LN — Single-minus Gluon Tree Amplitudes (Arkani-Hamed et al., 2026)
Full Reference: Arkani-Hamed, N., et al. (2026). Single-minus gluon tree amplitudes are nonzero. arXiv preprint arXiv:2602.12176. https://doi.org/10.48550/arXiv.2602.12176
Pass 1 — Bird’s Eye
Main Claim
Single-minus gluon tree amplitudes (traditionally assumed zero) are nonzero in specific regimes. GPT-5.2 Pro played a central role in deriving the closed-form expression, correcting an established belief in theoretical physics.
| Property | Detail |
|---|---|
| Type | Research paper — Theoretical Physics |
| Relevance to PUMA | ⭐⭐ Medium — key example for PUMA’s state-of-the-art section on AI-generated knowledge; demonstrates AI can correct human beliefs |
Pass 2 — Key Content
The Discovery
- Prior belief: Single-minus gluon helicity amplitudes = 0 at tree level (standard QFT result)
- New result: In “half-collinear” regime (Klein space or complex momenta), these amplitudes are non-zero
- Formula: Closed-form piecewise expression for n-point amplitudes derived and verified
AI’s Role (GPT-5.2 Pro)
According to OpenAI’s communications (openai.com/index/new-result-theoretical-physics/):
- Simplified extremely complex symbolic expressions
- Identified algebraic patterns invisible to human experts
- Proposed the general form of the closed-form formula
- Result verified against Berends-Giele recursion and standard soft theorems by a formal AI verifier
Significance
- One of the first documented cases of an LLM correcting an established belief in theoretical physics
- Goes beyond “pattern matching on existing data” — the result was counter to prevailing expert knowledge
- Challenges Felin & Holweg’s (2024) argument that AI cannot generate novel scientific theory
Pass 3 — PUMA Implications
This result is important for PUMA’s framing:
- Demonstrates LLMs can operate at the frontier of formal knowledge domains
- The verification chain (formal AI verifier + expert mathematicians) mirrors PUMA’s validation approach (statistical tests + human PM review)
- Reinforces that LLMs with sufficient reasoning depth can exceed interpolation
Note for PUMA Section 2: Cite this alongside Jumper 2021 (AlphaFold) and Degrave 2022 (plasma control) to build the case that AI-driven PM automation (PUMA) belongs to a broader validated trajectory.
APA7 Citation
Arkani-Hamed, N., et al. (2026). Single-minus gluon tree amplitudes are nonzero. arXiv preprint arXiv:2602.12176. https://doi.org/10.48550/arXiv.2602.12176