discussion remains open.

If our responses and revisions have addressed your concerns, we would appreciate it if you would consider updating your score.

Thank you again for your time and thoughtful feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.