Source-linked AI summary
The FERMIACC: Agents for Particle Theory
Prateek Agrawal, Nathaniel Craig, Amalia Madden, Iñigo Valenzuela Lombera
TL;DR
High energy physics data often admit many compatible theories, so the challenge is to generate and discriminate among plausible explanations rather than identify a unique theory. The FERMIACC uses scaffolded OpenAI agents, adversarial refinement, and deterministic simulation tools to generate and quantitatively test hypotheses. It demonstrates this approach on statistical fluctuations in LHC data, while identifying physics validation and higher-fidelity recasting as important limitations and future needs.
Problem
High energy physics has many distinct theories compatible with data at fixed precision and energy, creating a need to systematically navigate and discriminate among viable hypotheses.
Method
The FERMIACC couples OpenAI-based particle-theory agents, adversarial proposer–critic refinement, structured outputs, and deterministic simulation tools to generate and quantitatively test hypotheses.
Results
The FERMIACC produces and tests novel explanations for mild statistical fluctuations in ATLAS and CMS data, including the historical 750 GeV excess and analyses in varied final states.
Takeaways & Limitations
The FERMIACC provides a template for verifiable hypothesis generation that can explore competing explanations reproducibly and in parallel.
Takeaways & Limitations
The current workflow lacks a physics validation layer to reproduce published signal efficiencies and support higher-fidelity recasts.
Abstract
from arXiv · showhide
We present the FERMIACC, a scaffolded reasoning model built on OpenAI agents designed to autonomously generate and quantitatively validate theory hypotheses for high energy physics data at scale.
1 Introduction
High energy physics needs systems that navigate many theories compatible with existing data, rather than simply discovering one unique theory. The FERMIACC addresses this by combining scaffolded agents with deterministic simulation and adversarial validation to generate and test hypotheses at scale.
- Motivation: Effective field theory permits infinitely many theories consistent with fixed-precision data, making theory navigation more important than finding a single compatible theory.The central challenge is an overabundance of viable theories across the gap between the weak and Planck scales.
- Motivation: The proposed “AI Fermi” would characterize, navigate, test, and discriminate among viable theories while autonomously generating and quantitatively validating hypotheses for high energy physics data.The goal is to explore innumerable hypotheses across innumerable experiments and maximize the impact of existing data.
- Motivation: Current LLMs face hallucinations, probabilistic behavior, and weak long-horizon planning, motivating programmatic scaffolding with intermediate-state management, deterministic tools, and verification.Benchmark success alone does not establish a reliable scientific discovery agent.
- FERMIACC: The FERMIACC couples commercial agents to deterministic particle-physics simulation, using adversarial proposer–critic loops and iterative simulation feedback to generate and validate quantum field theory hypotheses.It reads experimental papers and produces structured hypotheses intended for quantitative testing.
- FERMIACC: The FERMIACC’s adversarial agent configuration is described as essential for generating novel particle-physics hypotheses and may extend to collaboration settings with full simulation tools and correlations.The authors present this configuration as a broader template for reasoning-model applications in high energy theory.
- Context: The work extends prior agentic particle-physics systems into hypothesis generation and theory interpretation while retaining the general-learning advantages emphasized by the bitter lesson.The paper positions the FERMIACC within developments spanning Monte Carlo workflows, auditable pipelines, autonomous simulations, and experiment-design loops.
2 Architecture
The FERMIACC uses scaffolded agent modules to turn experimental analyses into executable BSM hypotheses, simulate their collider signatures, and test their significance. Its architecture combines adversarial model refinement, deterministic software tools, structured outputs, and validation guardrails.
- Pipeline overview: The pipeline reads an experimental paper, proposes field-theoretic explanations, encodes them in FeynRules, simulates events, analyzes signal regions, and stores results for later hypotheses.The workflow proceeds from hypothesis generation through event generation, showering and detector simulation, selection cuts, statistical analysis, and database storage.
- Pipeline overview: The architecture is organized into three agent modules, including a Model Builder that proposes and formalizes Standard Model extensions.Figure 1 distinguishes agent modules from integrated software components.
- Model Builder: The Model Builder uses a proposer–critic–patching loop and transfers only converged proposals to FeynRules for validity checks.Intermediate proposals, critiques, and patches are logged, while accepted summaries enter a searchable database.
- Analysis generation and validation: The event-analysis stage uses MadGraph, Pythia, Delphes, and MadAnalysis under conservative guardrails, while modular state and recovery steps support complex threshold or category-dependent selections.A future physics validation layer is intended to compare reproduced and published signal efficiencies.
- Model Builder: Typed JSON schemas constrain BSM proposals for deterministic parsing, while ensembles at different sampling temperatures explore alternative hypotheses.The proposal database retrieves structurally similar prior models using field-content descriptors to support novelty checks.
- Model Builder: The executable benchmark is deliberately separated from the full model story because the current UFO/FeynRules workflow supports only restricted gauge-basis constructions.Each hypothesis specifies which fields and couplings enter the executable UFO and which ultraviolet features remain outside it.
3 Examples
The FERMIACC generated and tested diverse hypotheses for mild fluctuations in CMS and ATLAS data, including the 750 GeV diphoton excess and later dijet-related anomalies. Single-shot examples produced novel, quantitatively simulated explanations, although several fell short of the required signal rate.
- 750 GeV diphoton excess: The FERMIACC used the 750 GeV diphoton excess as a benchmark and produced viable proposals spanning most models proposed for the anomaly.It tested representative CMS and ATLAS proposals against the respective diphoton analyses.
- 750 GeV diphoton excess: It checked a pseudoscalar coupled to heavy vector-like quarks by evaluating EFT coefficients, production, diphoton branching, width, and simulated event histograms.The proposal was implemented in a UFO file and passed through the deterministic simulation pipeline for both CMS analysis regions.
- CMS paired dijet resonances: A scalar with EFT couplings reproduced the targeted diphoton features only with a small gluon coupling, which preserved the diphoton branching ratio but yielded a small signal cross section.The proposal was propagated through the deterministic simulation pipeline to produce the Figure 6 histogram.
- CMS paired dijet resonances: The FERMIACC generated novel candidate explanations for both CMS four-jet excesses, distinct from the interpretations in the experimental paper and later reinterpretations.Both candidates used conservative gluon couplings to preserve decays into daughter particles, limiting their signal cross sections.
- Overall examples: Together, the examples demonstrate first-pass hypothesis generation and validation, while several single-shot models failed to reproduce the necessary rate.The authors identify iterative runs as a potential way to improve these outcomes.
4 Conclusion
The FERMIACC combines scaffolded OpenAI agents with deterministic simulation tools to generate structured, testable explanations for LHC data. The framework enables rapid comparison of competing hypotheses, while its current examples remain single-shot and motivate further validation and iteration.
- Contribution: The FERMIACC is a scaffolded reasoning model that autonomously generates and quantitatively validates plausible hypotheses for high energy physics data.Its initial demonstrations target mild statistical fluctuations in ATLAS and CMS data across multiple final states.
- Architecture: Structured outputs can be passed directly to simulation tools, making proposed explanations reproducible and checkable step by step.This moves the system beyond unconstrained chat-based assistance.
- Capability: The framework supports exploring many competing explanations quickly and in parallel.The claim follows from converting proposals into executable, quantitatively tested outputs.
- Limitations and future work: The presented examples are single-shot runs, and further development is needed for self-validation, improved fast-detector emulation, related-analysis limits, and non-LHC simulations.The authors identify deployment within experimental collaborations as a longer-term direction for handling recasting and correlations.
- Implication: The FERMIACC provides a template for verifiable hypothesis generation in high energy physics.The paper suggests that similar adversarial agent configurations may support broader scientific reasoning systems.
A Implementation Details of the Model Builder
The hypothesis-generation stage uses structured representations, canonicalized retrieval, and an iterative propose–critique–patch loop to produce auditable BSM proposals constrained for downstream execution.
- Each hypothesis is represented as a structured object containing analysis anchors, new fields, kinetic terms, interactions, and other model information.
- Parameter justifications must contain one entry for every new field and coupling, linking benchmark values to explicit phenomenological reasoning.
- For each analysis PDF, the stage generates candidate proposals at different sampling temperatures and processes them independently through refinement.
- The critic evaluates proposals for novelty, physical consistency, UFO compatibility, and parameter-estimation quality, assigning PASS, REFINE, or FAIL labels.
- The procedure yields auditable proposal traces rather than a single opaque model suggestion.
- Canonicalized field signatures retrieve structurally similar prior proposals before the agent compares further theory and process details for genuine distinctness.
A.4 Executable Benchmark Contract
The executable benchmark contract separates the full theory proposal from the restricted model passed downstream and records the artifacts and decisions needed for reproducible reuse.
- The representation separates the full theory story from the executable benchmark because the downstream path supports only a restricted class of extensions.
- Executable benchmarks must avoid new gauge sectors, additional vacuum expectation values, explicit broken-phase authoring, and custom mixing or diagonalization metadata.
- The UFO contract records required fields and couplings, optional couplings that may be omitted, and acknowledged ultraviolet features left unsupported.
- Each proposal run stores the initial hypothesis, patched hypotheses, and final decision and summary as a structured event trace.
- Final proposal summaries are stored with field signatures, interaction terms, topology, process definitions, and critique outcomes for future novelty references.
B Implementation Details of Model Construction and Event Generation
After approval, the pipeline converts a proposal into explicit handoff artifacts, a restricted run card, and reproducible MadGraph inputs while preserving links to the analysis document.
- The stage begins with an approved proposal summary and structured benchmark definition, then writes proposal-specific artifacts for downstream execution.
- Artifacts include run-card settings, parameter mappings, UFO scope, build status, repair snapshots, process configuration, and explicit parameter points.
- These files make stage handoffs explicit and allow the workflow to be rerun or audited without recomputing the entire chain.
- The run-card helper infers beam settings, object-level cuts, and basic separation cuts from the analysis PDF and approved hypothesis.
- The helper uses a short whitelist of robust settings rather than tuning PDFs or making broad generator-level choices.
B.3 Executable Benchmark Construction
Executable benchmark construction translates approved hypotheses into restricted FeynRules and UFO models, validates their required couplings and processes, and then configures reproducible MadGraph generation.
- The pipeline assigns executable identifiers to particles and couplings while deriving a contract that specifies full or reduced benchmark scope and required content.
- The benchmark must remain in the unbroken Standard Model gauge basis and excludes unsupported broken-phase fields, gauge sectors, mixing machinery, and related constructs.
- The guardrails are conservative: they catch recurrent failures before symbolic processing but do not prove physical completeness.
- The symbolic workflow writes and loads the extension, sanitizes the exported UFO, verifies required couplings, and runs a minimal MadGraph generation test.
- If construction fails, diagnostics drive a repair loop that changes only executable model text until success or the repair limit is reached.
- A successful export undergoes sanitation, coupling-survival, and valid-diagram checks because syntactic validity alone may not realize the requested process.
- MadGraph commands combine the proposal process, analysis-derived run-card settings, and validated benchmark parameters so event generation uses the approved masses and couplings.
- The workflow generates parton-level events with MadGraph and passes them to Pythia8 by default, while Delphes remains available but off by default.
B.9 Benchmark Points and Parameter Scans
The pipeline supports a default benchmark point and optional parameter scans, with explicit safeguards against unbounded scan growth. Proposal-specific manifests preserve build, execution, warning, and success state for rerun-safe workflows.
- Benchmark points: The default stage-3 workflow runs only the benchmark point recorded in each proposal, while scan ranges remain optional and advisory.Requested scans are expanded only when explicitly enabled.
- Parameter scans: Scan requests are converted into Cartesian grids, capped by campaign-level point limits, written to scan plans, and executed with one MadGraph run per retained point.The cap prevents parameter scans from growing without bound.
- Execution state: Proposal manifests record UFO build attempts, the latest successful package, MadGraph runs, the latest successful run, and non-fatal warnings.These records make executable state explicit for each proposal.
- Pipeline organization: The recasting pipeline uses a shared AnalysisTemplate containing analysis metadata, reconstructed objects, regions, statistical inputs, and generated backend code.Major stages operate on this intermediate representation to reduce ambiguity and localize failures.
- Pipeline organization: Specialized agents extract analysis structure, identify numerical sources, and generate MadAnalysis 5 backend code alongside deterministic mapping, validation, archival, execution, and statistics stages.The components divide model-driven interpretation from deterministic processing.
C.2 Data-Source Layer
The data-source layer discovers candidate numerical inputs from structured records, official pages, digitized plots, and user mappings, then passes them through deterministic resolution into canonical recasting fields. Generated backends are compiled, executed, statistically analyzed, and archived for inspection and debugging.
- Source discovery: The DataSourceAgent searches HEPData, official analysis pages, digitized spectra, and user-provided CSV or JSON mappings.Its role is source discovery rather than final numerical arbitration.
- Source discovery: Web search locates public records and figures, dedicated readers retrieve structured tables, and digitization extracts approximate series when information is graphical.The output is a structured source plan, not final signal-region data.
- Numerical resolution: Deterministic mapping aligns candidate inputs with signal-region structure, while resolution selects observed values, background expectations, and uncertainties for canonical recasting fields.This separates discovery from numerical population of the analysis template.
- Backend execution: The generated backend includes analyzer code, region and histogram declarations, metadata, and archived artifacts, then undergoes topology normalization and consistency validation.The current backend target is MadAnalysis 5.
- Backend execution: Compiled packages run on Delphes-level samples, native MadAnalysis 5 statistics compute efficiencies and CLs-related outputs, and the full run is persisted with logs and artifacts.Archival makes each stage inspectable and supports targeted debugging after downstream failures.
C.5 Architectural Perspective
The architecture combines an agentic front end with deterministic infrastructure so planning, source discovery, code generation, validation, and reruns remain inspectable. The example shows the model proposing and justifying a UV-completion variant before reducing it to collider-relevant effective couplings.
- Architectural perspective: Specialized stages produce explicit artifacts, allowing failures to be localized to planning, acquisition, statistical resolution, code generation, or execution.This avoids hiding errors inside a single opaque generation step.
- Architectural perspective: The long-term design assigns planning and source discovery to agents while reserving topology construction, backend emission, validation, and staged reruns for a deterministic backbone.The architecture follows experience that region topology, cut naming, and backend metadata require deterministic ownership.
- Model-driven proposal: The model generates fields, interactions, couplings, mass estimates, and scan ranges, then passes these initial guesses to a deterministic pipeline for refinement.The example uses a CMS diphoton excess proposal as an illustration.
- Model-driven proposal: The proposal specifies a 750 GeV CP-odd singlet A and a 1.2e3 GeV vectorlike fermion T, with couplings and scan ranges encoded in the structured output.The fields and couplings are represented with explicit quantum numbers, benchmark values, and suggested scans.
- Model-driven proposal: The model justifies masses and couplings using the diphoton spectrum, decay-closure conditions, heavy-fermion matching, production estimates, and narrow-width considerations.It estimates yT = 10, cAG ∼3e-5 GeV^-1, and cAB ∼9e-6 GeV^-1 for the benchmark.
- Model-driven proposal: The coupling estimates imply a narrow resonance with gluon-fusion production and electroweak decay modes constrained by the same effective operators.The supplied reasoning estimates Gamma(A -> gg) ∼1e-2 GeV and Gamma(A -> aa) ∼1e-4 GeV.
D.3 Model interpretation and passing on to collider tools
The model recognizes that the collider-facing EFT benchmark is not new and frames the proposal's novelty as a linked UV completion. The deterministic pipeline therefore receives a reduced collider benchmark while retaining the distinction between effective phenomenology and UV structure.
- Model interpretation: The proposal's novelty lies in providing a UV completion for an existing EFT model rather than introducing a new collider-facing benchmark.The existing EFT describes a CP-odd singlet produced by gluon fusion and decaying to two photons.
- Model interpretation: The heavy vectorlike quark T generates effective A G Gtilde and A B Btilde operators, so the entry is stored as a UV-completion variant linked to the existing EFT benchmark.This classification avoids treating the collider phenomenology as a distinct standalone benchmark.
- Passing to collider tools: The reduced UFO contract retains A, cAG, and cAB while excluding T, yT, yH, and explicit one-loop matching from the collider benchmark build.The omitted UV features are not required for the target signature in the reduced pipeline.
- Passing to collider tools: The proposal is accepted as a physically and pipeline-consistent reduced CP-odd singlet EFT benchmark, with its UV completion correctly linked rather than registered as a new standalone benchmark.The recorded decision is PASS at temperature 0.7.