{"id":"b0431f71-7e9b-44ac-bdc8-9c97900dcbc0","arxiv_id":"2502.09544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FunSearch discovered a simple priority function for ordering IBP seeding integrals, reducing the number needed for multi-loop Feynman integral reductions by factors up to 3058.","lead":"Physicists used an AI search algorithm to discover a simple scoring rule that says which integration-by-parts equations should be generated first when reducing Feynman integrals, cutting the number of required equations by up to thousands of times. The heuristic makes some multi-loop calculations for particle colliders feasible that previously exceeded 400GB of memory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 3058x improvement is not isolated from the category-selection algorithm; no baseline with Laporta/Box/random priority under the same selection protocol is shown.","rationale":"The reader's verdict is CONDITIONAL, and I agree that no fatal internal error is visible. However, the most load-bearing weakness is not only external transferability of the one-loop bubble search; it is internal attribution. The measured reduction is produced by the full pipeline, F0 ordering plus M/G category selection plus binary/time search, and the paper provides no control where the selection component is held fixed and only the priority is varied. Since the baseline 'improved seeding' is a complete fixed set, the selection step alone can be expected to shrink it; without Laporta/Box/random controls we cannot tell whether F0 is essential or whether the selected subsets are just smaller because any subset search finds enough equations. This is concrete and easy to test, and it directly affects the headline 3058x and 1060x numbers. The reader's representativeness concern remains valid and is reinforced by Appendix B, where FunSearch on two-loop IBP did not surpass Box priority; but even before generalization, the two-loop evidence needs the ablation. I therefore keep the reader's CONDITIONAL verdict unchanged, with the additional requested check being a priority-function ablation.","tokens_in":21824,"tokens_out":10640,"duration_ms":105223,"concrete_test":"Rerun Table I row d5 (planar family) and Table II row d5 (non-planar) with the Sec. VI search algorithm unchanged, but replace F0 with (a) Laporta priority Eq. (14), (b) Box priority from Listing 3, and (c) random category order, using the same LiteRed/FiniteFlow pipeline. If any baseline yields a minimal seed count within roughly 2x of the F0 result (7109 planar, 10859 non-planar), the reported improvement is not attributable to F0; if all baselines fail or require orders of magnitude more seeds, the attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim, Tables I/II and the abstract's 3058x factor for d5, is attributed to the priority function F0 in Eq. (16), but the experiments compare F0 plus the category-selection/search algorithm of Sec. VI against the full improved-seeding set. They never run the same selection algorithm with a different ordering, such as Laporta priority Eq. (14), Box priority from Listing 3, or a random order. Because any complete seeding set contains many redundant equations, a subset-selection procedure can reduce 21,740,796 seeds to 7,109 even with a generic ordering; the giant ratio may therefore measure the per-target subset search rather than the discovered F0. This is load-bearing because the paper's stated contribution is F0 as a best-estimated priority function, and Appendix B concedes that FunSearch itself did not beat Box priority on two-loop problems, so human generalization plus selection is doing the work. Without an ablation isolating the priority, the 3058x and 1060x factors do not establish F0's efficacy.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to reduce the size of integration-by-parts (IBP) systems by ordering seeding integrals with a priority function discovered by the FunSearch algorithm on a one-loop massless bubble integral, then generalized by human experts to the Ellipse-type function F0 in Eq. (16). The authors apply F0, together with a category-based subset-selection algorithm, to planar and non-planar five-loop six-particle phase-space integral families. Tables I and II report large reductions in the number of seeding integrals, up to a factor of 3058 for a planar integral (d5) and 1060 for a non-planar integral (d5), compared to an improved-seeding baseline, and the paper claims these reductions make previously infeasible IBP reductions tractable.","tokens_in":22112,"tokens_out":4761,"duration_ms":46465,"significance":"If the reported reductions are genuinely attributable to the discovered priority function F0, this would be a practically valuable and interpretable heuristic for one of the main computational bottlenecks in multi-loop amplitude calculations. The combination of FunSearch on a simple problem with human transfer to harder problems is a promising methodology, and the target integrals (five-loop, six-particle phase-space, up to 46 master integrals) are nontrivial and of direct interest. The paper's strengths are the explicit, closed-form priority function (Eq. (16)), the use of FiniteFlow for the main comparisons, and the clear statement of the algorithm in Sec. VI. However, as detailed below, the current evidence does not isolate the effect of the priority function from the subset-selection procedure, so the quantitative claims require additional experiments before they can be fully credited.","major_comments":[{"comment":"The comparison labeled 'Priority function F0' in Tables I and II includes the full category-selection and search algorithm described in Sec. VI (steps 1–5, Eq. (17), with relaxation factor a=1/3), whereas the 'Improved seeding' baseline is the full minimal seeding set without this selection. The paper never runs the same selection algorithm with a different ordering, such as the Laporta priority of Eq. (14), the Box priority of Listing 3, or a random order. Because a complete seeding set contains many redundant equations, a subset-selection procedure can in principle reduce 21,740,796 seeds to 7,109 even with a generic ordering. Thus the reported factors of 3058 (planar d5) and 1060 (non-planar d5) do not currently establish the efficacy of Eq. (16) in isolation; they measure the combined pipeline. Please add an ablation that applies the identical Sec. VI search algorithm with alternative priority functions, including Box and Laporta, for at least the d5 cases in both families.","section":"Sec. VI, Tables I and II"},{"comment":"The priority function F0 in Eq. (16) is not itself the FunSearch output; it is a human generalization (with m chosen as 2) of an evolved one-loop function. Section IV states 'we hope to obtain a general best-estimated priority function under a simpler IBP problem and naturally extend it', and Appendix B concedes that on two-loop problems FunSearch did not beat Box priority. The multi-loop results therefore depend critically on the human transfer step, and the only direct evidence that Eq. (16) outperforms Box priority is the one-loop factor-of-five improvement in Sec. V (with a single reported target, I(15,10)). To support the central claim that F0 is a 'best-estimated priority function', the authors should compare Eq. (16) against Box priority and Laporta priority on the same multi-loop benchmarks using the same selection protocol.","section":"Sec. IV–V, Appendix B"},{"comment":"No code or data are shipped, and the exact procedure for constructing the categories used in Sec. VI is not specified beyond 'O(10^2) categories are produced for O(10^5) seeding integrals'. The 'improved seeding' minimal pattern search is also not described algorithmically. Timings in Tables I and II are single numerical sample points with no run-to-run variation or detailed machine characterization. For a computational paper whose central contribution is an algorithm, this is insufficient for independent verification. Please provide the implementation, or at least the exact category-generation rules, the optimized seeding vectors, and the binary-search/relaxation details, together with at least a few repeated timing runs for one representative case.","section":"General reproducibility (Secs. IV–VI)"}],"minor_comments":[{"comment":"The notation a/b and a/b/c is explained, but the columns NI/NP and TI/TP are not defined explicitly in the table captions; clarify that these are ratios of seed counts and timings between the F0 column and the improved-seeding column.","section":"Table I and Table II captions"},{"comment":"The quantities Ntot and Nm are defined in text, but the relation to the 'first M categories' is ambiguous when multiple sets of M categories could span all master integrals; state whether M is chosen by a deterministic rule and how ties are broken.","section":"Sec. VI, Eq. (17)"},{"comment":"The sentence claiming the d5 reduction 'can now be performed on a standard laptop in under two minutes' appears inconsistent with Table I, which reports 47.7 seconds to learn and 2.24 seconds to solve on a Xeon Gold 6148 with 768 GB RAM; please reconcile or remove the laptop claim.","section":"Sec. VI.A, last paragraph"},{"comment":"The text says 'the reduction including the reconstruction of D can now be performed...', but no details of the reconstruction step are given; define D and the reconstruction procedure, or cite the FiniteFlow description.","section":"Sec. VI.A, after Table I"},{"comment":"Figures 2 and 4 appear to contain garbled or mis-encoded text that is not readable; please regenerate these figures with proper font handling.","section":"Appendix A and Fig. 2 and Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is interesting and the target problems are significant, but the main quantitative claim is currently an attribution problem: the reported improvements are for a combined algorithm, not for the priority function alone. The requested ablation (same selection with Box/Laporta/random priorities) is straightforward and should be added before publication. I also note the overlap with Ref. [85], cited by the authors; the distinction in the Note Added is clear, but the referee report should not influence that. The manuscript fits the journal's scope as a computational hep-ph method paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper claims a learned priority function F0 (Eq. 16) that reduces IBP seeding integrals by up to 3058x on five-loop phase-space families. The core idea is clean: evolve a priority function with FunSearch on a simple one-loop bubble, then have humans generalize it to a closed form — an ellipse-type distance to the target plus a norm. That is genuinely new, and the multi-loop benchmarks (Tables I and II) are impressive: d5 planar goes from 21.7M seeds to 7,109, and previously OOM cases become laptop-solvable. The paper is also honest about what FunSearch did not do — Appendix B concedes it didn't beat Box priority on two-loop problems — and it carefully distinguishes seed-count optimization from solve-time optimization.\n\nThe soft spot is load-bearing. The comparisons always run F0 as the ordering inside their newly introduced category-selection/search algorithm, versus improved seeding (or FIRE6) without that selection. Never do they run Laporta priority, Box priority, or a random order through the same selection protocol. Since any complete seed set contains huge redundancy, a subset-search can shrink it dramatically even with a generic ordering; the 3058x factor may therefore measure the search, not F0. The abstract's wording ('compared to traditional methods') is also a bit loose: the 3058x is against the improved-seeding baseline, not Laporta seeding. Minor points: no code or data shipped, timings are single numerical sample points, and the a=1/3 relaxation factor is a human choice whose robustness is untested.\n\nIf a referee pushes for the ablation — same selection protocol with F0 vs Laporta/Box/random priorities — the paper's central claim will be either confirmed or reassigned. That is worth doing, because the problem matters and the paper is a serious attempt to make AI assistance explainable in this area.\n\nVerdict: send to peer review, but with a clear request for the ablation and, ideally, a release of the seed sets or code. For my own reading group, it is a good 'maybe' — the discussion of what exactly the LLM/generic algorithm contributed is instructive.\n\nBest,\n[Your name]","headline":"Useful and honest, but the headline improvement is not isolated from the category-selection algorithm; needs an ablation before the priority function gets credit.","tokens_in":22568,"tokens_out":2788,"would_cite":false,"duration_ms":24944,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An AI-designed ordering rule cuts the seed count in Feynman integral reduction by up to a factor of 3,058.","keywords":["Feynman integrals","integration-by-parts reduction","Laporta algorithm","priority function","FunSearch","large language models","genetic algorithms","phase-space integrals"],"falsifier":"Run $F_0$ with $m = 2$ on a multi-loop family not used in the paper — say a different non-planar five-loop phase-space family or a six-loop propagator integral with several dotted lines and numerators — and compare the required seed count and peak memory against the best improved-seeding scheme for that family; if the improvement factor collapses toward one, or the solver fails where improved seeding succeeds, the claimed transferability is refuted. A more direct check: evolve a priority function with FunSearch on a genuine two-loop IBP problem; if the best evolved function never beats the Ellipse-type $F_0$ carried over from one loop, the scalability argument loses its evidentiary base.","tokens_in":21702,"feed_emoji":"⚛️","tokens_out":12025,"duration_ms":99443,"temperature":0.7,"pith_summary":"This paper claims that the computationally hard part of Feynman integral evaluation — integration-by-parts (IBP) reduction, which solves large linear systems generated from 'seeding' integrals — can be made much cheaper by sorting the seeds with one explicit formula. The formula is a priority function that ranks each candidate seed by its distance from the origin in index space plus its distance from the target integral; an AI search called FunSearch discovered its shape on the simplest one-loop case, and the authors generalized it by hand to arbitrary numbers of loops and legs. Applied to five-loop six-particle phase-space integral families, the priority function cuts the required number of seeds by up to a factor of 3,058, turning reductions that previously needed more than 400 GB of memory into runs that finish on a laptop in under two minutes. If the transfer holds, IBP reduction stops being the main bottleneck for a wide class of higher-order calculations.","feed_headline":"One formula cuts Feynman reduction work by up to 3,058x","feed_subtitle":"An AI-found ordering rule trims seeding integrals, turning out-of-memory five-loop runs into laptop tasks.","key_machinery":"The load-bearing object is the Ellipse-type priority function $F_0$ of Eq. (16), which assigns each candidate seeding integral a rank equal to minus the sum of two geometric quantities in index space: an $\\ell^m$ norm of its propagator indices (a measure of how many dots and numerators it carries) and the Euclidean distance from the seeding point to the target integral. The name records its discovery path: on the one-loop massless bubble problem, FunSearch — a search that evolves short programs through a large language model combined with genetic algorithms — first reproduced the standard Laporta priority (a diagonal line in index space), then a Box priority, then converged to the Ellipse-type form, which the authors read off from the evolved code and generalized to any loop and leg count. Functionally, $F_0$ defines a nested chain of subsets $S_i$ of the seeding set, so the smallest subset sufficient to reduce the target can be located by binary search; the multi-loop applications refine this by sorting seeds into categories and selecting the first $M$ and $G$ categories with a relaxation factor $a = 1/3$.","core_discovery":"The paper's central claim is that the priority function $F_0$ of Eq. (16), $$F_0(I(n_i); I_{\\rm target}(t_i)) = -\\left[\\left(\\sum_i |n_i|^m\\right)^{1/m} + \\left(\\sum_i (n_i - t_i)^2\\right)^{1/2}\\right],$$ with $m \\in \\{1,2,4,\\infty\\}$ and the first term read as $\\max_i |n_i|$ for $m = \\infty$, is a near-universal way to order IBP seeding integrals: it penalizes seeds that carry dots or numerators (large $\\ell^m$ norm) and seeds far from the target integral in index space (large Euclidean distance). Ranked by $F_0$, the seeds generated by conventional schemes arrange into nested subsets, and a binary search locates the smallest subset that still solves the target; memory use is roughly proportional to the size of that subset. On the one-loop problem where the function was evolved, the required subset shrinks by about a factor of five; on the five-loop planar and non-planar families the seed count falls by up to 24.8 times for the hardest reachable top-sector integrals and by 3,058 times (planar) and 1,060 times (non-planar) for sub-sector integrals with many dots and numerators, converting out-of-memory failures into fast laptop-scale reductions.","pith_inferences":["If the one-loop-to-five-loop transfer is as general as the paper claims, the same recipe — evolve a priority function on a toy problem with FunSearch, then have an expert generalize the closed form — could be applied to other combinatorial bottlenecks in amplitude computation, such as seed selection for syzygy-based or block-triangular reduction schemes.","The paper optimizes seed count rather than solve time, and says so; a priority function tuned directly against equation-solving time could produce further speedups beyond the reported ratios, since the fastest seed set and the smallest seed set need not coincide.","The appendix's multi-target variant of Eq. (B1) was only shown to be at least as good as the single-target Ellipse priority; comparing it directly with the recursive single-target strategy on the 955-target problem is a cheap, natural next experiment.","Only the $m=2$ variant was used in the multi-loop benchmarks, so the sensitivity of the reported gains to the norm choice ($m=1,4,\\infty$) is untested; applying those variants to a hard non-planar family would show whether the improvement is robust to that choice."],"forward_implications":["The five-loop planar top-sector integrals $s_{10}$ and $s_{12}$, which the improved-seeding baseline cannot complete within 400 GB of memory, become computable, and the planar $d_5$ reduction finishes in under two minutes on a standard laptop.","The improvement factor grows with target complexity across the tested range (about 1.0x for the simplest planar top-sector seeds up to 24.8x for $s_{12}$ and 3,058x for $d_5$), so the method pays off most where previous algorithms failed outright.","The same hand-generalized priority function works for planar and non-planar integral families, for single-target reductions, and for a 955-target dimensional-recurrence problem in which it halves the seed count relative to improved seeding.","Because $F_0$ is a short, interpretable expression rather than a trained black box, it can be ported to new integral families and refined by human experts without rerunning the search."],"supporting_citations":[{"why":"Defines the Laporta algorithm and its seeding scheme, the standard baseline whose seed count the priority function must beat.","marker":"[6]"},{"why":"Introduces FunSearch, the large-language-model-plus-genetic-algorithm search used to discover the priority function's shape.","marker":"[34]"},{"why":"Provides the improved-seeding strategy used as the comparison baseline in the multi-loop examples.","marker":"[19]"},{"why":"FiniteFlow supplies the sparse linear-system solver used in all multi-loop reductions with priority-function seeds.","marker":"[31]"},{"why":"LiteRed generates the symbolic IBP identities from which the reduction systems are built.","marker":"[21, 22]"},{"why":"FIRE6 is the representative Laporta-style package whose seeding counts and timings form the comparison column of Table I.","marker":"[23, 24]"}],"fun_headline_variants":["AI-found priority function cuts Feynman reduction seeding 3058x","Evolved ordering rule makes five-loop Feynman integrals laptop-runnable","Explainable AI ranker slashes memory in Feynman integral reduction","Single formula from FunSearch trims Feynman seeds to 1/3058","New seed-ordering function reduces Feynman work by over 3000x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two-dimensional index space of the one-loop massless bubble is representative enough that a priority function optimized there, after human generalization, still works for five-loop six-particle phase-space families with 20 propagators and up to 46 master integrals — a transfer the paper asserts rather than proves, and one its own appendix undercuts slightly by reporting that running FunSearch directly on two-loop problems found nothing better than the Box priority.","fun_headline_variants_meta":{"raw":{"variants":["AI-found priority function cuts Feynman reduction seeding 3058x","Evolved ordering rule makes five-loop Feynman integrals laptop-runnable","Explainable AI ranker slashes memory in Feynman integral reduction","Single formula from FunSearch trims Feynman seeds to 1/3058","New seed-ordering function reduces Feynman work by over 3000x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000842,"raw_usage":{"total_tokens":3687,"prompt_tokens":986,"completion_tokens":2701,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2600}},"tokens_in":602,"tokens_out":2701,"duration_ms":19365,"temperature":1.0,"reasoning_tokens":2600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:06:09.952862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run $F_0$ with $m = 2$ on a multi-loop family not used in the paper — say a different non-planar five-loop phase-space family or a six-loop propagator integral with several dotted lines and numerators — and compare the required seed count and peak memory against the best improved-seeding scheme for that family; if the improvement factor collapses toward one, or the solver fails where improved seeding succeeds, the claimed transferability is refuted. A more direct check: evolve a priority function with FunSearch on a genuine two-loop IBP problem; if the best evolved function never beats the Ellipse-type $F_0$ carried over from one loop, the scalability argument loses its evidentiary base.","supporting_citations":[],"review_version":1}