{"id":"c47acb6c-f6d3-4c97-a2c4-42681bc65674","arxiv_id":"2507.20498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"MoKGR personalizes knowledge graph reasoning by gating over path lengths and over three pruning experts, and reports improved accuracy on transductive and inductive benchmarks.","lead":"This paper introduces MoKGR, a knowledge graph reasoning method that chooses different search depths and pruning strategies for each query instead of using one fixed recipe. It reports accuracy improvements over existing methods on six standard benchmarks, in both transductive and inductive settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"YAGO3-10's headline gains are not attributable to the expert mechanisms: PPR pre-filtering (App. D.1) bounds every later decision, and Table 3b's inference numbers appear to contradict the paper's efficiency claim.","rationale":"The central claim requires the proposed length and pruning experts to be the cause of the reported accuracy and efficiency gains. Appendix D.1 introduces a PPR-based subgraph extraction step that is applied before any expert mechanism runs on YAGO3-10, so the paper's marquee result is confounded unless PPR recall is shown to be effectively perfect and matched against the PPR-based baseline. This matches the reader's weakest_assumption. The same attribution problem affects Figure 3's efficiency claims: Table 3b appears to show MoKGR inference slower than one-shot-subgraph, contradicting the prose, so the efficiency half of the central claim is not yet supported by the submitted evidence. This is not a reason to reject outright: the transductive results on WN18RR, FB15k237, and NELL-995, plus the inductive results in Table 6, do not depend on the appendix PPR step, and the ablations do support complementarity of the expert combinations. Thus a conditional verdict is appropriate, pending code release, PPR controls, and corrected timing reports. Appendix F's \"theorems\" are assertions without derivations and conflict with Section 6's disclaimer about theoretical guarantees; they should be removed or explicitly labeled informal, but they are not the load-bearing threat to the empirical claim.","tokens_in":25917,"tokens_out":13828,"duration_ms":150247,"concrete_test":"Run one PPR-controlled YAGO3-10 evaluation: for each query, extract subgraphs at several retention budgets, report true-answer recall, and measure MoKGR vs. one-shot-subgraph MRR/Hit@k and wall-clock inference on identical subgraphs and hardware. If answer recall is materially below 100% or MoKGR's margin (and inference speedup) disappears when budgets are matched, the YAGO3-10 results cannot be attributed to the length/pruning experts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest empirical evidence for MoKGR is the YAGO3-10 row of Table 1 (MRR 0.657 vs. 0.606 for one-shot-subgraph) and the efficiency comparison in Section 4.2 and Figure 3. Both rest on an undocumented component: Appendix D.1 states that \"for ultra-large datasets such as YAGO3-10, we performed preliminary subgraph extraction using PPR in advance,\" and Algorithm 2 builds G_sub from the PPR cache before any length or pruning expert is applied. Hence every YAGO3-10 score is upper-bounded by PPR recall; if the true answer is not in the extracted subgraph, no expert routing can recover it. The main text reports no PPR recall, subgraph budget, or threshold, and Table 2's YAGO3-10 ablations are all conducted inside the same PPR subgraph, so they cannot separate the PPR contribution from the expert contribution. Since one-shot-subgraph is itself PPR-based, the reported +0.051 MRR margin could reflect a different PPR construction or budget rather than the length/pruning experts. The efficiency claim is additionally troubled by Table 3b, which lists MoKGR inference as 358.3 versus 71.9 for one-shot-subgraph and 84.2 for AdaProp while Section 4.2 claims \"significantly faster inference times\"; unless the units or protocol differ, the paper's own table contradicts the efficiency claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"MoKGR proposes a mixture-of-experts framework for knowledge-graph reasoning whose two components are (1) a mixture of length experts that adaptively selects and weights reasoning-path lengths per query and (2) a mixture of pruning experts that combines scoring, attention, and semantic similarity to retain per-layer entity sets. The paper reports experiments on six transductive benchmarks and twelve inductive splits, claiming superior accuracy and efficiency over NBFNet, RED-GNN, A*Net, AdaProp, and one-shot-subgraph. Ablations on WN18RR and YAGO3-10 support the contribution of the balancing terms, the noise term, and the individual experts, and a case study on Family illustrates query-adaptive path selection.","tokens_in":26247,"tokens_out":4739,"duration_ms":48202,"significance":"If the accuracy and efficiency claims held, MoKGR would be a useful contribution to query-adaptive reasoning over knowledge graphs, and the expert-wise ablations are well-motivated. The paper's strengths include evaluation on standard transductive and inductive benchmarks, a thorough ablation of the proposed components, and a case study that concretely demonstrates personalized path selection. However, the attribution of the headline YAGO3-10 gains to the expert mechanisms is not established because PPR pre-filtering bounds every later decision, the reported inference numbers in Table 3b appear to contradict the efficiency claim, and the theoretical appendix does not contain valid derivations. These issues are load-bearing for the central claims and must be resolved before the results can be fully credited.","major_comments":[{"comment":"The YAGO3-10 results in Table 1 cannot be attributed to MoKGR's length and pruning experts because Algorithm 2 constructs G_sub from the PPR cache before any expert mechanism is applied, and Appendix D.1 states that PPR pre-filtering is performed in advance for YAGO3-10. Thus every later score is upper-bounded by PPR recall. The paper does not report PPR recall, the subgraph budget, the PPR threshold, or the number of retained entities, and Table 2's YAGO3-10 ablations are all conducted inside the same PPR subgraph, so they cannot separate the PPR contribution from the expert contribution. Since one-shot-subgraph is itself PPR-based, the reported +0.051 MRR margin could reflect a different PPR construction or budget rather than the proposed experts. Please report PPR recall on YAGO3-10, include an ablation with and without PPR, and control the PPR budget across baselines if a comparative efficiency or accuracy claim is retained.","section":"§4.2 / Appendix D.1 / Algorithm 2"},{"comment":"The efficiency claim in Section 4.2 that MoKGR achieves 'significantly faster inference times' is contradicted by Table 3b, which reports MoKGR inference time of 358.3 versus 71.9 for one-shot-subgraph and 84.2 for AdaProp. Unless the units or evaluation protocol differ between the rows, the paper's own table undermines the inference-efficiency claim. Please clarify the units and protocol, provide a unified timing setup, and correct the claim if it is not supported by the measurements.","section":"§4.2 / Table 3b / Figure 3"},{"comment":"The theoretical analysis in Appendix F is not a valid derivation. Theorem 2 assumes without proof that the learned compatibility score [Q(c_q)]_l correlates with negative expected reasoning error. Theorem 3 assumes that at least one of the three pruning experts ranks every optimal-path entity in its top-K set and that expert failures are independent, which is exactly the property needing proof. Theorem 4 assumes that the compatibility score estimates mutual information between the answer and path length and cites Huang et al. 2025, an unrelated code-debugging paper. Section 6 correctly states that 'developing theoretical guarantees for the optimality of the selected paths remains challenging,' which is consistent with my assessment that these theorems should be removed or explicitly labeled as conjectures or empirical motivations rather than presented as proofs.","section":"Appendix F, Theorems 2–4"},{"comment":"Several hyperparameters that directly control the method's behavior are not reported: the truncation threshold T in Eq. (18), the adaptive sampling parameters K_s, K_h, K_l, l_i, and a in Eq. (24), and the PPR exploration scope in Appendix D.1. These parameters influence early stopping, exploration breadth, and candidate retention at every layer, so the reported results cannot be reproduced without them. Please include a complete hyperparameter table with final values for each dataset in Appendix A.1.","section":"Eq. (18), Eq. (24), Appendix D.1"},{"comment":"The central empirical claim is that MoKGR consistently outperforms strong baselines, but no error bars, multiple seeds, or significance tests are reported. Several margins in Tables 1 and 6 are small relative to the reported precision, so the claim of 'exceptional performance across all benchmark datasets' is stronger than the evidence. Please report standard deviations over at least three seeds or another appropriate uncertainty measure, and temper the wording of the performance claim accordingly.","section":"§4.2 / Tables 1 and 6"}],"minor_comments":[{"comment":"'On diverse benchmark' should be 'on diverse benchmarks'.","section":"Abstract"},{"comment":"Eq. (3) defines epsilon as Gaussian noise epsilon ~ N(0,1), but the ablation in Table 2 treats epsilon as a scalar (epsilon = 0 or epsilon = 0.2). Please clarify how the scalar epsilon is used in the fixed-noise setting and how the Gaussian sample is drawn in the dynamic setting.","section":"Eq. (3) / Table 2"},{"comment":"Use consistent names: 'Adaprop' appears in Section 4.1 while 'AdaProp' is used elsewhere, and 'Y AGO3-10' contains inconsistent spacing in Tables 1 and 3.","section":"Throughout"},{"comment":"The captions of Figure 5 contain garbled tokens such as 'Ours/uni' and '/uni00000017/uni00000018'; these should be repaired.","section":"Figure 5"},{"comment":"The case study states that MoKGR 'emphasizes aunt for niece queries' and 'brother, nephew, and uncle for brother queries,' but the heatmaps plot relation-type ratios and need a short explanation of how the reported emphasis is quantified.","section":"§4.4 / Figure 6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take. MoKGR is a real, if incremental, combination: it couples hop-dependent length experts with three learned pruning experts and query-level gating. The ablations are consistent with the story, and the non-YAGO results on WN18RR, FB15k-237, and NELL-995 are plausible. That is the good part.\n\nThe soft spots are serious. First, the YAGO3-10 headline is not attributable to the expert mechanisms. Appendix D.1 and Algorithm 2 show that for ultra-large datasets the model runs inside a PPR-precomputed subgraph before any length or pruning expert is applied. No PPR recall or subgraph budget is reported. Since one-shot-subgraph is itself PPR-based, the +0.051 MRR margin could just reflect a different PPR construction or budget. Table 2 ablations on YAGO run inside the same PPR subgraph, so they cannot separate the PPR contribution from the expert contribution. This is a load-bearing confound, not a minor omission.\n\nSecond, the efficiency claim is contradicted by the paper's own numbers. Table 3b lists MoKGR inference at 358.3 versus 71.9 for one-shot-subgraph and 84.2 for AdaProp, while the text claims \"significantly faster inference times.\" Unless the units differ, that is the opposite. That paragraph should not be in the paper as written.\n\nThird, Appendix F's theorems are assertions, not derivations. Theorem 2 assumes the gating scores learn to encode the error they are supposed to minimize; Theorem 4 calls the compatibility score an \"estimate\" of mutual information without proof. The Limitations section even admits that theoretical guarantees remain challenging, so the appendix overclaims. It should be dropped or rewritten as a limitations discussion.\n\nFinally, there is no code, no seeds, no error bars, and several key hyperparameters are unspecified. That makes the central empirical claim non-reproducible as submitted.\n\nWho is this for? People working on path-based KG reasoning and MoE-style pruning. They will find the combination interesting and the ablations useful, if code and corrected results appear.\n\nMy recommendation: this deserves a serious referee rather than a desk reject, but only because the underlying method and the non-YAGO results are worth checking. Ask the authors to release code with a commit hash, move PPR into the main method with recall numbers, fix the efficiency table, and replace Appendix F with an honest limitations paragraph. Without those changes, I would not accept.","headline":"A plausible within-subfield MoE combination for KG reasoning, but the YAGO3-10 headline is confounded by PPR prefiltering and the efficiency table contradicts the text.","tokens_in":26766,"tokens_out":3270,"would_cite":false,"duration_ms":33114,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"MoKGR adapts reasoning depth and pruning per query, beating fixed-depth GNNs on knowledge graph benchmarks.","keywords":["knowledge graph reasoning","mixture of experts","path reasoning","graph neural networks","adaptive path length","pruning","personalized page rank","inductive reasoning"],"falsifier":"Run MoKGR on YAGO3-10 two ways: with its PPR subgraph pre-filtering and with a wider or exact subgraph (or the same PPR budget given to NBFNet and AdaProp). If the MRR gap over baselines shrinks or disappears, then the claimed gains are attributable to PPR retention rather than to expert-based length and pruning adaptation.","tokens_in":25664,"feed_emoji":"🧩","tokens_out":3463,"duration_ms":29790,"temperature":0.7,"pith_summary":"The paper claims that knowledge-graph reasoning is held back by rigid, query-agnostic path exploration: most GNN-based methods propagate messages for a fixed number of hops and prune candidate paths with a single uniform criterion. To fix this, it introduces MoKGR, a mixture-of-experts framework in which a set of length experts selects the path depth that fits each query, and a set of three pruning experts (scoring, attention, semantic similarity) jointly decides which paths to keep. The claimed payoff is two-fold: accuracy improves across transductive and inductive benchmarks, including a large gain on YAGO3-10, and computation drops because the model can stop early on short paths and explore fewer candidates. The paper's central bet is that query-specific path length and query-specific pruning, rather than any single fixed strategy, is what makes reasoning paths informative.","feed_headline":"Adaptive path experts lift knowledge-graph reasoning accuracy","feed_subtitle":"Query-specific depth and pruning beat fixed-depth GNNs on six benchmarks, including a large YAGO3-10 gain.","key_machinery":"A mixture-of-experts gating system layered on top of message-passing path encoders. Length experts assign a weight $g_q(\\ell)$ to each candidate path length so the final entity score is $\\Psi(e_a) = \\sum_{\\ell\\in\\mathcal{A}} g_q(\\ell)\\, s_\\ell(q,e_a)$; pruning experts each define a scoring function $\\phi^\\ell_i(e_a)$ (layer-wise score, max attention over incoming edges, cosine similarity to the query relation), and entities retained at layer $\\ell$ are the union over selected experts of their individual top-$K_\\ell$ sets. A Gumbel-Sigmoid binary gate on each layer's message update lets the model stop early when the path already carries enough evidence, and a PPR-based pre-filtering step is used for very large graphs.","core_discovery":"MoKGR's central claim is that personalizing path exploration to the query improves knowledge-graph reasoning. It does this with two complementary mechanisms: length experts, which use a noisy top-k gating to select a subset of path lengths per query and weight each selected length's score; and pruning experts, which score candidate entities from three complementary views, a global scoring view, a local structural attention view, and a semantic-similarity view, then retain the union of the top entities per selected expert. A layer-wise Gumbel-Sigmoid gate encourages early termination when shorter paths suffice, and this is combined with an adaptive sampling curve that widens exploration in middle layers and narrows it later. The experiments report state-of-the-art or near-state-of-the-art numbers across Family, UMLS, WN18RR, FB15k-237, NELL-995, and YAGO3-10 in the transductive setting and across the inductive versions of WN18RR, FB15k-237, and NELL-995.","pith_inferences":["One testable implication the author leaves implicit: the length experts could be interpreted as an estimator of query difficulty, so their gating weights on a new query could be used to predict which queries are already answerable from a small local subgraph, a cheap pre-filter for any downstream reasoner.","The three pruning experts bear a family resemblance to known complementary signals, global score, local attention, and embedding similarity, and the paper's ablated results suggest that each alone underperforms the mixture; that is evidence for a broader principle that path-quality scoring is multi-dimensional.","A natural falsifying experiment the paper does not run: hold the length experts fixed at a single length and let only pruning experts adapt, versus hold pruning fixed and let only length adapt, on a dataset where true answer distance varies widely like NELL-995, to see which mechanism contributes the majority of the MRR gain.","The reliance on PPR pre-filtering on YAGO3-10 means the headline efficiency numbers should not be read as the cost of the expert mechanism alone; a fair comparison would run a fixed-depth GNN inside the same PPR subgraph."],"forward_implications":["If query-adaptive lengths carry the gain, then fixed-depth GNNs like NBFNet and RED-GNN are over-exploring easy queries and under-exploring hard ones; MoKGR's selective depth should dominate whenever query difficulty is heterogeneous.","If pruning experts are complementary, then the union-of-experts criterion should be able to keep high-recall path sets at smaller size than any single score, transferring to other link-prediction and subgraph-retrieval pipelines.","If PPR pre-filtering plus expert pruning handles YAGO3-10 within one 48GB GPU, then the same recipe should scale to larger KGs such as full YAGO or Wikidata slices without the per-layer blow-up of full-exploration GNNs.","The length-selection signal, query relation embedding plus local structure around the query entity, gives a reusable heuristic for when to stop path expansion, beyond this specific model."],"supporting_citations":[{"why":"RED-GNN supplies the relational directed-graph message passing and the scoring formulation that MoKGR extends with length and pruning experts.","marker":"(Zhang and Yao, 2022)"},{"why":"AdaProp is the node-wise pruning base built upon, the hyperparameter default, and the key pruning baseline MoKGR compares against.","marker":"(Zhang et al., 2023b)"},{"why":"NBFNet is a primary full-exploration fixed-depth baseline providing the path-encoding scoring framework and one of the accuracy baselines in Tables 1 and 6.","marker":"(Zhu et al., 2021)"},{"why":"one-shot-subgraph demonstrates the PPR pre-filtering idea on large KGs and is both the accuracy and efficiency baseline for the YAGO3-10 results.","marker":"(Zhou et al., 2024)"},{"why":"Supplies the noisy top-k gating formulation that both length and pruning expert selection are built on.","marker":"(Shazeer et al., 2017)"},{"why":"Graph MoE is the source of the balance and importance losses, including the coefficient-of-variation regularization MoKGR adopts.","marker":"(Wang et al., 2023a)"},{"why":"GraIL provides the inductive dataset split setup used to build the 12 inductive reasoning subsets.","marker":"(Teru et al., 2020)"},{"why":"Gumbel-Softmax is the basis for the Gumbel-Sigmoid layer-wise gate that enables the early-stopping behavior.","marker":"(Jang et al., 2017)"},{"why":"A*Net is a pruning-based baseline whose path-scoring mechanism MoKGR compares against and also builds on conceptually.","marker":"(Zhu et al., 2023)"}],"fun_headline_variants":["Length and pruning experts adapt KG reasoning per query","Query-specific path depths and pruning boost KG inferencing","Mixture of experts personalizes knowledge-graph reasoning depth","Adaptive experts select path lengths and prune noisy paths","MoKGR: Query-aware length and pruning experts for KGs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The largest-scale results on YAGO3-10 assume the pre-computed Personalized PageRank subgraph contains the true answer entity for every test query; if PPR drops the answer, no later expert can recover it.","fun_headline_variants_meta":{"raw":{"variants":["Length and pruning experts adapt KG reasoning per query","Query-specific path depths and pruning boost KG inferencing","Mixture of experts personalizes knowledge-graph reasoning depth","Adaptive experts select path lengths and prune noisy paths","MoKGR: Query-aware length and pruning experts for KGs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3846,"prompt_tokens":925,"completion_tokens":2921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2842}},"tokens_in":541,"tokens_out":2921,"duration_ms":21351,"temperature":1.0,"reasoning_tokens":2842,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:42:45.928564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MoKGR on YAGO3-10 two ways: with its PPR subgraph pre-filtering and with a wider or exact subgraph (or the same PPR budget given to NBFNet and AdaProp). If the MRR gap over baselines shrinks or disappears, then the claimed gains are attributable to PPR retention rather than to expert-based length and pruning adaptation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NBFNet is a primary full-exploration fixed-depth baseline providing the path-encoding scoring framework and one of the accuracy baselines in Tables 1 and 6."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the noisy top-k gating formulation that both length and pruning expert selection are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gumbel-Softmax is the basis for the Gumbel-Sigmoid layer-wise gate that enables the early-stopping behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A*Net is a pruning-based baseline whose path-scoring mechanism MoKGR compares against and also builds on conceptually."}],"review_version":2}