How much of a five-loop form factor must a model learn?
Velie and colleagues report reconstructing a known five-loop form-factor symbol from as little as 5% labeled input. A transformer proposes coefficients. Exact relations fill in others. A small calculation using the original reference data shows how one coefficient can determine another.
What the labels represent
The observable is a three-gluon form factor of the chiral stress-tensor multiplet in planar four-dimensional N = 4 super Yang–Mills theory. An operator insertion carries momentum into three on-shell states. At five loops the symbol encodes the form factor’s iterated logarithmic structure in words of length ten over six letters.
Each coefficient labels an ordered sequence of letters. Adjacency rules forbid some sequences, and integrability and symmetry relate the coefficients of others. A symbol retains information about iterated discontinuities. Recovering it still leaves information beyond the symbol to determine the full function.
Where the search enters
Velie, Jha, Ganesh and Ghosh grouped the last five letters into a compact basis. Given a subset of the known coefficients, their algorithm propagates consequences of exact relations until it reaches assignments those relations alone cannot fill. A transformer proposes values for an unassigned key. Monte Carlo tree search tries assignments and uses the relations to reject inconsistent branches.
In §5 and Figure 2 the authors report assigning all 12,543 compact keys at each tested seed fraction from 5% to 50%. They report agreement with the earlier bootstrap result up to a cyclic transformation. Their transformer-only comparison uses a six-loop benchmark. It does not isolate the five-loop gain from adding search under otherwise identical conditions.
An exact prediction from one symmetry
The reference data make a useful distinction visible. In one encoding, the final four letters select a block and the first six select a coefficient inside it. In another, the final eight select a block and only a two-letter prefix remains. Both can name the same ten-letter word. Dixon and colleagues supply both encodings for five loops.
Write C(w) for the coefficient of word w in the reference file’s alphabet. A reflection of the form factor exchanges b with c and e with f while leaving a and d fixed. Apply it to the word beginning with b followed by nine d’s. Once that coefficient is known, symmetry predicts the value for the word beginning with c. Here the prediction can be checked against a second tabulated entry.
Reading the same coefficient two ways
We selected the eight-letter endings bbcbdddd and dddddddd. Both also end in the listed four-letter block dddd. For each ending we compared all 36 possible two-letter prefixes across the two encodings. All 72 coefficients agreed exactly: 15 were nonzero and 57 were zero. Changing the sign of the octuple-block value for aabbcbdddd, from −1/4 to +1/4, produced a single mismatch.
For the reflection example, the value predicted from the b word was −26,880. The c word had that coefficient in the table. The data use rational coefficients in their published alphabet. A comparison to the learned integer keys would also require the letter normalization, the compact basis and the cyclic relabeling. These 72 comparisons use only the bootstrap reference. The calculation package includes the rational inputs and an option to extract them again from the pinned ancillary file.
Limits of this result
- The reported recovery from 5% input is a benchmark on a known five-loop symbol with exact relations and supplied labels.
- The methods paper v2 reports cyclic agreement but provides no linked learned coefficient file or specific alignment map. The site has not compared all 12,543 learned keys with the reference.
Checks performed
Exact comparison of two bootstrap reference tables
Independently extracted two selected eight-letter suffix blocks and matched all 72 possible ten-letter coefficients against a four-letter suffix table. Checked one reflection prediction. No transformer training, search or comparison to learned output was run.
Sources & citations
- Learning Scattering Amplitudes with Transformer Reinforcement LearningPhilip Velie, Piyush Jha, Vijay Ganesh and Aishik Ghosh · 2026-10-31
arXiv:2609.32096v2. §2 and Eq. (1), observable and symbol alphabet. §§3–4, compact suffix basis, propagation, transformer policy and tree search. §5 and Fig. 2, five-loop reconstruction and six-loop baseline.
- Bootstrapping a Stress-Tensor Form Factor through Eight LoopsLance J. Dixon, Ömer Gürdoğan, Andrew J. McLeod and Matthias Wilhelm · 2023-03-19
arXiv:2204.11901v2. §2, operator and normalization. §4, closing discussion of ancillary files Esymb.txt and FFmultifinalentry.txt.
- Esymb.txt — versioned symbol coefficientsLance J. Dixon, Ömer Gürdoğan, Andrew J. McLeod and Matthias Wilhelm · 2023-03-19
Header on alphabet and dihedral symmetry. quadic8 and Esymbquad[5], first block. octic93 and Esymboct[5], blocks 47 and 93 (one-based).
Index terms form factor · five-loop · three-gluon · planar N=4 SYM · super Yang–Mills · stress tensor · symbol bootstrap · symbol coefficients · transformer reinforcement learning · Monte Carlo tree search · MCTS · suffix compression · arXiv:2609.32096