Searching for New Physics with Reinforcement Learning
Jacky Kumar, Marianne Bouchard, David London
hep-ph · cs.LG
Quick verdict
RL actor-critic autonomously searches 912 SMEFT operators to explain anomalies; reproduces W-mass results and beats random search, but no code release.
Summary
The authors formulate the search for SMEFT operators that explain low-energy anomalies as a reinforcement-learning problem: a transformer-based policy sequentially proposes operators from a set of 912 baryon-number, lepton-number and lepton-flavour conserving dimension-6 operators, each scored by fitting its Wilson coefficient with Minuit over observables computed with flavio, with reward given by Δχ² minus a penalty for ineffective proposals. For the CDF W-mass anomaly (27 Z-pole observables plus the m_W world average; χ²_SM = 53.8), the policy reproduces the known [O_ll]1221 ⊕ [O_HD] solution (Δχ² = 27.6, χ²/d.o.f. = 1.05) and finds an additional [O_lu]1133 ⊕ [O_Hu]33 combination (Δχ² = 20.6) absent from the earlier analysis. On a synthetic 32-observable dataset with ten fictitious 3-5σ anomalies (χ²_SM = 178.3), actor-critic outperforms REINFORCE-only (final average reward ≈153 vs ≈141) and uniform random search: 75K evaluations over the full 912-operator space yield 560 unique sets with Δχ² ≥ 120 (random: 0), the best being a 10-operator solution with Δχ² = 156.4. Limitations: no code release, unspecified hyperparameters (penalty c_P, architecture), no ground-truth recovery test for the synthetic anomalies, and unequal evaluation budgets (75K vs 25K) across search-space comparisons.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An actor-critic RL framework with a transformer policy can autonomously search 912 SMEFT dimension-6 operators using only a Δχ²-based reward, without human physics intuition.
- For the CDF W-mass anomaly the method reproduces the published [O_ll]1221 ⊕ [O_HD] solution (Δχ² = 27.6) and identifies an additional [O_lu]1133 ⊕ [O_Hu]33 combination (Δχ² = 20.6) missed by the earlier analysis.
- On a synthetic 32-observable dataset with ten fictitious 3-5σ anomalies, RL finds 560 unique operator sets with Δχ² ≥ 120 in 75K evaluations over the full 912-operator space, where uniform random search finds none.
- Adding a critic improves training, raising the final average reward from ≈141 (REINFORCE-only) to ≈153.
Strengths
Principled RL formulation
The policy/action/state/reward mapping onto the operator-search problem is explicit; the reward (Eq. 1) is the physically meaningful Δχ² with a penalty c_P for dead proposals (Δχ² improvement ≤ 0.001), and the search is driven solely by the χ² landscape with no injected physics intuition.
Externally validated case study
For the CDF W-mass anomaly (χ²_SM = 53.8 over 27 Z-pole observables plus the m_W world average), the policy recovers the [O_ll]1221 ⊕ [O_HD] solution of Bagnaschi et al. with Δχ² = 27.6, χ²/d.o.f. = 1.05 and WCs of order 10⁻⁸ GeV⁻² at Λ = 1 TeV, plus a second [O_lu]1133 ⊕ [O_Hu]33 combination (Δχ² = 20.6, m_W = 80404 MeV).
Quantified advantage over random search
Algorithm-level ablation
Weaknesses
No ground-truth recovery in the synthetic study
Confounded Table II comparisons
Reproducibility gaps
Thin baseline coverage
Editorial carelessness
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv:2609.10382, UdeM-GPP-TH-26-311) introduces a reinforcement-learning (RL) procedure that autonomously searches the dimension-6 SMEFT operator space for combinations explaining low-energy "anomalies". The search space is 912 operators conserving baryon number, lepton number, and individual lepton flavour (out of 2499 flavour-indexed Warsaw-basis operators). Two central claims: (i) on the CDF W-mass anomaly the method reproduces the known [O_ll]1221 ⊕ [O_HD] solution of Bagnaschi et al. (Δχ² = 27.6) and additionally finds [O_lu]1133 ⊕ [O_Hu]33 (Δχ² = 20.6) missed there; (ii) on a synthetic 32-observable dataset with ten invented 3-5σ anomalies, RL identifies hundreds of high-Δχ² operator sets where uniform random search finds none.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the paper 'Searching for New Physics with Reinforcement Learning' about?
It trains a reinforcement-learning agent (actor-critic with a transformer policy) to automatically pick combinations of SMEFT dimension-6 operators — from 912 candidates — that best explain measured deviations from Standard Model predictions, using the improvement in a global χ² fit as the learning reward and tools like flavio and Minuit for the physics evaluation.
What results does this RL new-physics search achieve?
On the CDF W-boson mass anomaly it reproduces a published two-operator solution (Δχ² = 27.6) and finds an extra viable combination (Δχ² = 20.6). On a synthetic set of 32 observables with ten invented 3-5σ anomalies, it found 560 operator sets with Δχ² ≥ 120 in 75,000 evaluations over the full 912-operator space, while uniform random search found none; the best solution used 10 operators with Δχ² = 156.4.
Is code available for this paper?
No public code release is mentioned. The paper relies on public tools (flavio for observable predictions, Minuit for fitting), but its own policy network implementation, training hyperparameters, and the exact synthetic dataset values are not published, so independent reproduction currently requires reimplementation from scratch.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1's plotted curves are present only as caption text in the extraction; visual details of the learning curves (e.g., exact plateau shapes) could not be assessed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.