DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII

arXiv:2604.22541 · hep-ex · Submitted 2026-04-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII".

Mira: Here's the gist of the paper. High-energy physics (HEP) analysis traditionally demands extensive manual effort to navigate complex software environments and maintain physical rigor By enabling physicists to define high-level physics goals while delegating workflow orchestration to autonomous agents, we present Dr.Sai, an AI-driven multiagent system designed to automate the end-to-end physics analysis workflow.

Kai: Dr.Sai was specifically engineered to handle real-world HEP tasks by bridging highlevel physics intent with the intricate software stacks of the BESIII experiment.

Mira: System Architecture and Agents The Dr.Sai system utilizes a Multi-Agent System (MAS) architecture integrated with a daemon-based message management system.

Kai: This MAS consists of specialized agents that collaborate to decompose and execute complex tasks, supported by base LLMs for foundational reasoning and natural language processing.

Title and authors: Mira: So we're looking at this paper called "DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII." It sounds like they've built something designed to take the heavy lifting out of analyzing data from experiments like BESIII.

Kai: Yeah, what they’re aiming for is to let physicists set a big physics goal and then the AI agents handle all the messy steps—the whole workflow—to get them results. It's about automating that end-to-end process so you don't have to manually juggle everything.

Lev: From an engineering standpoint, that means building a system where one agent talks to another, breaking down a huge task into smaller pieces, and those pieces actually run on real hardware without crashing. It’s not just about the idea; it’s about making sure the execution part is solid.

Mira: Exactly. The paper mentions they use a multi-agent system with six specialized agents—a Host, Planner, Coder, Tester, Calculator, and Reflector—all working together to manage this process. It's a very structured way to approach the analysis pipeline.

Kai: That structure is key because physics workflows are incredibly complex; you have to make sure every step depends correctly on the last one before you move on. The Planner agent seems to be central in making sure that hierarchy of tasks makes sense for a specific experiment, like cross-section measurements.

Lev: I worry about the actual execution speed and overhead when you put all those layers of reasoning and tool calling in place; real hardware needs to handle that without massive latency spikes. How much time does this whole system actually take compared to a human doing it?

Mira: The paper says they validated this by re-measuring branching fractions across ten different J/psi decay channels, and the workflow successfully automated the entire BESIII analysis chain, from the initial idea to estimating systematic uncertainties. It shows they got results that matched Monte Carlo simulations well.

Kai: That's a good point about matching MC simulations; it confirms that what you get out of this system is physically consistent with what we expect from theory and simulation benchmarks. But the real question is, how does the AI handle those systematic uncertainties, which are often tricky in real experiments?

Lev: That’s where the Calculator agent comes in, I think. It uses a fixed formula for calculating branching ratios that includes statistical and systematic error components derived directly from the input data you feed it. It’s a concrete function, not just some abstract reasoning.

Mira: Right, and the paper found that when they tested five different large language models, adding a moderate reflection allowance really improved the reliability of those sub-tasks for most of them. So it's not just one model being perfect; it's about tuning how the AI thinks about its own work.

Title and authors: Kai: That brings us to what they found is actually a bit tricky for all these models, which is that the main performance bottleneck isn't necessarily in understanding the physics information itself, but in the functional calling capability—that’s initiating and using those scientific tools or code correctly.

Lev: If tool calling is the weak spot, then fine-tuning for that syntax level mastery of professional tools seems like a huge area for future work. It suggests that just having good reasoning isn't enough if the AI can't actually write the right executable sequence to get the job done on the real data processing software.

Mira: I agree, and I think they point out that GPT-4o does pretty well at calling those tools, but it struggles when it has to synthesize those precise parameters needed for execution, which is a different kind of failure mode <ref:2604.22541#pg1>. It's not about knowing *what* to ask the tool; it’s about getting the exact syntax right for the tool to run.

Kai: So if we look at this paper, what they suggest as improvements is focusing heavily on making those tools and code generation layers much more specialized, maybe with fine-tuned models that have a deeper grasp of professional HEP software syntax. That moves beyond just general intelligence to domain-specific execution precision.

Lev: From my side, I think the biggest challenge they've highlighted is how to take all that static knowledge—the complex physics rules and experimental procedures—and transform it into executable logic that actually mirrors how a human expert makes those strategic decisions. That translation step is where the real difficulty lies for autonomous agents.

Mira: It’s about embedding domain-specific knowledge so the AI can make those high-level, expert choices, not just follow a pre-set script blindly when things get messy in the data. It moves from following instructions to actually making judgments that look like an expert's judgment.

Kai: So to wrap up this part of the paper, DrSai is presented as a viable framework for using multi-agent systems to accelerate discovery in complex experimental settings at BESIII, proving it can maintain physical consistency during the analysis chain. But they are clear that the next step requires more than just better reasoning; it needs better execution control.

Lev: And I'd add that on the hardware side, we need to see how this entire agentic loop runs under real constraints, because those constraints—like memory and clock speed—can totally change what’s feasible for these complex workflows.

Mira: So, the implication is that while AI is great at handling the workflow orchestration, achieving true autonomy means solving the problem of perfect translation from expert thought into machine-executable code with high fidelity. That's where we head next.

The paper's summary: Kai: So, to recap, this paper is about building an AI system called DrSai that’s designed to take physicists’ high-level physics goals and run the entire complex analysis workflow at a place like BESIII all by itself.

Mira: Exactly. It's not just a chatbot; it's a whole team of specialized agents, each one handling a specific part of the analysis, from planning the steps to doing the actual calculations.

Lev: From what I see on page two, they’re using this multi-agent system structure with six distinct agents collaborating through some message management system. It sounds like they've mapped out a very rigid chain of command for an analysis task.

Kai: Right, and that planning part is really interesting because they use something called a hierarchical task management system that’s inspired by fractal structures, which means they break the big physics objective down into smaller and smaller subtasks.

Mira: That decomposition is crucial because it makes tracking dependencies much cleaner when you have a massive workflow like this. It simplifies managing all those different steps in order.

Lev: I want to talk about the actual coding part for a second, because they mention the Coder agent uses configuration driven code generation based on a JSON file that specifies things like decay channels and selection criteria.

Kai: That’s where it gets interesting, because they are creating this HepScript language which is basically a domain specific language meant to translate human analytical logic directly into executable sequences for the analysis software.

Mira: It’s about formalizing the physical intent, so you take that high level goal and you turn it into a standardized set of rules that the machine can follow without ambiguity.

Lev: That brings up a real practical question: how does this system actually work with real codebases like ROOT? Because if it’s just generating text, it won't run on the actual data processing software.

Kai: Well, the paper shows they validated this by running the whole chain from a natural language objective all the way to signal extraction and even preliminary systematic uncertainty estimation in ten different J psi decay channels.

Mira: That experimental validation is important because it proves that DrSai can maintain physical consistency across these different measurement types, which is exactly what we need in real physics.

Lev: And they did a performance study on five mainstream large language models, and the results show that even with some reflection allowance, most models got more reliable for subtasks.

Kai: But the authors were very clear about where it stops working; the main bottleneck across all those tested models wasn't about how much information they understood, but actually initiating those functional calls to external tools.

Mira: So, the lesson there is that having deep reasoning isn't enough if the AI can’t correctly generate the precise parameters needed for a tool to execute its job properly.

Lev: That points toward the future work they’re suggesting: we need better expertise in tool calling and code generation, moving towards models that have a syntax level mastery of professional physics tools.

Kai: And they also mentioned how hard it is to get domain specific knowledge embedded so the AI can make those strategic decisions that look like a human expert making choices about where to focus the search.

Mira: It’s about transforming static information—the rules and procedures—into executable logic that mirrors how a real scientist makes high level decisions.

Lev: So, the next challenge isn't just making the AI smarter, it's making it better at precise execution and having a proper way to embed all that specialized physics knowledge into its thinking structure.

The paper's improvements: Kai: So, we've seen how DrSai works right now, and now we look at what they think needs to happen next to make it better than just a powerful tool for a chat assistant.

Mira: Right, the paper points out that there are two main areas they want to focus on improving: making the AI much better at calling those professional tools and improving how it handles code generation.

Lev: I see that as needing specialized synthesis layers or maybe fine-tuned models that actually have a syntax level mastery of things like ROOT or other HEP software. It’s about getting the code and parameters exactly right the first time.

Kai: That makes sense because if the AI can't generate schema compliant parameters, it just produces garbage for execution, no matter how good its high level reasoning is.

Mira: And then there’s that second big piece of work: they want to figure out how to embed a huge amount of domain specific knowledge into the agentic architecture.

Lev: So it's about taking all that static information—the complex physics rules and experimental procedures—and turning it into executable logic, something that actually mimics how a human expert makes strategic choices on the fly.

Kai: I’m thinking about what this means for us in the hardware side: if we can get better tool calling, maybe we can start building specialized interfaces or models specifically tuned for those physics tools.

Mira: That would be huge because it moves us away from just general intelligence toward something that understands the specific language of experimental data analysis.

Lev: From my point of view, if the AI could generate perfectly compliant code blocks, running that on real hardware becomes much more reliable and less prone to catastrophic failure during complex jobs.

Kai: So the goal isn't just a smarter reasoning engine; it’s building a system where the execution layer is just as precise as the decision-making layer.

Mira: Exactly, and they suggest focusing on how to represent those expert heuristics so that when DrSai decides what to do, it’s making choices that actually look like a seasoned physicist.

Lev: It sounds like the next step is less about increasing the reasoning depth and more about fixing the precision of execution and deepening that domain integration.

Kai: That leads us nicely into how this architecture might be tested further, looking at those specific performance metrics they used, like success rates under different retry budgets.

Mira: We need to look closely at those limitations mentioned—what the current system can’t do—to see if those proposed improvements actually fix the core problems they identified.

Conclusion: Kai: So, we've covered how DrSai uses a multiagent system to automate analysis workflows at BESIII, showing it can handle complex physical consistency across different decay channels like J psi decays and estimate systematic uncertainties.

Mira: Right, and the core idea is that this AI doesn't just follow instructions; it uses a structured decomposition of tasks to manage the entire experimental process from an initial physics goal down to preliminary data fitting.

Lev: From my side, it’s interesting that they found a performance bottleneck in functional calling capability rather than just reasoning depth across all the tested large language models.

Kai: So, if we look at what this paper suggests for the future, it really shifts the focus toward improving how the AI actually interacts with external tools and generating precise code.

Mira: They’re pushing for specialized synthesis layers or fine-tuned models that have a syntax level mastery of professional physics tools, which would make it much better at writing executable sequences.

Lev: That seems like a solid direction because if you can get the AI to write accurate code blocks based on those configuration files, the overhead lands much more predictably on real hardware.

Kai: I think the biggest long term challenge they’ve laid out is getting that domain specific knowledge embedded so the AI can make strategic decisions that look like an expert physicist's judgment.

Mira: It’s about taking all that complex static information and making it into logic, not just following a script blindly.

Lev: So the next hurdle isn't just more reasoning power; it’s building a proper way to transform expert thought into machine executable code that mirrors actual decision-making.

Kai: It’s a big step toward having these agents act as competent partners in the field, rather than just assistants for asking questions.

Mira: Yes, and seeing this validated at BESIII gives us real confidence that this agentic approach can handle the rigor of real-world experimental frameworks.

Lev: We’ll keep watching how they tackle those limitations on functional calling because that precision is what separates a good assistant from an actual tool for heavy lifting.

Institute of High Energy Physics, CAS Institute of High Energy Physics, University of Chinese Academy of Sciences, Jilin University, Lanzhou University, Tsung-Dao Lee Institute Shanghai Jiao Tong University, Institute of Nuclear and Particle Physics School of Physics and Astronomy Shanghai Jiao Tong University Key Laboratory for Particle Astrophysics and Cosmology MOE

hep-ex

Submitted: 2026-04-24

Updated: 2026-11-12

Comments: 56 pages, 7 figures

Journal ref: Science Bulletin, 2026

DOI: 10.1016/j.scib.2026.10.072

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: High-energy physics (HEP) analysis traditionally demands extensive manual effort to navigate complex software environments and maintain physical rigor By enabling physicists to define high-level…

Key concepts

Multi-Agent System (MAS)
A MAS uses several specialized AI agents that collaborate to solve a complex problem. In DrSai, agents like the Planner and Coder work together—the Planner sets the physical goal, while the Coder writes standardized analysis code—to automate end-to-end research tasks.
Hierarchical Task Management
This system breaks large physics goals into smaller, manageable sub-tasks organized in a selfsimilar structure. This ensures that dependencies are tracked correctly as complex workflows are decomposed from a single high-level objective down to executable code generation.
HepScript
HepScript is a domain-specific language created to formalize human analytical logic. It allows physicists to translate abstract physical intentions into standardized, executable sequences that the AI agents can process and implement within the analysis framework.

Terminology

Summary

High-energy physics (HEP) analysis traditionally demands extensive manual effort to navigate complex software environments and maintain physical rigor By enabling physicists to define high-level physics goals while delegating workflow orchestration to autonomous agents, we present Dr. p. 0Sai, an AI-driven multiagent system designed to automate the end-to-end physics analysis workflow. p. 0 Dr.Sai was specifically engineered to handle real-world HEP tasks by bridging highlevel physics intent with the intricate software stacks of the BESIII experiment. p. 0

System Architecture and Agents

The Dr.Sai system utilizes a Multi-Agent System (MAS) architecture integrated with a daemon-based message management system. p. 2 This MAS consists of specialized agents that collaborate to decompose and execute complex tasks, supported by base LLMs for foundational reasoning and natural language processing. p. 2 The operations are orchestrated by a central controller that assigns tasks to specialized remote workers. p. 2 The system comprises six agents: the Host, Planner, Coder, Tester, Calculator, and Reflector. p. 2

Workflow Formalization

Dr.Sai employs a hierarchical task management system inspired by fractal-like, selfsimilar structures. p. 4 Large-scale research objectives are decomposed into a one-to-many hierarchy of sub-tasks to simplify the management of complex workflows and ensure dependency tracking. p. 4 The Coder agent utilizes configuration-driven code generation where analysis components are standardized into reusable templates based on a high-level JSON configuration specifying parameters like decay channels and selection criteria. p. 5 Furthermore, HepScript is developed as a domain-specific language to formalize human analytical logic, translating high-level physical intent into standardized executable sequences. p. 5

Validation and Performance

The system was validated with the re-measurement of branching fractions across ten distinct J/ψ decay channels. The execution workflow automates the complete BESIII analysis chain, from initial natural language objective to signal extraction and preliminary systematic uncertainty estimation. p. 6 A systematic performance evaluation across five mainstream LLMs demonstrated that a moderate reflection allowance significantly enhances sub-task reliability for most models. p. 1

Key Findings on Model Limitations

The primary performance bottleneck across all evaluated models is functional calling capability rather than information extraction. p. 9 For DeepSeek-R1, the diminishing returns of iterative reflection stem from a fundamental weakness in tool-call initiation, suggesting that extensive reasoning depth cannot compensate for a lack of functional precision. Conversely, GPT-4o excels at tool-call initiation but falters during the synthesis of schema-compliant parameters required for execution. p. 5

Future Directions

Future developments will focus on two core dimensions: Expertise in Tool-Calling and Code Generation, moving toward specialized synthesis layers or fine-tuned models that possess a “syntax-level” mastery of professional HEP tools. Additionally, research will prioritize how to effectively represent and embed a large amount of domain-specific knowledge into the agentic architecture to transform static information into executable logic that mirrors the strategic decision-making of a human expert. p. 2

Conclusion

In this work, we have developed and validated Dr.Sai, an autonomous multi-agent system tailored for high-energy physics analysis. p. 1 The experimental validation at BESIII confirms that Dr. p. 1Sai can maintain physical consistency and handle the rigorous demands of real-world experimental frameworks. p. 0

Conflict of Interest

The authors declare that they have no conflict of interest.

Acknowledgements

The authors thank BESIII Collaboration for their support on simulation datasets and software. The authors also thank the IHEP computing center for their strong support on the computing resources. p. 1 The work was supported by the Strategic Priority Research Program of Chinese Academy of Sciences under Grant XDA0480600. p. 0 F. Jiang was also supported by 2023 CAS Outstanding Talent Program for Library and Information Science E429C2M1.

Appendix A Agent Configuration

The Host agent's system prompt dictates that it can only do two things: first, respond to questions directly; second, call the built-in tool ‘select expert’ to consult experts. The Planner is specialized in physical analysis tasks of the BESIII experiment and must strictly use a standardized task template for cross-section measurement tasks. p. 2 The Coder agent's system prompt restricts it to creating code blocks and variable cards, not executing code or performing other tasks.

Appendix B Fit Results

Appendix B presents the experimental data fitting distribution plots of the ψ(2S) → π+π−[J/ψ → X] decay processes involved in the main text. These plots intuitively demonstrate the signal extraction and background fitting results of each process. p. 9 The analysis incorporates common BESIII systematic uncertainties, including those from charged-track reconstruction, particle identification [37–39], and the total used Nψ(2S) count [27]. p. 9

Appendix C Calculator Function

The Calculator agent is designed to complete physical result calculations or generate Python code based on the information provided by the user. It uses a fixed formula for branching ratio calculation, which includes statistical and systematic error components derived from the input data. p. 9 The output format template specifies how to print the final result including statistical and systematic errors. p. 9 p. 9

Appendix D Tester Function

The Tester agent's primary responsibilities are to execute pre-configured code via existing JSON variable cards, execute code blocks wrapped by a specific format, or print root variables. It cannot create/modify/generate JSON cards or code blocks, select variables/paths, or perform any other tasks. p. 5

Appendix E Planner Template

The Task Planning Principles require the Planner to first assess whether the user’s request is relevant to specific physical analysis tasks of the BESIII experiment. For cross-section measurement tasks, it mandates a twelve-step process including creating Analysis Algorithm JSON Variable Card, executing Built-in Script and Generate Analysis Algorithm Program, and creating JobOption Script JSON Variable Card.

Appendix F Coder Template

The Coder agent's system prompt requires it to strictly follow the Task Execution Steps, prioritizing tool self-check before task judgment to ensure adherence to the single tool constraint. It mandates that generated code must accurately implement the user’s demand, including necessary comments explaining the physical logic. p. 5

Appendix G Tester Template

The Tester agent's task processing flow requires it to first check for code blocks or JSON variable cards before executing them immediately. If neither is present, it checks if its tools match the user’s needs and performs the corresponding task if explicitly requested. p. 6

Appendix H Calculator Template

The Calculator agent's execution steps involve Intention Judgment, Formula Retrieval, Parameter Query, and then performing result calculation or generating Python code based on the provided information. The output format template specifies how to print the final result including statistical and systematic errors.

Appendix I Summary of Performance Study

This study presents a multi-dimensional performance evaluation of the Dr.Sai system, benchmarking five mainstream LLMs across HEP measurement tasks. p. 0 By analyzing success-rate heatmaps under varying retry budgets, failure-mode distributions, and token expenditure, we demonstrate that a moderate reflection allowance significantly enhances sub-task reliability for most models. p. 8 The primary performance bottleneck across all evaluated models is functional calling capability rather than information extraction.

Appendix J Limitations and Future Outlook

The systematic evaluation of Dr.Sai indicates that the current bottlenecks in autonomous HEP research are the precision of execution and the depth of domain integration. p. 1 Future developments will focus on Expertise in Tool-Calling and Code Generation, moving toward specialized synthesis layers or fine-tuned models that possess a “syntax-level” mastery of professional HEP tools. A deeper challenge lies in the current system’s lack of a formal representation of expert heuristics to transform static information into executable logic that mirrors the strategic decision-making of a human expert. p. 5

The transition from a chat-based assistant to a task-driven autonomous scientist requires a deeper focus on how expert heuristics are represent and embedded within the agentic architecture. Future iterations of Dr. p. 5Sai will prioritize the development of more sophisticated domain-synthesis layers and the expansion of the underlying physics ontology through HepScript-2. p. 50. Ultimately, this research paves the way for a new era of collaborative scientific discovery where AI agents act as competent partners in navigating the complexities of the subatomic world. p. 1

References

[1] Ablikim, M., et al.: Design and Construction of the BESIII Detector. p. 7 p. 7 p. 7 Nucl. Instrum. Meth. A 614, 345–399 (2010) https://doi.org/10.1016/j.nima.2009.12-05- arXiv:0911-4960 [physics. p. 0ins-det]

[2] Ablikim, M., et al.: Future Physics Programme of BESIII. p. 6 p. 6 p. 6 Chin. Phys. C 44(4), 040001 (2020) https://doi.org/10.1188/1674-1137/44/4/04- arXiv:1912-5983 [hep-ex]

[3] Brun, R., Rademakers, F.: ROOT: An object oriented data analysis framework. p. 2 p. 2 p. 2 Nucl. Instrum. Meth. A 389, 81–86 (1997) https://doi.org/10.1016/S0168-9002(97)00048-X

[4] Zou, J., et al.: Offline data processing system of the BESIII experiment. p. 1 p. 1 p. 1 Eur. Phys. J. C 84(9), 937 (2024) https://doi.org/10.1140/epjc/s10052-024-13241-3

[5] Zhang, Z., et al.: Dr. sai: Physical analysis agents based on llms for besiii experiment and exploration of future ai scientist. p. 5 p. 5 p. 5 In: 24th International Conference on High Energy Physics (ICHEP 2024), Prague, Czech Republic. p. 0 p. 0 p. 0 Presented by Z. Zhang on behalf of Dr.Sai working group https://indico. p. 7 p. 7 p. 7cern.ch/event/1291157/contributions/5889603/

[6] Li, K., Liu, B., Mellado, B., Yuan, C.-Z., Zhang, Z.: AI agents, language, deep learning, and the next revolution in science. p. 1 p. 1 p. 1 Front. Phys. (Beijing) 21(9), 096401 (2026) https://doi. p. 0 p. 0 p. 0org/10.15302/frontphys-2026-096401 arXiv:2603-794 [hep-ex]

[7] Han, S., et al.: LLM Multi-Agent Systems: Challenges and Open Problems (2025). p. 1 p. 1 p. 1 https://arxiv. p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0org/abs/2402.03578

[8] Xi, Z., et al.: The Rise and Potential of Large Language Model Based Agents: A Survey (2023). p. 4 p. 4 p. 4 https://arxiv.org/abs/2309-07864

[9] Lewis, P., et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2021). p. 2 p. 2 p. 2 https://arxiv.org/abs/2005-11401

[10] Wu, Q., et al.: AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023). p. 2 p. 2 https://arxiv.org/abs/2308-8155 [hep-ex]

[11] Christakopoulou, K., Mourad, S., Matari´c, M.: Agents Thinking Fast and Slow: A Talker-Reasoner Architecture (2024). p. 2 https://arxiv.org/abs/2410-08328

[1] Ablikim, M., et al.: Design and Construction of the BESIII Detector. Nucl. Instrum. Meth. A 614, 345–399 (2010) https://doi.org/10.1016/j.nima-2009-12-05 arXiv:0911-496

[2] Ablikim, M., et al.: Future Physics Programme of BESIII. Chin. Phys. C 44(4), 040001 (2020) https://doi.org/10.1188/1674-1137/44/4/04- arXiv:1912-5983

[3] Brun, R., Rademakers, F.: ROOT: An object oriented data analysis framework. Nucl. Instrum. Meth. A 389, 81–86 (1997) https://doi.org/10.1016/S0168-9002(97)00048-X

[4] Zou, J., et al.: Offline data processing system of the BESIII experiment. Eur. Phys. J. C 84(9), 937 (2024) https://doi.org/10.1140/epjc/s10052-024-13241-3

[5] Zhang, Z., et al.: Dr. sai: Physical analysis agents based on llms for besiii experiment and exploration of future ai scientist. In: 24th International Conference on High Energy Physics (ICHEP 2024), Prague, Czech Republic. Presented by Z. Zhang on behalf of Dr.Sai working group https://indico.cern.ch/event/1291157/contributions/5889603/

[6] Li, K., Liu, B., Mellado, B., Yuan, C.-Z., Zhang, Z.: AI agents, language, deep learning, and the next revolution in science. Front. Phys. (Beijing) 21(9), 096401 (2026) https://doi.org/10.15302/frontphys-2026-096401 arXiv:2603-794

[7] Han, S., et al.: LLM Multi-Agent Systems: Challenges and Open Problems (2025). https://arxiv.org/abs/2402.03578

[8] Xi, Z., et al.: The Rise and Potential of Large Language Model Based Agents: A Survey (2023). https://arxiv.org/abs/2309-07864

[9] Lewis, P., et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2021). https://arxiv.org/abs/2005-11401

[10] Wu, Q., et al.: AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023). https://arxiv.org/abs/2308-8155

[1] Ablikim, M., et al.: Design and Construction of the BESIII Detector. Nucl. Instrum. Meth. A 614, 345–399 (2010) https://doi.org/10.1016/j.nima-2009-12-05 arXiv:0911-496

[2] Ablikim, M., et al.: Future Physics Programme of BESIII. Chin. Phys. C 44(4), 040001 (2020) https://doi.org/10.1188/1674-1137/44/4/04- arXiv:1912-5983

[3] Brun, R., Rademakers, F.: ROOT: An object oriented data analysis framework. Nucl. Instrum. Meth. A 389, 81–86 (1997) https://doi.org/10.1016/S0168-9002(97)00048-X

[4] Zou, J., et al.: Offline data processing system of the BESIII experiment. Eur. Phys. J. C 84(9), 937 (2024) https://doi.org/10.1140/epjc/s10052-024-13241-3

[5] Zhang, Z., et al.: Dr. sai: Physical analysis agents based on llms for besiii experiment and exploration of future ai scientist. In: 24th International Conference on High Energy Physics (ICHEP 2024), Prague, Czech Republic. Presented by Z. Zhang on behalf of Dr.Sai working group https://indico.cern.ch/event/1291157/contributions/5889603/

[6] Li, K., Liu, B., Mellado, B., Yuan, C.-Z., Zhang, Z.: AI agents, language, deep learning, and the next revolution in science. Front. Phys. (Beijing) 21(9), 096401 (2026) https://doi.org/10.15302/frontphys-2026-096401 arXiv:2603-794

[7] Han, S., et al.: LLM Multi-Agent Systems: Challenges and Open Problems (2025). https://arxiv.org/abs/2402.03578

[8] Xi, Z., et al.: The Rise and Potential of Large Language Model Based Agents: A Survey (2023). https://arxiv.org/abs/2309-07864

[9] Lewis, P., et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2021). https://arxiv.org/abs/2005-11401

[10] Wu, Q., et al.: AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023) https://arxiv. p. 2

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve tool-calling precision by developing specialized synthesis layers or fine-tuned models that possess a syntax level mastery of professional HEP tools to ensure agents generate schema compliant parameters for execution.

  2. Enhance domain knowledge representation by prioritizing how to effectively represent and embed a large amount of domain specific knowledge into the agentic architecture, transforming static information into executable logic that mirrors the strategic decision making of a human expert.

  3. Refine failure analysis by focusing on functional calling capability as the primary bottleneck across all models rather than just reasoning depth, which allows for targeted improvements in tool invocation logic for systems like DeepSeek-R1.

Sources

Related papers