Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis

arXiv:2606.10381 · hep-ex, cs.AI, cs.CL, cs.IR, physics.ins-det · Submitted 2026-06-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis".

Tom: Agentic hybrid RAG presents an evidence-grounded framework for scientific question answering in muon collider research,

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So we’ve covered a lot about this paper, "Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis," and we’ve seen how it uses a hybrid retriever combined with agentic reasoning to tackle the challenges of finding precise, well-grounded answers in high-energy physics literature.

Jane: It really boils down to the authors' claim that integrating these two components results in better retrieval effectiveness and higher quality answers than standard RAG baselines for this specific type of research. They are focused on making sure the AI is not just pulling text, but actively using that text to construct a sound argument.

Lu: The implication we see here is that for fields where evidence is heterogeneous, like muon collider research spanning accelerator physics and detector design, a system that can intelligently decompose complex queries into manageable retrieval steps offers a path toward more systematic scientific inquiry.

Meng: I’m thinking about the practical takeaway: this approach provides a blueprint for building analysis assistants that are more reliable because they enforce strong grounding constraints during the answer generation phase, ensuring the output strictly relies on what was retrieved.

Lalam: My perspective is that this work gives us a model for how AI can be used to organize vast scientific knowledge bases effectively, which could seriously enhance how our community manages and utilizes the existing body of literature.

Tom: Absolutely, Lalam; it’s about creating a system where the reasoning isn't just unstructured exploration but is carefully applied to organization and synthesis of that retrieved evidence. It’s an advancement in how we can make AI assistants useful for actual scientific discovery workflows.

Conclusion: Tom: So, we’ve seen how this Agentic Hybrid RAG framework works in practice, and now we need to wrap up by really focusing on what this paper actually means for us in the long run.

Jane: Exactly, Tom; it's important to take a moment to really look at the title itself—"Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis"—and understand what that combination of ideas actually signifies.

Lu: From a research standpoint, this paper shows how we can move beyond simple information retrieval and build systems where the AI actively reasons about how to gather and synthesize evidence across complex, specialized domains like HEP.

Meng: I’m thinking about the practical impact; if this framework can consistently pull high-quality evidence grounded in specific technical language, it could significantly reduce the time engineers spend manually cross-referencing documents.

Lalam: I see a huge cultural shift here, Tom and Jane; having an AI that doesn't just give you an answer but shows its work through structured decomposition and strict grounding really elevates the standard of what we expect from scientific tools.

Tom: That’s a great way to put it, Lalam; moving from passive search to active evidence organization is a big deal for how we use these tools daily.

Jane: I agree with Lalam; thinking about the authors, they clearly focused on building this system step-by-step, which makes the methodology very transparent for us as listeners trying to understand it.

Lu: The authors did a smart job of fusing sparse lexical retrieval with dense semantic retrieval using that weighted rank fusion score; it’s a solid technical choice for handling both exact jargon and conceptual similarity simultaneously.

Meng: And from an engineering side, the system's pipeline—decomposition followed by independent hybrid retrieval—is very structured, which means we can actually debug where a query might be failing in the evidence gathering process.

Lalam: The implication I see is that this method provides a blueprint for how future AI tools should structure their internal processes when dealing with highly specialized, jargon-heavy scientific literature.

Tom: So, it’s not just about getting answers; it’s about building a more rigorous and trustworthy mechanism for extracting knowledge from the data we collect on arXiv.

Jane: Right; and that leads us perfectly into how these kinds of evidence-grounded systems might fundamentally alter the way physicists approach detector design and background studies in the next few years.

State Key Laboratory of Nuclear Physics and Technology, Peking University

hep-ex, cs.AI, cs.CL, cs.IR, physics.ins-det

Submitted: 2026-06-09

Updated: 2026-09-28

Comments: 23 pages, 5 figures, and 6 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Agentic hybrid RAG presents an evidence-grounded framework for scientific question answering in muon collider research, combining hybrid retrieval with agentic reasoning to improve retrieval

Key concepts

Hybrid Retrieval Backbone
This component combines two search methods: sparse lexical retrieval (BM25) to find exact technical terms like acronyms, and dense semantic retrieval using embeddings (sentence-transformers) to find conceptually similar ideas. They are fused using a weighted reciprocal rank fusion score to get the best of both worlds.
Agentic Query Decomposition
For hard questions, an agent uses three sequential language model prompts to break the main query into smaller, focused subqueries. This helps explore different angles like mechanism or motivation. Each subquery is then searched independently by the hybrid retriever.
Evidence-Grounded Answer Generation
The final answer generation step strictly limits the language model's output to only use information found in the retrieved chunks. It forces the model to cite evidence and abstain if it lacks sufficient proof, ensuring factual accuracy for scientific applications.

Terminology

Summary

Agentic hybrid RAG presents an evidence-grounded framework for scientific question answering in muon collider research, combining hybrid retrieval with agentic reasoning to improve retrieval effectiveness and answer quality over standard baselines.

The gist

Agentic hybrid RAG consistently outperforms representative retrieval and RAG baselines in retrieval effectiveness, answer quality, evidence coverage, and factual grounding by integrating a hybrid retriever with an agentic reasoning module for query decomposition and evidence expansion.

Hybrid Retrieval Backbone

The framework utilizes a hybrid retriever combining sparse lexical retrieval (BM25) and dense semantic retrieval to create a robust evidence backbone. The sparse component uses BM25 to preserve exact technical terminology commonly used in HEP literature, effectively handling acronyms like BIB, MDI, VBS, and aQGC. The dense component employs sentence-transformers/all-MiniLM-L6-v2 embeddings indexed with FAISS under cosine similarity to capture conceptual similarity even when surface forms differ significantly. These two signals are fused using the weighted reciprocal rank fusion (RRF) score, SRRF(c) = w d r d(c) + w s r s(c), where the default configuration sets wd = 0.9 and ws = 0.1, reflecting a stronger semantic coverage of dense retrieval while using sparse retrieval as a high-precision fallback for exact terminology.

Agentic Query Decomposition

For complex scientific queries, the system employs an agentic query expansion layer to decompose the original query into targeted subqueries. This process involves three sequential lightweight language model prompts: first, a domain tagging prompt identifies relevant physics domains using a controlled vocabulary; second, a query classification prompt assigns the query to one of three strategies—precise fact, broad synthesis, or reasoning; and third, a subquery generation prompt produces the final set. For instance, for reasoning queries, the decomposition is structured along mechanism, motivation, and limitation dimensions. The resulting subqueries are independently processed by the hybrid retrieval pipeline to produce a ranked list of candidate chunks. These retrieved lists are then merged via taking the union and deduplicating by chunk identifier to form a final evidence pool constrained to a fixed budget of top-M chunks.

Evidence-Grounded Answer Generation

The answer generation module receives the original query and the consolidated evidence set from both the original query and its decomposed subqueries. The language model is instructed to produce responses strictly grounded in the provided chunks, to cite supporting evidence, and to abstain when the retrieved material is insufficient. This grounding constraint is crucial for detector and physics applications. When multiple subqueries retrieve overlapping evidence, duplicates are removed while preserving the highest-scoring occurrence according to the hybrid ranking function. The final output is conditioned on this constrained evidence set to produce the response.

Evaluation and Results

The framework was evaluated using a self-constructed benchmark comprising a retrieval benchmark (58 questions) and an answer-generation benchmark (40 questions). Retrieval performance metrics, such as Precision@1, Recall@5, MRR, and gNDCG@5, demonstrated that the hybrid retriever achieved the strongest overall performance, outperforming both standard BM25 and dense vector retrieval. However, end-to-end answer generation results showed that Agentic Hybrid RAG achieved the strongest overall answer-generation performance, with a Good Rate increasing from 50.0% (Vanilla RAG) to 60.0% and Key-Point Coverage rising substantially from 55.1% to 79.3%, indicating substantially more complete utilization of retrieved evidence. The study concludes that agentic reasoning is most effective when applied to evidence organization, contextualization, and answer synthesis, rather than as an unconstrained replacement for retrieval.

System Overview

The proposed system follows a pipeline organized into three tightly coupled stages: agentic query decomposition, hybrid retrieval, and evidence aggregation. The purpose of the decomposition stage is to explicitly expand the original information need into multiple complementary retrieval perspectives, increasing coverage over heterogeneous evidence sources. The second stage applies the same hybrid retrieval pipeline to each decomposed subquery independently. The third stage merges all retrieved chunks into a unified evidence pool through deduplication and rank-based selection, ensuring a compact and controllable context size for downstream generation. This end-to-end design integrates decomposition-driven query expansion, hybrid sparse-dense retrieval, and evidence fusion.

Limitations and Outlook

The work acknowledges limitations, including the fact that the benchmark is self-constructed and that answer evaluation relies partly on an LLM-as-a-judge framework. Future work should focus on incorporating community-reviewed benchmarks and assessing the framework in realistic scientific workflows, such as detector-background studies, to provide a more direct measure of its utility for future HEP analysis agents.

Improvements for AI systems

Here are specific, actionable improvements derived from the Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis paper, detailing what an improved AI system could achieve:

)Improved AI System Capabilities: Agentic Hybrid RAG for High-Stakes Scientific Literature Analysis (Muon Collider Domain)

The proposed framework moves beyond simple retrieval to a sophisticated evidence-aware reasoning agent, capable of performing complex, multi-step scientific inquiry that requires cross-referencing fragmented knowledge across massive technical corpora.

Here are the specific improvements and resulting capabilities:

Abstract

Muon collider research spans accelerator physics, detector instrumentation, and high-energy phenomenology, with relevant evidence scattered across a rapidly expanding and heterogeneous body of scientific literature. As high-energy physics (HEP) increasingly explores agent-assisted analysis workflows, efficiently locating, integrating, and verifying scientific evidence becomes an essential capability. While retrieval-augmented generation (RAG) offers a promising framework for scientific question answering, integrating agentic reasoning without compromising retrieval precision remains a key challenge. In this work, we present agentic hybrid RAG, an evidence-grounded RAG framework for muon collider research. The framework combines a hybrid retriever, integrating sparse lexical and dense semantic retrieval, with an agentic reasoning module for query decomposition, evidence expansion, and grounded answer generation. To enable systematic evaluation, we construct the first benchmark for retrieval-augmented scientific question answering in the muon collider domain, comprising a curated literature corpus together with dedicated retrieval and answer-generation benchmarks covering major detector and physics research topics. Extensive evaluation shows that hybrid retrieval provides the strongest retrieval backbone, while agentic reasoning is most effective for controlled evidence expansion and answer synthesis. Built on this principle, agentic hybrid RAG consistently outperforms representative retrieval and RAG baselines in retrieval effectiveness, answer quality, evidence coverage, and factual grounding. Together, the benchmark and framework provide a foundation for evidence-grounded scientific question answering and future HEP analysis agents operating over large-scale scientific literature. Code is available at this URL.

Sources

Related papers