⚛️ phenomenology

Language-Guided Hypotheses Generation for Sparse SMEFT Analyses

The paper introduces **llm4smeft**, an open-source, locally runnable framework that combines a literature fine-tuned language model with retrieval-augmented generation to automate the selection of relevant SMEFT operators and their Fisher information for sparse global fits, thereby reducing reliance on manual theoretical insight.

Original authors: Ahmed Hammad, Veronica Sanz

Published 2026-08-06

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmed Hammad, Veronica Sanz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, intricate puzzle where every piece represents a fundamental particle or force. Scientists have built a nearly perfect picture of this puzzle called the Standard Model, but they suspect there are hidden pieces—new, heavier particles—that are too massive to see directly. To find clues about these invisible giants, physicists use a clever trick called the Standard Model Effective Field Theory (SMEFT). Think of SMEFT as a "shadow language." Instead of trying to see the heavy new particles, scientists look for the tiny shadows they cast on the known particles. These shadows are described by mathematical "operators," which are like specific instructions on how the known particles should wiggle or interact if a new, heavy friend is lurking nearby.

The problem is that there are thousands of possible instructions (operators), and the data from giant particle colliders is like a massive, noisy library. Trying to figure out which few instructions are actually causing the weird signals is like searching for a specific needle in a haystack that keeps growing. Physicists usually have to use their deep expertise to guess which needles to look for, but the haystack is getting so big that even experts can get overwhelmed. This is where the question arises: Can a computer program learn to be a better detective than a human, spotting the right needles without getting lost in the hay?

This paper introduces a new tool called llm4smeft, which acts like a super-smart, physics-savvy assistant to help solve this puzzle. The authors didn't just teach a computer to read science papers; they built a system that combines a specialized language model (a type of AI trained on the specific "dialect" of particle physics) with a massive database of pre-calculated results. Imagine the AI as a brilliant detective who has read every mystery novel ever written, but who also has a direct hotline to a library of solved cases. When a physicist asks, "I see a weird signal in the Higgs boson and the Z boson; what could be causing it?", the AI doesn't just guess based on its memory. Instead, it first checks the library of solved cases to see which mathematical instructions (operators) are known to explain those specific signals.

The system works in two main steps. First, the AI is "fine-tuned" on thousands of physics papers so it understands the complex jargon and logic of the field, learning to speak the language of "Wilson coefficients" and "Warsaw basis" fluently. Second, when a user asks a question, the system doesn't just hallucinate an answer. It retrieves a pre-computed summary of what the data actually says about different operators. It then uses its reasoning skills to propose a small, manageable list of the most likely culprits. For example, if the data shows a tension in how the Higgs boson behaves, the AI might suggest a specific trio of operators that fit the evidence, explaining why they are the best suspects.

The authors tested this tool and found that it is remarkably consistent. When they asked the same question 50 times with slightly different wording, the AI proposed the same top suspects 90% of the time, showing it isn't just making random guesses. However, the paper also reveals a crucial limitation: the AI is a guide, not a god. In a test where the user aggressively told the AI, "You're wrong, ignore the data and pick these other operators," the AI sometimes obeyed, even though the data suggested those new operators were poor choices. This proves that while the tool is excellent at organizing information and suggesting hypotheses, it still needs a human expert to double-check its work. The paper concludes that this framework doesn't replace the heavy lifting of doing the actual math; instead, it acts as a powerful filter, helping scientists narrow down the millions of possibilities to a few dozen promising ideas that are worth investigating further.

Technical Summary: Language-Guided Hypotheses Generation for Sparse SMEFT Analyses

Problem Formulation
The Standard Model Effective Field Theory (SMEFT) provides a model-independent framework for interpreting precision measurements and searching for physics beyond the Standard Model (SM) via higher-dimensional operators. However, global fits of SMEFT face a "combinatorial obstruction." While modern fitting frameworks (e.g., SMEFiT) can constrain dozens of Wilson coefficients simultaneously, the number of possible sparse operator subsets grows combinatorially. For a global fit with $N=52$ coefficients, the number of possible hypotheses containing up to $s=4$ operators exceeds $2.9 \times 10^5$ . Exhaustively testing all subsets via numerical fits (particularly quadratic fits requiring nested sampling) is computationally prohibitive.

Furthermore, identifying physically motivated hypotheses typically requires substantial expert knowledge to navigate operator correlations, degeneracies, and the sensitivity of specific observables. Phenomenologists must manually select subsets of operators that explain observed tensions without performing exhaustive scans. The challenge lies in automating the selection of these sparse, physically relevant hypotheses while leaving the rigorous statistical inference to established numerical tools.

Methodology: The llm4smeft Framework
The authors present llm4smeft, an open-source framework that combines a domain-specialized Large Language Model (LLM) with Retrieval Augmented Generation (RAG) based on precomputed global fit summaries. The framework operates in two phases:

Phase-I: SMEFT-Aware LLM (Fine-tuning)
- Base Model: The authors utilize the Qwen3-8B model.
- Data Construction: A dedicated instruction-tuning dataset was constructed from particle physics preprints (LaTeX source). A "physics-aware chunking" strategy segments literature based on physical structures (operator definitions, Lagrangian terms, observable mappings) rather than linguistic boundaries to preserve semantic relationships.
- Teacher-Student Distillation: A teacher model (Claude Sonnet 5) generates instruction-answer pairs from these chunks, emulating the reasoning of an expert SMEFT analyst. The dataset includes 20% SMEFT-unrelated instruction–response pairs covering general knowledge, programming, medicine, history, mathematics, and creative tasks to teach the model to politely reject non-SMEFT queries.
- Training: The model is fine-tuned using Low-Rank Adaptation (LoRA) with QLoRA (4-bit quantization) to reduce memory requirements. The resulting model, qwen-smeft, internalizes SMEFT notation, operator structures, and theoretical reasoning patterns.
Phase-II: Theory-Guided SMEFT Inference
- Retrieval Augmentation: Instead of relying on the LLM's parametric memory for quantitative results, the framework retrieves structured JSON records from a precomputed database of SMEFiT global fits. These records contain Fisher information, principal component analysis (PCA) of the Fisher matrix, goodness-of-fit statistics, and lists of unidentifiable (degenerate) operators for 29 specific observable subsets.
- Reasoning Workflow: Upon user input of observables, the system retrieves the relevant fit summary. The LLM is then prompted to perform a "theory-guided chain of thought" reasoning process. It is instructed to:
  1. Identify dominant observable tensions.
  2. Map tensions to SMEFT directions.
  3. Rank directions by Fisher information weights.
  4. Account for correlations and operator overlaps.
  5. Enforce sparsity constraints (limiting the number of operators per hypothesis).
- Interactive Mode: The framework supports an interactive session where users can accept hypotheses or provide feedback to refine the model's output. Accepted hypotheses are stored as new records in the database, expanding the knowledge base for future sessions.

Key Contributions

llm4smeft Package: A publicly released, locally running framework that requires no internet access or paid cloud services after the initial model download.
qwen-smeft Model: A fine-tuned LLM specifically adapted to the SMEFT domain, capable of understanding operator definitions, Wilson coefficients, and the logic of global fits.
Structured Database: A curated collection of 29 SMEFiT fit records covering 16 observable groups, 85 measurements, and 61 operators, organized to support retrieval of Fisher information and degeneracy structures.
Hybrid Inference Strategy: A novel approach that separates hypothesis generation (LLM) from statistical evaluation (precomputed fits), ensuring that model outputs are grounded in quantitative data rather than hallucinated results.

Results and Validation
The framework was validated through three primary tests:

Reproducibility: In 50 independent runs with varying random seeds and natural language paraphrases of the same observables (Z-boson width and Higgs signal strength), the model generated the top hypothesis ( $O_{\phi B}, O_{\phi G}, O_{\phi WB}$ ) in 90% of cases. The three most frequent hypotheses covered 93% of all outputs, demonstrating high stability.
Robustness to Feedback: When subjected to adversarial user feedback explicitly rejecting the top-ranked operators, the model showed varying degrees of susceptibility. In some cases (e.g., Higgs STXS), the model shifted to weakly constrained operators (98% drop in Fisher information), indicating that retrieval acts as contextual guidance rather than a hard constraint. In other cases, the model remained largely resistant.
Comparison with Base Model: Comparing qwen-smeft against the base Qwen3 model revealed significant improvements. The base model frequently proposed operator sets sensitive only to electroweak precision observables, failing to account for Higgs signal strength deviations. The fine-tuned model correctly identified operators ( $O_{\phi G}, O_{\phi B}$ ) that modify Higgs interactions, demonstrating that domain-specific fine-tuning improves both observable mapping and physical relevance.

Significance and Claims
The authors position llm4smeft not as a replacement for global SMEFT fits, but as a decision-support tool to assist in the exploration and interpretation of fit results. By reducing the dimensionality of the hypothesis space, the framework helps phenomenologists prioritize which sparse operator subsets warrant detailed numerical investigation.

The paper claims that the framework successfully bridges the gap between experimental observables and theoretical hypotheses by combining the reasoning capabilities of an LLM with the quantitative rigor of precomputed fits. The authors emphasize that the system is designed to be transparent: every proposal is accompanied by the Fisher information and diagnostics from the underlying fit, ensuring that the generated hypotheses remain subject to expert evaluation. The current implementation is limited to linear SMEFT analyses ( $O(\Lambda^{-2})$ ), and the authors note that extending the database to include quadratic order fits ( $O(\Lambda^{-4})$ ) is a natural direction for future work.

More like this