Pith. sign in

REVIEW 3 major objections 5 minor 54 references

This paper establishes that the one-loop five-gluon BCJ numerator problem is exactly solved by a 920-rank cut matrix whose 207-dimensional kernel is annihilated by the specified color-ring observable map, making the entire fiber one observa

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:15 UTC pith:FSANTBEV

load-bearing objection Five-point linear algebra is solid; the S·K=0 certificate is load-bearing and not yet publicly verifiable. the 3 major comments →

arxiv 2607.13412 v1 pith:FSANTBEV submitted 2026-07-15 hep-th

An adaptive inverse-problem framework for one-loop five-gluon BCJ numerators

Lin Mai , Yaobo Zhang This is my paper
classification hep-th
keywords BCJ numeratorscolor–kinematics dualityone-loop five-gluon amplitudegeneralized unitarity cutsinverse problemnumerator ambiguitydimensional regularizationcolor-ring observable
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that Bern–Carrasco–Johansson (BCJ) numerator construction can be posed as a finite, exact inverse problem whose solution is fully characterized by two null spaces: one that flags missing ansatz directions and one that measures which cut equations still add information. For the one-loop five-gluon pure Yang–Mills case, it shows that after kinematic and graph-symmetry reduction the candidate numerator space has 1127 coefficients, generalized maximal, box, triple, and double cuts have rank 920, and the remaining 207-dimensional affine fiber of valid numerators is collapsed to a single observable class by the specified R12345 color-ring readout. If correct, the paper has completely classified the one-loop five-gluon BCJ numerator ambiguity: every valid numerator choice differs only by cut-invisible deformations that do not change the stated observable. This would matter because it turns the notoriously representation-dependent BCJ numerator problem into a deterministic, reproducible linear-algebra construction that can be adapted to other amplitudes.

Core claim

The central claim is Theorem 7.2 together with Corollary 7.10: for one-loop five-gluon pure Yang–Mills theory, the cut matrix satisfies rank A = 920, dim ker A = 207, and the specified R12345 color-ring observable map S annihilates every kernel direction, S·K = 0. This means the complete affine fiber of numerator solutions, xp + Kβ(Ds), maps to a single equivalence class under S. The paper constructs the fiber exactly, identifies a published forward-limit numerator as a point in it with 15 active kernel directions, and validates the result on independent cuts, integral reductions, and helicity amplitudes. It also documents that the raw fixed-routing and forward-limit prescription maps differ

What carries the argument

The carrying mechanism is the compiled inverse-problem tuple Pσ = (A, b; S; T). The matrix A maps the 1127 symmetry-reduced numerator coefficients to generalized-cut data; its left null space identifies inconsistencies that new ansatz columns must repair, while its right kernel K spans coefficient changes invisible to the current cuts. The observable map S projects through the R12345 color-ring coefficient, a canonical loop-momentum routing table, an LSZ projector, and stated scaleless-integral identities. The pivotal identity is S(ka) = 0 for all 207 kernel vectors, equivalently ker A ⊆ ker S, which collapses the 207-dimensional fiber to one S-class.

Load-bearing premise

The one-class conclusion rests on the claim that after the prescribed canonical loop-momentum shifts all 296 categorized scaleless integral sources are genuinely scale-free and therefore exactly zero in dimensional regularization; if any of those sources carries a hidden external-scale dependence, some kernel direction would survive the observable map and the fiber would split into multiple observable classes.

What would settle it

Compute one of the 296 route-scaleless sources, say an adjacent massless external bubble, after its stated canonical loop shift at a generic five-point kinematic point and check whether the shifted integral is identically zero as a function of the external invariants; alternatively, directly evaluate S(ka) for a single kernel direction after a non-canonical loop-momentum shift to see whether it vanishes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If S·K = 0 is correct, every valid one-loop five-gluon BCJ numerator reconstructed from these cuts gives the same R12345 color-ring observable after LSZ and dimensional-regularization rules.
  • The 207-dimensional numerator fiber is a concrete measure of BCJ representation freedom: the cut data fix all observable content, while the kernel vectors parameterize harmless reshufflings of contact terms and loop-momentum shifts.
  • The same diagnostics transfer to other amplitudes: left-null obstructions certify when an ansatz must be enlarged, and the right-kernel visibility rank tells which cut topology adds genuinely new information.
  • The paper's nonzero difference between the raw fixed-routing and forward-limit prescription maps means that the one-class statement does not automatically extend to integrated amplitudes; a separately specified surface, external-field, and scheme completion is needed.
  • The adaptive loop provides a deterministic interface where a proposed basis change, cut addition, or validation step is compiled into an exact finite problem and evaluated before being accepted.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the one-class result is likely sensitive to the choice of canonical routing, and the same fiber-observable separation may fail for other color-ring projections or for observables that keep loop-momentum information, suggesting the equivalence is tied to the paper's specific S.
  • Inference: the rank-920 to nullity-207 split offers a quantitative 'BCJ representability index' for other loop orders and multiplicities—the gap between symmetry-reduced ansatz dimension and cut rank is exactly the freedom that S must classify.
  • Inference: a testable extension is to repeat the kernel classification for parity-odd numerator sectors or for N=4 super-Yang–Mills, where the dimension-shift and scaleless-integral identities differ; the framework predicts the analogue of S·K = 0 will hold or fail depending on the chosen observable.
  • Inference: the nonzero master-integral difference between prescription maps suggests that, beyond the paper's point, even a single equivalence class under S can branch into multiple integrated classes under different surface-term choices, so the relation between S-equivalence and integration-equivalence deserves explicit study.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper develops an adaptive inverse-problem framework for constructing BCJ numerators. The central case study is the one-loop five-gluon pure-Yang-Mills amplitude. A 1127-dimensional parity-even local numerator basis is built, generalized cuts (maximal, box, triple, double) give rank 920 and a 207-dimensional affine solution fiber, and a specified color-ring observable map S is claimed to annihilate the entire kernel, so the fiber is a single observable class. A published forward-limit numerator is located inside the fiber, and the result is checked against fresh cuts, independent integral reductions, and standard helicity amplitudes. The paper also proposes an agent-compatible workflow with exact diagnostics for basis and measurement revision.

Significance. If correct, the paper gives the first complete characterization of the coefficient ambiguity for one-loop five-gluon BCJ numerators, and it demonstrates a reproducible exact-algebra pipeline for amplitude-representation inverse problems. The strengths are real: exact finite-field computation with two independent solvers, two tree-amplitude implementations, 43,056 kernel residual checks, standard amplitude benchmarks, and a detailed release manifest with compact sparse vectors. The main significance, however, rests on the untested scaleless-source classification used in Theorem 7.9; the paper's own release boundary defers the per-direction term ledgers to a future v0.2 archive. There is also a tension between the declared parity-even ansatz and the reported tr5 coefficient in the all-plus benchmark. Both issues are load-bearing and require resolution before the central claim can be accepted.

major comments (3)
  1. [§7.4, Eqs. (7.22)–(7.24); App. F/G] The proof of Theorem 7.9 and Corollary 7.10 depends on the exhaustive statement that 296 of 646 kernel conditions vanish as 'route scaleless' sources, split as 21 polynomials, 175 massless tadpoles, and 100 adjacent external bubbles. This is the load-bearing part of ker A ⊆ ker S. However, the per-direction term ledgers and routing tables that would verify this enumeration are explicitly deferred to the planned v0.2 payload; v0.1 contains only a global Stage-S summary certificate. A single external-scale-dependent source after the canonical shift would break S(k_a)=0 and produce |F_b/∼S| > 1. Please provide the complete source-level enumeration and the explicit route shifts, or release the v0.2 production data, so that the 296-source classification is independently checkable.
  2. [§8.2, Eq. (8.10), with §5.1] The ansatz of Sec. 5.1 is declared parity-even, containing only parity-even scalar contractions and excluding Levi–Civita structures. Yet the all-plus benchmark in Eq. (8.10) reports a nonzero coefficient for the parity-odd invariant tr5 in the basis (sum s_i^2, sum s_i s_{i+1}, sum s_i s_{i+2}, tr5). A parity-even integrand cannot produce a linear tr5 term in the integrated amplitude. As written this appears internally inconsistent. Please state explicitly how a tr5 coefficient can arise from the parity-even ansatz, or correct the reported coefficient; if tr5 arises from the published forward-limit representative, reconcile that with the claim that this representative lies in the parity-even affine fiber.
  3. [§7.2, Cor. 7.4] The membership of the published forward-limit numerator in the fiber is stated via Eq. (7.12), with 15 active kernel directions. The kernel vectors are combinations of the parity-even ansatz basis. If the published numerator [12] contains parity-odd structures—which is suggested by the tr5 benchmark—the equality x_FL = x_p + K(α0 + D_s α1) would be impossible in the declared coordinate space. Please verify that every raw tensor entering the comparison is within the parity-even basis, or specify which additional sector is being used. This is a consistency check on the main reconstruction claim, not just a benchmark detail.
minor comments (5)
  1. [Throughout] The text uses 'parity-even' in multiple senses (Lorentz parity vs. graph reflection parity). Define the term at first use and use distinct names for the two notions.
  2. [Abstract and App. F] The compact v0.1/v0.2 boundary should be disclosed in the abstract, since the central certificate is not yet in the public release; the current abstract overstates what is reproducible from the distributed material.
  3. [Eq. (2.29) and surrounding text] A few displayed equations have spacing/formatting artifacts, and §9 contains a run-together phrase ('formulaisthereforeacompactanalyticrepresentative'). These should be corrected in the final version.
  4. [§8.2] The paper notes that adjacent-collinear and soft-limit factorization checks are pending. That is acceptable, but the wording in §9 ('all held-out residuals vanish') should be qualified to exclude the explicitly pending factorization targets.
  5. [§4.3 / §5.1] The claim that 300 omitted repeated-contact directions were discovered by exact diagnostics would be easier to verify if the corresponding left-null witness or the exact dimension count of the square-free quotient were listed in a table; the present text states the count but not the specific witness.

Circularity Check

0 steps flagged

No significant circularity: the cut fiber is solved from Ax=b alone, S·K=0 is a conditional computation with deferred ledgers, and the same-author forward-limit numerator enters only as post-hoc validation.

full rationale

The derivation chain is not circular. The affine fiber is obtained by exact solution of Ax=b in the fixed 1127-coordinate basis, with rows built from maximal/box/triple/double cuts and targets computed independently from tree amplitudes; neither the published numerator nor the observable map S enters the solve. The claim S K=0 is then a computation on the already-determined kernel: S is fixed in the specification as the R12345 projection followed by LSZ amputation and dimensional-regularization scaleless-integral identities (eqs. (7.20)-(7.24)), and Theorem 7.9 reduces S(k_a) to 350 exact coefficient identities and 296 route-scaleless sources. The only fragile point is that the 296-source classification is asserted without per-direction ledgers; App. F explicitly defers those to v0.2. That is an evidence/reproducibility gap of the type that could invalidate Cor. 7.10 if a source carries hidden external scale, but it is not a reduction of the conclusion to its inputs. The same-author forward-limit numerator [12] is located inside the already-computed fiber via (7.12) and is not used to select the basis or cuts; external benchmarks (Kira/FIRE; [22,51,52]) provide independent validation. No fitted quantity is renamed as a prediction, and no load-bearing premise rests solely on a self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The reconstruction is a computed solve, not a fit: no target-derived parameters appear in the ledger. The listed 'free parameters' are procedural choices (stop rules, sample sets, reconstruction bounds, witness points) that the paper documents, and they do not tune the central claim. The load-bearing content sits in the domain axioms: generalized-unitarity completeness for this cut hierarchy, sufficiency of the parity-even power-counting ansatz, and especially the scaleless-integral identities underlying S (axiom 4), which carry 296/646 of the kernel-quotient proof. No new physical entities are postulated.

free parameters (5)
  • Double-cut row-selection tie-break
    Four independent double-cut rows selected by a 'fixed tiebreak' among candidate families (Listing 4, §4.3). The rank gain rank(R_K)=4 is criterion-independent, but which four kernel directions get resolved can depend on the tie-break; reported nullity 207 is robust to the choice.
  • Cut-schedule stopping rule (64 dependent consistent rows) = 64
    Rows are added until each physical topology has at least three successful trials and 64 consecutive dependent-and-consistent rows occur (§4.3, Listing 5). The number 64 is a hand-chosen convergence heuristic, not derived from physics.
  • Rational-reconstruction bounds = A=B=10^8, M=1100178405778335937
    Bound guaranteeing a unique rational lift (App. C.2, eq. C.7). Standard technique; all 235,543 coordinates satisfy the bound. Chosen, not fitted to the result.
  • Sample-D_s interpolation set = 36 sample–Ds inputs; Ds=12 held out
    Chosen finite set of internal-polarization dimensions used to interpolate the affine dependence x(Ds)=x^(0)+Ds x^(1) (App. C.1). Extrapolation at Ds=12 is a held-out check.
  • Prescription-comparison witness point = (s12,s23,s34,s45,s51)=(-172,-79,7,-231,320), Ds=3
    Preregistered kinematic point for the nonzero master-integral-coefficient witness C_FR−FL(D) (eq. 8.17). A demonstration point, not a fitted parameter; the non-vanishing is claimed as an exact rational function.
axioms (7)
  • domain assumption Generalized unitarity completeness: products of on-shell tree amplitudes on maximal/box/triple/double cuts give exact linear constraints that determine the integrand within the chosen ansatz.
    Used throughout §3.2–§6 to build the cut rows; standard at one loop, but the completeness of this particular cut hierarchy for pinning the numerator fiber is part of the construction, not proven independently of the cut compilation.
  • domain assumption The parity-even local ansatz (mass dimension 5, at most two inverse-propagator factors, one power of each external polarization) contains a valid BCJ numerator for the one-loop five-gluon amplitude.
    Ansatz envelope fixed in Ω (§3.3, §5.1); parity-odd structures are excluded and the paper notes helicity checks must retain both parity-conjugate roots. Consistency obA(b)=0 supports adequacy for the cut data, but universality across helicities and schemes is assumed.
  • standard math Free-Lie/Jacobi graph generation (the Jac map) produces all graph numerators with exact antisymmetry and Jacobi relations under the fixed routing.
    Proposition 5.2 (proof sketch) plus free-Lie identities; the kinematic Jacobi relations hold symbolically before numerical rows are constructed (§5.2, eq. 5.18).
  • domain assumption Dimensional-regularization scaleless-integral identities under canonical loop-momentum shifts: ∫d^Dℓ (ℓ²)^r = 0, massless tadpoles vanish, ∫d^Dℓ P(ℓ)/(ℓ+Δ)² = 0 when no external scale is present.
    Load-bearing for S(k_a)=0: 296 of 646 kernel conditions vanish by these identities (§7.3–7.4, eqs. 7.21, 7.24). If any 'route-scaleless' source is not genuinely scale-free after its shift, the observable quotient collapses. This is the paper's most exposed premise (App. B.1 routing tables).
  • domain assumption The LSZ projector removes external-line bubble graphs; the post-assembly registry has 31 regular and 25 formal objects with rank P_LSZ = 31.
    Introduced in §7.3 (eq. 7.20); standard amputation, but the specific 56-object registry and the 31/25 split are construction-specific.
  • domain assumption Generic-D kinematic relations: momentum conservation, on-shell conditions, transversality are the only algebraic relations; strictly four-dimensional Gram and Levi-Civita relations are excluded from the reduction.
    §3.2, §5.1: required to retain D-dimensional rational terms and to define the 10,724-dimensional quotient; dropping 4D Gram relations is a deliberate scheme choice.
  • standard math Finite-field sampling with Schwartz–Zippel bounds: ranks, pivots, and kernel computed over several primes lift uniquely to Q(Ds).
    App. C (eqs. C.1–C.7); standard exact-reconstruction methodology with stated uniqueness bounds.

pith-pipeline@v1.3.0-alltime-deepseek · 41301 in / 21455 out tokens · 193533 ms · 2026-08-02T05:15:48.988333+00:00 · methodology

0 comments
read the original abstract

An inverse problem comprises a matrix equation together with its unknown space, physical data, equivalence relation, and validation tests. We formulate Bern--Carrasco--Johansson (BCJ) numerator construction as an exact adaptive inverse problem. A fixed scientific specification determines the theory, graph conventions, coefficient field, locality and power counting, cut data, observable equivalence, and independent checks. Each finite working specification compiles to $\mathcal{P}_\sigma=(A,b;\mathcal{S};\mathcal{T})$, where $Ax=b$ reconstructs numerator coefficients, $\mathcal{S}$ classifies the solution fiber, and $\mathcal{T}$ tests it on held-out information. Left-null obstructions identify candidate numerator-basis directions needed for consistency, while the action of candidate measurements on the right kernel identifies informative new cut equations. We illustrate these steps by hand at four points and apply them to one-loop five-gluon pure Yang--Mills theory. After kinematic and graph-symmetry reduction, the candidate numerator basis contains 1127 independent coordinates. The combined maximal, box, triple, and double cuts have rank 920, giving a 207-dimensional affine solution fiber. Exact reconstruction determines a particular solution and the complete ordered kernel. The specified $R_{12345}$ color-ring readout $\mathcal{S}$ annihilates every kernel direction, so the full fiber represents one observable class. A published forward-limit numerator lies in this fiber, and fresh cuts, independent integral reductions, and helicity-amplitude benchmarks validate the result. Explicit search rules and agent interfaces can propose revisions. Deterministic compilation and exact evaluation assess each proposal.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 47 linked inside Pith

  1. [1]

    Bern, J.J.M

    Z. Bern, J.J.M. Carrasco and H. Johansson,New relations for gauge-theory amplitudes,Phys. Rev. D78(2008) 085011 [0805.3993]

    Pith/arXiv arXiv 2008

  2. [2]

    Bern, J.J.M

    Z. Bern, J.J.M. Carrasco and H. Johansson,Perturbative quantum gravity as a double copy of gauge theory,Phys. Rev. Lett.105(2010) 061602 [1004.0476]

    Pith/arXiv arXiv 2010

  3. [3]

    Huang, J

    R. Huang, J. Rao, B. Feng and Y.-H. He,An algebraic approach to the scattering equations, JHEP12(2015) 056 [1509.04483]

    Pith/arXiv arXiv 2015

  4. [4]

    K. Zhou, J. Rao and B. Feng,Derivation of Feynman rules for higher-order poles using cross-ratio identities in CHY construction,JHEP06(2017) 091 [1705.04783]

    Pith/arXiv arXiv 2017

  5. [5]

    Feng and Y

    B. Feng and Y. Zhang,Note on the labelled tree graphs,JHEP12(2020) 096 [2009.02394]

    Pith/arXiv arXiv 2020

  6. [6]

    B. Feng, C. Hu and Y. Zhang,Note on solutions of scattering equations,Phys. Rev. D105 (2022) 036015 [2105.03700]

    Pith/arXiv arXiv 2022

  7. [7]

    Li and Y

    S. Li and Y. Zhang,Inverse Cachazo–He–Yuan integrand construction from factorization channels with exact integer constraint propagation,Phys. Rev. D113(2026) 105006 [2508.02248]

    arXiv 2026

  8. [8]

    Fu, Y.-J

    C.-H. Fu, Y.-J. Du and B. Feng,Note on symmetric BCJ numerator,JHEP08(2014) 098 [1403.6262]

    Pith/arXiv arXiv 2014

  9. [9]

    Bjerrum-Bohr, J.L

    N.E.J. Bjerrum-Bohr, J.L. Bourjaily, P.H. Damgaard and B. Feng,Manifesting color–kinematics duality in the scattering equation formalism,JHEP09(2016) 094 [1608.00006]

    Pith/arXiv arXiv 2016

  10. [10]

    Huang, Q

    R. Huang, Q. Jin, J. Rao, K. Zhou and B. Feng,The Q-cut representation of one-loop integrands and unitarity cut method,JHEP03(2016) 057 [1512.02860]

    Pith/arXiv arXiv 2016

  11. [11]

    Zhou,On differential operators and unifying relations for one-loop Feynman integrands, JHEP10(2021) 150 [2108.04025]

    K. Zhou,On differential operators and unifying relations for one-loop Feynman integrands, JHEP10(2021) 150 [2108.04025]

    Pith/arXiv arXiv 2021

  12. [12]

    Q. Cao, S. He, Y. Zhang and F. Zhu,Loop-level double copy relations from forward limits,Sci. China Phys. Mech. Astron.69(2026) 261011 [2509.25129]

    arXiv 2026

  13. [13]

    Parke and T.R

    S.J. Parke and T.R. Taylor,An amplitude forn-gluon scattering,Phys. Rev. Lett.56(1986) 2459

    1986

  14. [14]

    Carrasco and H

    J.J.M. Carrasco and H. Johansson,Five-point amplitudes inN= 4Super-Yang–Mills theory andN= 8supergravity,Phys. Rev. D85(2012) 025006 [1106.4711]

    Pith/arXiv arXiv 2012

  15. [15]

    Z. Bern, S. Davies, T. Dennen, Y.-t. Huang and J. Nohle,Color-kinematics duality for pure Yang–Mills and gravity at one and two loops,Phys. Rev. D92(2015) 045041 [1303.6605]

    Pith/arXiv arXiv 2015

  16. [16]

    Boels, R.S

    R.H. Boels, R.S. Isermann, R. Monteiro and D. O’Connell,Colour-kinematics duality for one-loop rational amplitudes,JHEP04(2013) 107 [1301.4165]

    Pith/arXiv arXiv 2013

  17. [17]

    Bern, J.J

    Z. Bern, J.J. Carrasco, M. Chiodaroli, H. Johansson and R. Roiban,The duality between color and kinematics and its applications,Phys. Rept.873(2020) 1 [1909.01358]

    Pith/arXiv arXiv 2020

  18. [18]

    He and O

    S. He and O. Schlotterer,New relations for gauge-theory and gravity amplitudes at loop level, Phys. Rev. Lett.118(2017) 161601 [1612.00417]

    Pith/arXiv arXiv 2017

  19. [19]

    S. He, O. Schlotterer and Y. Zhang,New BCJ representations for one-loop amplitudes in gauge theories and gravity,Nucl. Phys. B930(2018) 328 [1706.00640]. – 67 –

    Pith/arXiv arXiv 2018

  20. [20]

    Mafra and O

    C.R. Mafra and O. Schlotterer,Towards one-loop SYM amplitudes from the pure spinor BRST cohomology,Fortsch. Phys.63(2015) 105 [1410.0668]

    Pith/arXiv arXiv 2015

  21. [21]

    Edison, S

    A. Edison, S. He, H. Johansson, O. Schlotterer, F. Teng and Y. Zhang,Perfecting one-loop BCJ numerators in SYM and supergravity,JHEP02(2023) 164 [2211.00638]

    Pith/arXiv arXiv 2023

  22. [22]

    Bern, L.J

    Z. Bern, L.J. Dixon, D.C. Dunbar and D.A. Kosower,One-loopn-point gauge theory amplitudes, unitarity and collinear limits,Nucl. Phys. B425(1994) 217 [hep-ph/9403226]

    Pith/arXiv arXiv 1994

  23. [23]

    Bern, L.J

    Z. Bern, L.J. Dixon, D.C. Dunbar and D.A. Kosower,Fusing gauge theory tree amplitudes into loop amplitudes,Nucl. Phys. B435(1995) 59 [hep-ph/9409265]

    Pith/arXiv arXiv 1995

  24. [24]

    Britto, F

    R. Britto, F. Cachazo and B. Feng,Generalized unitarity and one-loop amplitudes inN= 4 Super-Yang–Mills,Nucl. Phys. B725(2005) 275 [hep-th/0412103]

    Pith/arXiv arXiv 2005

  25. [25]

    Forde,Direct extraction of one-loop integral coefficients,Phys

    D. Forde,Direct extraction of one-loop integral coefficients,Phys. Rev. D75(2007) 125019 [0704.1835]

    Pith/arXiv arXiv 2007

  26. [26]

    Giele, Z

    W.T. Giele, Z. Kunszt and K. Melnikov,Full one-loop amplitudes from tree amplitudes,JHEP 04(2008) 049 [0801.2237]

    Pith/arXiv arXiv 2008

  27. [27]

    Badger,Direct extraction of one-loop rational terms,JHEP01(2009) 049 [0806.4600]

    S.D. Badger,Direct extraction of one-loop rational terms,JHEP01(2009) 049 [0806.4600]

    Pith/arXiv arXiv 2009

  28. [28]

    Peraro,Scattering amplitudes over finite fields and multivariate functional reconstruction, JHEP12(2016) 030 [1608.01902]

    T. Peraro,Scattering amplitudes over finite fields and multivariate functional reconstruction, JHEP12(2016) 030 [1608.01902]

    Pith/arXiv arXiv 2016

  29. [29]

    Peraro,FiniteFlow: Multivariate functional reconstruction using finite fields and dataflow graphs,JHEP07(2019) 031 [1905.08019]

    T. Peraro,FiniteFlow: Multivariate functional reconstruction using finite fields and dataflow graphs,JHEP07(2019) 031 [1905.08019]

    Pith/arXiv arXiv 2019

  30. [30]

    Abreu, F

    S. Abreu, F. Febres Cordero, H. Ita, M. Jaquier, B. Page and M. Zeng,Two-loop four-gluon amplitudes from numerical unitarity,Phys. Rev. Lett.119(2017) 142001 [1703.05273]

    Pith/arXiv arXiv 2017

  31. [31]

    Zhang,PJ-RoPE: A Fourier–jet–affine position space for relative attention,2606.05345

    Y. Zhang,PJ-RoPE: A Fourier–jet–affine position space for relative attention,2606.05345

    Pith/arXiv arXiv

  32. [32]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan et al.,ReAct: Synergizing reasoning and acting in language models, inInternational Conference on Learning Representations, 2023 [2210.03629]

    Pith/arXiv arXiv 2023

  33. [33]

    Schick, J

    T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer et al.,Toolformer: Language models can teach themselves to use tools, inAdvances in Neural Information Processing Systems, vol. 36, 2023 [2302.04761]

    Pith/arXiv arXiv 2023

  34. [34]

    Yang, C.E

    J. Yang, C.E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan et al.,SWE-agent: Agent–computer interfaces enable automated software engineering,2405.15793

    Pith/arXiv arXiv

  35. [35]

    Wang et al.,OpenHands: An open platform for AI software developers as generalist agents, inInternational Conference on Learning Representations, 2025 [2407.16741]

    X. Wang et al.,OpenHands: An open platform for AI software developers as generalist agents, inInternational Conference on Learning Representations, 2025 [2407.16741]

    Pith/arXiv arXiv 2025

  36. [36]

    C. Lu, C. Lu, R.T. Lange, J. Foerster, J. Clune and D. Ha,The AI scientist: Towards fully automated open-ended scientific discovery,2408.06292

    Pith/arXiv arXiv

  37. [37]

    Gendreau-Distler, J

    E. Gendreau-Distler, J. Ho, D. Kim, L.T. Le Pottier, H. Wang and C. Yang,Automating high energy physics data analysis with LLM-powered agents,2512.07785

    arXiv

  38. [38]

    Menzo, A

    T. Menzo, A. Roman, S. Gleyzer, K. Matchev, G.T. Fleming, S. Höche et al.,HEPTAPOD: Orchestrating high energy physics workflows towards autonomous agency,2512.15867

    Pith/arXiv arXiv

  39. [39]

    S. Qiu, Z. Cai, J. Wei, Z. Li, Y. Yin, Q.-H. Cao et al.,An end-to-end architecture for collider physics and beyond,2603.14553. – 68 –

    arXiv

  40. [40]

    J. Jiao, T. Liu, K. Li, W. Song, Y. Liao, B. Zhang et al.,HepScript: A dual-use DSL for human–AI collaborative data analysis workflows in high-energy physics,2605.01423

    Pith/arXiv arXiv

  41. [41]

    Badea, Y

    A. Badea, Y. Chen, M. Maggi, Y.-J. Lee and Electron-Positron Alliance,Agentic AI–physicist collaboration in experimental particle physics: A proof-of-concept measurement with LEP open data,2603.05735

    arXiv

  42. [42]

    Moreno, S

    E.A. Moreno, S. Bright-Thonney, A. Novak, D. Garcia and P. Harris,AI agents can already autonomously perform experimental high energy physics,2603.20179

    Pith/arXiv arXiv

  43. [43]

    He et al.,Dr.Sai: An agentic AI for real-world physics analysis at BESIII,2604.22541

    M. He et al.,Dr.Sai: An agentic AI for real-world physics analysis at BESIII,2604.22541

    Pith/arXiv arXiv

  44. [44]

    Faroughy, S

    D.A. Faroughy, S. Palacios Schweitzer, I. Pang, S. Mishra-Sharma and D. Shih,Collider-Bench: Benchmarking AI agents with particle physics analysis reproduction,2605.13950

    Pith/arXiv arXiv

  45. [45]

    Luo, T.-Z

    M.-x. Luo, T.-Z. Yang, H.X. Zhu and Y.J. Zhu,Quark transverse parton distribution at the next-to-next-to-next-to-leading order,Phys. Rev. Lett.124(2020) 092001 [1912.05778]

    Pith/arXiv arXiv 2020

  46. [46]

    von Hippel and M

    M. von Hippel and M. Wilhelm,Refining integration-by-parts reduction of Feynman integrals with machine learning,JHEP05(2025) 185 [2502.05121]

    Pith/arXiv arXiv 2025

  47. [47]

    Song, T.-Z

    Z.-Y. Song, T.-Z. Yang, Q.-H. Cao, M.-x. Luo and H.X. Zhu,Explainable AI-assisted optimization for Feynman integral reduction,JHEP06(2026) 225 [2502.09544]

    Pith/arXiv arXiv 2026

  48. [48]

    Tooby-Smith,HepLean: Digitalising high energy physics,Comput

    J. Tooby-Smith,HepLean: Digitalising high energy physics,Comput. Phys. Commun.308 (2025) 109457 [2405.08863]

    Pith/arXiv arXiv 2025

  49. [49]

    Brenner, V

    M.P. Brenner, V. Cohen-Addad and D. Woodruff,Solving an open problem in theoretical physics using AI-assisted discovery,2603.04735

    arXiv

  50. [50]

    Bern, L.J

    Z. Bern, L.J. Dixon and D.A. Kosower,Dimensionally regulated one-loop integrals,Phys. Lett. B302(1993) 299 [hep-ph/9212308]

    Pith/arXiv arXiv 1993

  51. [51]

    Bern, L.J

    Z. Bern, L.J. Dixon and D.A. Kosower,One-loop corrections to five-gluon amplitudes,Phys. Rev. Lett.70(1993) 2677 [hep-ph/9302280]

    Pith/arXiv arXiv 1993

  52. [52]

    Bern, L.J

    Z. Bern, L.J. Dixon and D.A. Kosower,On-shell recurrence relations for one-loop QCD amplitudes,Phys. Rev. D71(2005) 105013 [hep-th/0501240]

    Pith/arXiv arXiv 2005

  53. [53]

    Bern, L.J

    Z. Bern, L.J. Dixon, D.C. Dunbar and D.A. Kosower,One-loop self-dual andN= 4 Super-Yang–Mills,Phys. Lett. B394(1997) 105 [hep-th/9611127]

    Pith/arXiv arXiv 1997

  54. [54]

    Bern and G

    Z. Bern and G. Chalmers,Factorization in one-loop gauge theory,Nucl. Phys. B447(1995) 465 [hep-ph/9503236]. – 69 –

    Pith/arXiv arXiv 1995