acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Heterogeneous-State Guided Exact Search for Multi-Paper Reviewer Allocation

Abstract

Reviewer allocation requires assigning reviewer groups to papers while satisfying expertise and operational constraints. When multiple papers are assigned simultaneously, reviewer-load constraints couple their assignments, and optimizing aggregate reviewer–paper affinity may leave some papers with substantially lower expertise coverage. We therefore formulate the Multi-Paper Reviewer Allocation (MPRA) problem, which assigns a fixed-size reviewer group to each paper under keyword-relevance, reviewer-load, and group-level social-distance constraints, with the objective of maximizing the minimum keyword coverage over all papers. We prove that MPRA is NP-hard and propose Dynamic Heterogeneous-State Guided Exact Search (HGES), which combines rareness-aware greedy initialization with exact branch-and-bound search. HGES derives paper demand, reviewer availability, keyword rareness, and social feasibility from each partial allocation to guide paper expansion and reviewer-group enumeration, while admissible upper-bound pruning preserves optimality. We further develop an Index-Enhanced Search Acceleration (IESA) module that uses keyword inverted lists, paper-level candidate precomputation, and -hop neighborhood indexing to reduce repeated search operations. On academic-network and attributed social-network instances, exact methods obtain matching average objectives on jointly solved instances. HGES+IESA reduces reported runtime relative to unguided branch-and-bound, while dynamic-ordering ablations show that fewer expanded nodes need not yield shorter runtime. General-purpose mixed-integer optimization remains faster in several settings.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multi-Paper Reviewer Allocation, Dynamic Heterogeneous State, Exact Search, Branch-and-Bound, Max-Min Keyword Coverage.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X9D7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the Multi-Paper Reviewer Allocation (MPRA) problem, which assigns a fixed-size reviewer group to each paper with the objective of maximizing the minimum keyword coverage across papers. The authors formulate MPRA, prove its NP-hardness, and propose HGES, an exact branch-and-bound method that dynamically uses paper demand, reviewer availability, keyword rareness, and social feasibility to guide the search. They further introduce an Index-Enhanced Search Acceleration (IESA) module to reduce repeated candidate retrieval and social-feasibility checking, and evaluate the proposed method on academic and attributed social-network datasets against heuristic and exact baselines.

Critical strengths
  1. The paper studies a meaningful reviewer allocation problem by jointly considering keyword coverage, reviewer workload, and social-distance constraints.

  2. The paper provides extensive experiments on multiple datasets, including comparisons with heuristic and exact baselines, ablation studies, parameter sensitivity analysis, and evaluation of the indexing component.

Critical weaknesses
  1. The novelty and motivation of the studied problem are not sufficiently established. The discussion of related work is insufficient. Several existing studies already consider key ingredients of MPRA, including max-min reviewer fairness [1], multi-task team formation with coverage and workload constraints [2], and max-min allocation with conflict constraints [3]. The paper should more clearly explain what is fundamentally new in MPRA beyond combining these existing components. Second, the motivation for the max-min objective is also limited. The paper argues that aggregate objectives may disadvantage some papers, but does not explain why optimizing the worst-case coverage is preferable to average or total coverage. Since only MinCov is reported, the potential loss in overall allocation quality is unclear. Moreover, the NP-hardness result is also relatively weak. The reduction uses only one paper, so the multi-paper coupling and max-min objective disappear, and the hardness mainly comes from the social-distance constraint through Independent Set. Similar hardness results under conflict constraints already exist [3].

  2. The methodological novelty of HGES is unclear. The main components of HGES, including branch-and-bound, upper/lower bounding, greedy initialization, and dynamic search ordering, have been studied in existing exact optimization methods [4,5]. The heterogeneous state mainly changes the order in which papers and reviewers are explored, while pruning still relies on standard feasibility checks and an admissible upper bound. Therefore, HGES appears more like a problem-specific combination of existing exact-search techniques. The paper should more clearly explain what new algorithmic principle is introduced beyond existing branching and ordering strategies.

[1] Stelmakh et al. PeerReview4All: Fair and Accurate Reviewer Assignment in Peer Review. ALT, 2019.

[2] Vombatkere et al. Forming Coordinated Teams that Balance Task Coverage and Expert Workload. DMKD, 2025.

[3] Chiarelli et al. Fair Allocation of Indivisible Items with Conflict Graphs. Algorithmica, 2023.

[4] Sandholm et al. CABOB: A Fast Optimal Algorithm for Winner Determination in Combinatorial Auctions. Management Science, 2005.

[5] Berktaş and Yaman. A Branch-and-Bound Algorithm for Team Formation on Social Networks. INFORMS Journal on Computing, 2021.

Questions for authors
  1. Could the authors clarify the fundamental novelty of MPRA compared with prior work on max-min reviewer fairness, team formation with coverage/workload constraints, and allocation under conflict constraints? In particular, what new technical challenge arises from MPRA beyond combining these existing components?

  2. Why is the max-min keyword coverage objective preferable to average or total coverage in this setting? Could the authors provide additional results showing whether improving MinCov leads to a substantial loss in overall allocation quality?

  3. Since the NP-hardness reduction uses only one paper and therefore removes the multi-paper coupling and max-min structure, could the authors clarify what this hardness result reveals specifically about MPRA beyond the known hardness induced by conflict/social-distance constraints?

  4. What is the main algorithmic novelty of HGES beyond standard branch-and-bound techniques such as bounding, greedy initialization, feasibility pruning, and dynamic branching/search ordering? Could the authors identify the specific new algorithmic principle introduced by the heterogeneous-state design?

Reviewer_WTghOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers the Multi-Paper Reviewer Allocation (MPRA) problem, where the goal is to maximize the minimum keyword coverage across papers while satisfying reviewer workload and social-distance constraints. The authors propose HGES, a branch-and-bound framework that uses greedy initialization and dynamic state information to guide the search. They also introduce IESA to reduce the cost of candidate retrieval and feasibility checking. I think the problem is meaningful, and the paper is generally clear and easy to follow. The formulation is well explained, and the authors provide theoretical arguments for the correctness of their exact search procedure. I also appreciate the distinction between using heuristics to guide the search and using valid bounds to prune branches without losing optimality. My main concern is the novelty of the proposed method. Most of its components, including greedy initialization, heuristic search ordering, branch-and-bound, and indexing, are well-established techniques. While combining them for this particular problem is reasonable, I am not fully convinced that the dynamic heterogeneous-state design offers a sufficiently new algorithmic contribution. I also have some concerns about the experimental results. The dynamic search strategy reduces the number of expanded nodes in some cases, but this does not consistently lead to faster execution. For example, the fixed-order variant performs much better than the full method on AMiner in the ablation study. The MILP baseline is also considerably faster on Gowalla and Brightkite. These results make it difficult to understand when the proposed approach would be preferable to existing optimization methods. Another limitation is the scale and realism of the evaluation. The experiments use relatively small allocation instances, and keyword coverage and social distance are simplified measures of reviewer expertise and relationships. I would have liked to see stronger evidence that the method can handle larger, more realistic reviewer allocation settings. Overall, I find the problem interesting and the technical presentation solid. However, I am not convinced that the paper establishes enough methodological novelty or practical advantage over existing approaches. Given these concerns, I lean toward rejection in its current form.

Critical strengths

1. Meaningful problem formulation. The paper addresses a practical reviewer allocation problem by considering expertise coverage, reviewer workload, and social-distance constraints together. I particularly appreciate the max-min coverage objective, which focuses on improving the allocation quality of the least-covered paper rather than simply maximizing overall matching scores. This provides a reasonable motivation for jointly optimizing assignments across multiple papers.

2. Clear algorithmic framework with optimality guarantees. The proposed method is well formulated, and the authors provide theoretical justification for its correctness. I appreciate that the dynamic heuristics are used to guide search ordering, while feasibility checks and admissible bounds preserve optimality. This separation makes the approach easier to understand and analyze, and the ability to certify optimal solutions for the considered problem is a useful property.

Critical weaknesses

1. Limited methodological novelty. My main concern is that the proposed approach largely combines existing techniques, including greedy initialization, heuristic search ordering, branch-and-bound, and indexing. Although the dynamic heterogeneous-state design is reasonable, it mainly changes the order in which search branches are explored, without introducing a fundamentally new search or pruning mechanism. I am therefore not fully convinced that the paper offers a sufficiently distinct algorithmic contribution beyond adapting established methods to this particular problem.

2. Insufficient evidence of practical advantages. The experimental results do not consistently support the effectiveness of the proposed dynamic search strategy. In Table 2, the fixed-order variant substantially outperforms the full HGES method on AMiner, while reductions in expanded nodes do not always translate into faster execution. Moreover, the MILP baseline is significantly faster on several datasets, particularly Gowalla and Brightkite. Given these results, it remains unclear under what conditions HGES offers a meaningful advantage over existing exact solvers. This is a major concern because improved search efficiency is one of the paper's central contributions.

Questions for authors

1. Could the authors clarify the large performance difference in the AMiner ablation results (Table 2)? The variant without dynamic updates expands only one node and takes 0.16 seconds, while the full HGES expands 4,086 nodes and takes 149.24 seconds. Since these variants reportedly share the same initialization, feasibility conditions, and upper-bound formulation, I would like to understand what accounts for this difference. Were the results obtained on exactly the same instances under identical experimental settings? This clarification would help me assess the reliability of the ablation results and the actual benefit of dynamic state guidance.

2. Could the authors clarify the main algorithmic contribution beyond existing branch-and-bound techniques? The dynamic heterogeneous state appears to primarily influence search ordering, while optimality is preserved through conventional feasibility checks and admissible bounds. I would be interested in understanding what makes this particular state-guided strategy substantially different from standard problem-specific branching heuristics. A clearer explanation of this distinction would help me better assess the methodological novelty.

3. Under what conditions does HGES offer a meaningful advantage over general-purpose exact solvers? In Table 1, MILP is considerably faster on Gowalla and Brightkite, and also slightly faster on DBLP. Could the authors clarify whether there are identifiable instance characteristics where HGES consistently performs better, and whether preprocessing costs are included in the efficiency comparison? This would help establish the practical significance of the proposed approach.

Reviewer_F6ZfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies reviewer allocation when multiple papers are assigned at the same time. The authors formalize the Multi-Paper Reviewer Allocation (MPRA) problem. Under keyword relevance constraints, per reviewer load limits, and group level social distance constraints, a fixed size reviewer group is assigned to each paper, and the objective is to maximize the minimum keyword coverage over all papers (a max min coverage objective). The authors prove that MPRA is NP hard via a reduction from Independent Set, and propose Dynamic Heterogeneous-State Guided Exact Search (HGES). A rareness aware greedy procedure (GREEDYINIT) first constructs a feasible allocation that serves as an initial lower bound, and an exact branch and bound search then follows. During the search a heterogeneous state is constructed dynamically from the current partial allocation, capturing paper side demand, reviewer availability, keyword rareness, and social feasibility. This state only guides the order of paper expansion and reviewer enumeration, while an admissible upper bound preserves optimality. An Index Enhanced Search Acceleration (IESA) module with keyword inverted lists, paper level candidate precomputation, and k hop neighborhood indexing is further proposed to reduce online lookup cost. Experiments are conducted on DBLP and AMiner (academic networks) and Gowalla and Brightkite (social network stress tests), comparing against Sequential-KTG, MILP, unguided branch and bound, and fixed ordering branch and bound, with component ablations and parameter sensitivity analyses.

Critical strengths

The constraint system of MPRA (group size, keyword relevance, within group social distance, and cross paper load coupling) combined with the max min objective has a clear practical motivation. The argument in Section 1 that pairwise affinity alone fails to capture complementary expertise at the group level is well made. The NP hardness proof and the proofs of upper bound admissibility and exactness (Appendices E and F) appear complete and correct, which supports the reliability of the claim that the method returns an optimal solution.

Critical weaknesses
  1. Limited novelty with a thin technical delta over prior work. The individual ingredients of MPRA, namely coverage based reviewer group selection (Kou et al. 2015, Fu et al. 2025), max min fairness objectives (Stelmakh et al. 2019), group level social constraints (Li et al. 2023, Zhu et al. 2023), and allocation under load coupling, are all well established in the literature. The single paper case is structurally close to the KTG query (Zhu et al. 2023) with a coverage objective, and the NP hardness result (Appendix E, reduced from Independent Set) mainly inherits this known hardness. On the method side HGES is a classical branch and bound framework that combines a greedy initial lower bound, hand crafted ordering heuristics, and admissible upper bound pruning, without introducing any new algorithmic principle. The paper's own positioning (Section 2, Positioning of MPRA) concedes that the distinction lies in the combined formulation and the search procedure rather than in introducing coverage or max min fairness individually. At this level of increment both the problem and the method read as a recombinant application of existing techniques in a new setting, which falls short of a substantive research contribution.

  2. No advantage over the strongest baseline, so the main claim is unsupported. Table 1 shows that a general purpose MILP solver is faster than HGES+IESA on DBLP, Gowalla, and Brightkite (for example 1.49 s versus 1.63 s on DBLP and 0.81 s versus 190.67 s on Gowalla), and Fixed-B&B is also faster than HGES+IESA on Gowalla and Brightkite. In other words, across all scales tested in the paper, simply formulating the equivalent MILP (Appendix D) and handing it to an off the shelf solver is a more efficient or equally efficient choice. The paper positions itself as a problem specific exact search framework whose traversal behavior can be explicitly controlled and analyzed (Section 5.2), yet it never exhibits a regime where HGES outperforms a general purpose solver, such as a scalability curve on larger instances or under particular constraint structures. If no such regime can be demonstrated, the incremental contribution over writing a MILP and calling a solver is hard to justify for publication. Moreover, all methods achieve nearly identical MinCov on the jointly solved instances (Table 1) and GREEDYINIT alone already reaches near optimal solution quality (0.485 versus 0.489 in Table 9), so the practical solution quality gain brought by the whole exact search framework is very limited.

  3. Unconvincing experiments, with several results pointing against the paper's own motivation. Beyond the comparisons with MILP and Fixed-B&B in weakness 2, the ablation (Table 2) shows that the proposed dynamic heterogeneous state guidance does not consistently translate into runtime gains. On DBLP the full HGES (127.50 s) is not faster than the frozen ordering variant (121.90 s), and on AMiner the frozen variant needs only 0.16 s with one expanded node while the full HGES needs 149.24 s with 4,086 nodes. The abstract itself concedes that fewer expanded nodes need not yield shorter runtime. At the same time solution quality is essentially indistinguishable across methods (identical MinCov, with GREEDYINIT at 0.485 versus 0.489). In short, the paper demonstrates neither an efficiency advantage over a general purpose solver nor a solution quality advantage over a fast greedy heuristic, and the central claim that dynamic state guidance accelerates exact search lacks positive evidence in the authors' own experiments.

  4. Experiments are small in scale and the baseline selection is insufficient. (a) The default setting of 16 papers and at most 200 reviewers is two to three orders of magnitude away from any realistic reviewing scenario, the sensitivity analysis (Tables 5 to 8) only goes up to 28 papers, and SolvedRate already degrades substantially within this small range (dropping to 20 percent on Gowalla at 24 papers), so extrapolation to realistic scales is unclear. (b) No recent reviewer allocation or fairness aware method is actually compared. The paper cites directly related works such as Kobren et al. (2019), Stelmakh et al. (2019), and Fu et al. (2025), but the only problem specific baseline is Sequential-KTG, adapted from a database query paper (Zhu et al. 2023, KTG) and deliberately configured as a weak sequential version that cannot revisit earlier groups. Without comparisons against stronger heuristic baselines from fairness aware allocation or coverage based group recommendation, the implicit claim that MPRA requires exact search rather than a good heuristic remains unverified, especially given that GREEDYINIT already achieves near optimal quality.

Questions for authors
  1. Table 1 shows that MILP is faster than HGES+IESA on three of the four datasets. Under what conditions (instance scale, constraint structure, or time limits) does HGES have a practical advantage over a general purpose solver? Could you report a comparison of solving time and time to optimality certificate on larger instances, for example with 100 or more papers? If no such regime exists, how do you justify the necessity of a problem specific exact search framework?

  2. In the ablation of Table 2, the w/o Dynamic Update variant on AMiner needs only 0.16 s and one expanded node, while the full HGES needs 149.24 s and 4,086 nodes, and on DBLP the full HGES is not faster than the frozen ordering variant. Please clarify on which instance characteristics dynamic state updates bring genuine runtime benefits, and provide a profiling breakdown of the overhead of state construction and candidate sorting, which the paper repeatedly notes as needing profiling.

  3. The runtime magnitudes differ greatly across Table 1 (HGES+IESA at 1.63 s on DBLP), Table 10 (HGES at 4.27 s and HGES+IESA at 1.63 s), and Table 2 (HGES at 127.50 s on DBLP). Please confirm whether the ablation experiments (Table 2) used a different instance set or configuration than the main comparison (Tables 1 and 10). If so, please state this explicitly in the main text.

  4. The related work section cites several fairness aware reviewer allocation and coverage based group recommendation methods (Kobren et al. 2019, Stelmakh et al. 2019, Fu et al. 2025), yet none of them is compared with in the experiments. How would a natural two stage baseline, for example a fairness or coverage based allocation that ignores social distance followed by a postprocessing repair step for social feasibility, compare with HGES in solution quality and running time?

  5. All methods achieve nearly identical MinCov on the jointly solved instances, and GREEDYINIT alone is already near optimal (0.485 versus 0.489 in Table 9). In realistic deployments where only a high quality feasible solution is needed, what is the marginal value of exact search over a fast greedy heuristic? Under what instance structures does the gap between the optimal solution and a greedy solution under the max min objective become substantial?

  6. In the single paper case with m equal to 1, what is the essential difference between MPRA and a KTG query (Zhu et al. 2023) with a coverage objective? Is cross paper load coupling the only new algorithmic challenge of MPRA relative to solving KTG per paper? Please explain concretely how it changes the search structure, that is, which search behaviors cannot be realized by solving KTG per paper with a global capacity conflict check and instead require the joint exact search proposed here?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.