{"id":"1cbf7658-1ee7-4e33-bd94-69ec927f733b","arxiv_id":"2607.28628","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Hybrid graph-transformer networks guiding A* and beam search find Seiberg-duality paths between ~10-node quivers more efficiently than BFS or pure physics heuristics, with a measured complexity breaking point.","lead":"Neural networks plus pathfinding algorithms can trace chains of Seiberg dualities between quiver gauge theories more efficiently than blind search for modest-sized quivers. The work supplies a concrete benchmark and a practical tool for measuring how hard duality search is in theoretical physics.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The central claim is a carefully scoped empirical statement about relative efficiency/success of GNN-guided hybrid pathfinders versus BFS and LCA on quivers of O(10) nodes. The generation bias noted by the reader is real and load-bearing in principle, yet the authors quantify its effect (path-length agreement ~84 %, absolute node savings still positive for K≥5, OOD infinite-mutation families still favor the hybrid, and an explicit complexity breaking-point study). Public code/checkpoints further allow independent verification. No hidden assumption, algebraic error, or over-claim undermines the reported tables or the usefulness of the benchmark. The reader's ACCEPT / high-confidence verdict is therefore unchanged.","tokens_in":45730,"tokens_out":480,"duration_ms":10800,"concrete_test":"Re-generate a modest ID subset (e.g., C3/Z5 and C(dP1) pairs) after full WL-hash canonicalization of every mutant, recompute true geodesic distances under the automorphism group, retrain DGNN/AGNN on the de-biased labels, and re-evaluate Table 2b EER of Hybrid-LCA vs LCA; if the average EER falls below 1.0 the bias concern lands, otherwise the claim is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption correctly flags that training distances (Eq. 3.5) are upper bounds from BFS path prefixes without full automorphism quotienting, and that this biases pairs toward low-rank LCAs that the deterministic baseline also exploits. The paper itself discloses this (§3.1, §6.1.5–6.2.2, Fig. 24) and still shows Hybrid/Hybrid-LCA gains of ~1.1–1.2\times on ID data plus consistent outperformance on infinite-mutation OOD families (Fig. 38) that were never seen in training. The controlled complexity study (§7) further shows the hybrid remains superior up to ~1.5–2\times training complexity before degrading. These checks already bound the practical impact of the shared-generation bias for the modest-quiver regime claimed. No stronger internal inconsistency or unsupported leap in the central comparative claim is apparent.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies the computational problem of deciding whether two 4d N=1 quiver gauge theories are related by a sequence of Seiberg dualities (quiver mutations) and of finding short duality paths. Starting from toric Calabi–Yau seed quivers, the authors generate mutation trees by BFS, train a Distance GNN (DGNN) to regress mutation distance and an Adviser GNN (AGNN) to propose the next node to dualize, and embed both networks as heuristics/costs inside bidirectional A*, beam search, a physics-inspired Lowest-Common-Ancestor (LCA) rank heuristic, and hybrid combinations. On in-distribution and out-of-distribution quivers with O(10) nodes they report that transformer+MLP architectures, especially Hybrid and Hybrid-LCA pathfinders, substantially outperform unguided BFS and improve on pure LCA in efficiency while retaining near-100% success; a controlled complexity study (C = D log10 K) estimates the hybrids remain superior up to roughly 1.5–2× the training complexity before degrading.","tokens_in":46068,"tokens_out":1228,"duration_ms":34231,"significance":"The work supplies a concrete, reproducible benchmark that sits at the intersection of Seiberg duality, cluster-algebra mutations, and modern graph ML. Strengths that should be credited explicitly include: (i) public checkpoints and pathfinder code, (ii) systematic ID/OOD tables, success heatmaps and efficiency ratios (Tables 2a–2b, §§6–7), (iii) an honest failure-mode section that retrains at lower complexity to locate the breaking point (Eq. 7.3), and (iv) transparent disclosure that training distances are upper bounds from BFS path prefixes and that the dataset is biased toward low-rank common ancestors. If the empirical claims hold under the stated regime, the paper offers both a practical tool for duality searches and a useful stress-test for frontier AI models applied to theoretical physics.","major_comments":[{"comment":"§3.1, Eq. (3.5): duality distance is defined from BFS path-prefix LCAs without full quotienting by quiver automorphisms, so the reported d is an upper bound and the training set is biased toward pairs that share a low-rank common ancestor—the same criterion the LCA baseline exploits. The authors flag this (§6.1.5–6.2.2, Fig. 24) and still show gains on infinite-mutation OOD families never seen in training, but a quantitative bound on the residual bias (e.g., symmetry-reduced distances on a subsample, or ER recomputed after random node relabeling) is needed before the ~1.1–1.2× Hybrid-vs-LCA claim can be treated as fully independent of the generation procedure.","section":"§3.1, Eq. (3.5); §6.2.2"},{"comment":"§5.2 and Table 1: the Hybrid-LCA cost weights (c_det dec/eq/inc, λ_det cost, λ_AR, λ_DGNN, λ_LCA) are stated to have been optimized for 100% success and maximal ER. It is unclear whether this optimization used a held-out validation split distinct from the 500-pair evaluation samples of §6. Without that separation the modest efficiency gains over pure LCA risk mild contamination by the evaluation distribution; a short statement of the hyperparameter protocol (or a re-evaluation with frozen weights chosen only on a validation slice) would remove the ambiguity.","section":"§5.2, Table 1"}],"minor_comments":[{"comment":"§2.2: typographical slip “the far more generic cas is” → “case”.","section":"§2.2"},{"comment":"Fig. 8c and related MAE plots: the median error is negative (ˆd < d). A one-sentence remark that the network systematically under-estimates large distances (consistent with the non-monotonicity discussion) would help the reader interpret the heuristic quality.","section":"§4.1.1, Fig. 8"},{"comment":"Appendix A.2 / Fig. 37: several quivers are anomalous as 4d N=1 theories. The text already notes they remain valid as BPS quivers; a brief cross-reference in the main OOD discussion (§6.2) would prevent confusion.","section":"§6.2, Appendix A.2"},{"comment":"The complexity measure C = D log10 K is introduced only in §7. A forward pointer in the introduction or §5.3 would make the later breaking-point analysis easier to anticipate.","section":"§7"},{"comment":"References: the concurrent DualityCert work [25] is noted; a one-sentence clarification of the complementary goals (path-finding/complexity vs. verifier-gated claim repair) already present in the Note Added could be moved into the introduction for readers who skip the note.","section":"Note Added / §1"}],"recommendation":"minor_revision","confidential_remarks":"The central comparative claim is empirically well-supported within the modest-quiver regime the authors advertise; the two major points are clarifications rather than structural flaws. Fit for hep-th is good—the physics motivation (duality cascades, warped throats, computational complexity of dual phases) is genuine and the ML is used as a tool rather than an end in itself. I see no citation or novelty issues that would require editorial intervention."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a practical, well-executed paper that builds GNN pathfinders (DGNN distance estimator + AGNN policy) and hybrids them with A*/beam search and a physics-informed lowest-common-ancestor rank heuristic to find Seiberg-duality sequences on quivers with ~10 nodes. The new piece is the first systematic treatment of the generic infinite-mutation case (prior ML work stayed with finite-mutation type), plus concrete efficiency numbers versus unguided BFS and a deterministic LCA baseline, plus a controlled breaking-point study that retrains at lower complexity and shows the hybrids stay ahead out to roughly 1.5–2× training complexity.\n\nWhat they do well: the ID/OOD tables, success heatmaps, EER/GE metrics, and §7 failure-mode analysis are thorough and transparent. They ship checkpoints and code. They flag the main structural caveat themselves—training distances come from BFS path prefixes (Eq. 3.5) that are upper bounds without full automorphism quotienting and that bias pairs toward low-rank common ancestors, the same logic LCA exploits. Even so, the hybrids still show ~1.1–1.2× gains on ID data and hold up on infinite-mutation OOD families never seen in training. Non-monotonicity of the DGNN heuristic (~32%) and AGNN inversions are measured and mitigated by the hybrid design rather than papered over. Citations to the finite-mutation ML papers and learning-to-unknot are appropriate; the math of mutations is standard and correctly used.\n\nSoft spots are real but proportionate: the shared generation logic between labels and the strongest baseline means the “outperformance of LCA” contest is partly on home turf, the free hyperparameters (Hybrid-LCA weights, beam width, architecture sizes) are tuned rather than derived, and everything is empirical for modest K. None of that breaks the central comparative claim for the regime they advertise.\n\nThis is for people who actually need to enumerate dual phases or want a clean AI-for-physics benchmark. It deserves a serious referee. I would engage with it and expect to cite the efficiency numbers and the public models.","headline":"Solid empirical ML tool for tracing Seiberg dualities on modest quivers; hybrids beat BFS and LCA with public code and honest failure analysis.","tokens_in":46687,"tokens_out":533,"would_cite":true,"duration_ms":11142,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Neural networks plus pathfinders find Seiberg-duality chains for modest quivers faster than blind or pure physics search.","keywords":["Seiberg duality","quiver mutations","graph neural networks","pathfinding","computational complexity","supersymmetric gauge theories","A* search","toric Calabi-Yau"],"falsifier":"Retrain the same architectures on a dataset that fully quotients by quiver automorphisms and re-measures efficiency ratios against LCA; if the hybrid advantage disappears or success rates fall below the pure LCA baseline at comparable complexity, the claimed outperformance is an artifact of the biased distance definition.","tokens_in":46595,"feed_emoji":"🔀","tokens_out":891,"duration_ms":17695,"temperature":0.7,"pith_summary":"Given two supersymmetric quiver gauge theories, deciding whether they are related by a chain of Seiberg dualities (quiver mutations) and finding a short chain is combinatorially hard once the number of nodes grows. This paper trains graph networks with transformers on mutation trees grown from D-brane seed theories, then uses the networks as distance estimators and next-move advisers inside A* and beam-search pathfinders. For quivers of roughly ten nodes the hybrid searchers outperform both unguided breadth-first search and a pure rank-minimizing physics heuristic while keeping near-perfect success rates. The same machinery yields a practical complexity measure and a benchmark task for AI applied to theoretical physics.","feed_headline":"AI pathfinders beat pure physics search on Seiberg dualities","feed_subtitle":"Hybrid networks plus rank heuristics find short duality chains for ~10-node quivers faster than blind or LCA baselines","key_machinery":"Hybrid pathfinders that combine a Distance GNN (heuristic estimating mutation distance), an Adviser GNN (policy over which node to dualize next), and a physics-informed Lowest-Common-Ancestor cost that prefers rank-reducing moves; the networks are trained on BFS-generated mutation trees from toric Calabi–Yau seed quivers.","core_discovery":"For quivers with a modest number of nodes (order 10), transformer-plus-MLP graph networks used as policies inside bidirectional A* and beam pathfinders, especially when hybridized with a lowest-common-ancestor rank heuristic, find Seiberg-duality paths more efficiently than unguided BFS or pure deterministic LCA search while maintaining near-100% success.","pith_inferences":["The same mutation-pathfinding setup could be reused for other cluster-algebra or BPS-quiver problems where the mutation graph is infinite and non-monotonic heuristics appear.","Once automorphism-aware distances are available, the residual gap between Hybrid LCA and pure LCA would quantify how much genuine pattern learning the networks have acquired beyond the rank-minimizing bias built into the training trees.","Scaling the training set to larger node counts would test whether the observed complexity threshold (roughly 1.5–2× training C) continues to hold or collapses, giving a concrete scaling law for this class of physics search tasks."],"forward_implications":["Computational complexity of duality chains can be read off as C = D log10 K and used to estimate how far a holographic RG flow proceeds down a warped throat.","The same hybrid pathfinders give a practical tool for deciding whether two given quivers are Seiberg dual and for enumerating short connecting sequences.","The task supplies a concrete benchmark for frontier AI models on a well-defined theoretical-physics search problem.","Hybrid NN-plus-physics search remains superior to pure LCA up to roughly 1.5–2 times the training complexity before degrading."],"fun_headline_variants":["Hybrid networks find Seiberg duality paths faster than BFS on ~10-node quivers","Transformers plus MLPs beat pure LCA search for quiver mutations","AI pathfinders with rank heuristics trace Seiberg dualities more efficiently","Graph networks in A* outperform unguided search on Seiberg dualities","ML policies plus LCA heuristics shorten duality chains for modest quivers"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"Distances used for training and scoring are taken from BFS trees via longest common path prefixes and are only upper bounds once quiver symmetries are ignored, biasing the data toward pairs that share a low-rank common ancestor—the same criterion the physics baseline exploits.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid networks find Seiberg duality paths faster than BFS on ~10-node quivers","Transformers plus MLPs beat pure LCA search for quiver mutations","AI pathfinders with rank heuristics trace Seiberg dualities more efficiently","Graph networks in A* outperform unguided search on Seiberg dualities","ML policies plus LCA heuristics shorten duality chains for modest quivers"]},"model":"grok-4.5","effort":"low","cost_usd":0.00525,"raw_usage":{"total_tokens":1442,"prompt_tokens":796,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":52504000,"prompt_tokens_details":{"text_tokens":796,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":545,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":796,"tokens_out":101,"duration_ms":10390,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T01:43:46.750012+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the same architectures on a dataset that fully quotients by quiver automorphisms and re-measures efficiency ratios against LCA; if the hybrid advantage disappears or success rates fall below the pure LCA baseline at comparable complexity, the claimed outperformance is an artifact of the biased distance definition.","supporting_citations":[],"review_version":1}