One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing

arXiv:2511.12484 · eess.SY, cs.AI, cs.SY · Submitted 2026-08-24 · 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: Next we'll be talking about the paper "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing".

Jane: The paper was written by Xu Yang, Chenhui Lin, Haotian Liu, Qi Wang, Yue Yang et al. from Tsinghua University, Beijing 100084, China. and Hong Kong Polytechnic University, Hong Kong. and Hefei University of Technology, Hefei 230009, China..

Tom: Stay tuned as we take you through the paper and discuss its implications.

Summary: Tom: We’ve established that the problem is complex and diverse, so let's talk about what "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing" actually does to solve it. Jane, can you summarize the core methodology for our listeners?

Jane: The paper introduces the ADN-Agent architecture, which is essentially a three-part system: a Planner, the suite of DSM experts, and a Summarizer. This framework is designed specifically to handle those multi-scenario operational requests that plague grid managers.

Tom: And it's not just about routing; it’s about intelligent decomposition—the Planner takes the user request and breaks it down into a series of manageable subtasks that need to be executed.

Lu: The mechanism they use for this is incredibly elegant because it means the LLM isn't trying to understand every single DSM's internal workings; it just needs to know what each tool can do and where.

Meng: That leads us straight into the communication mechanism, which is where I see the real engineering innovation—how we get those disparate tools to talk without requiring a lot of complex coding from the main LLM.

Lalam: The implication of this workflow is that AI has become a true conductor, managing a symphony of specialized functions to deliver one coherent result.

Tom: So, Jane, the Planner identifies the need for specific tools?

Jane: Yes. For example, if an operator asks about peak voltage in the Valley District on October 12th, the Planner knows that requires a data tool followed by a power flow simulation tool.

Lu: It’s about aligning user intent with functional capabilities of various components. The system is adapting to the request rather than forcing the request into a static model.

Meng: This adaptive routing means we don't need to predict every possible scenario and build one giant model for every single case; we just invoke the specific tools needed.

Lalam: It’s a shift from having a single "brilliant but limited" AI to having a system that manages human-level complexity.

Improvements: Tom: We have the structure and the summary, so now let's talk about what improvements this paper suggests in "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing." Jane, where does it improve on existing solutions?

Jane: It improves dramatically by addressing two major obstacles that typically hinder large-scale AI deployment: the interface heterogeneity of various tools and the reliance on domain expertise. The authors provide a practical solution to both problems.

Tom: And that's where I think the communication mechanism really shines, Jane, because it’s designed to overcome those interface differences. It’ provides a unified and flexible way for all these diverse tools to interact with the Planner at all times.

Lu: The key insight here is the Translator module within each DSM; it handles translating high-level natural language subtasks into executable commands, which is a huge reduction in cognitive load for other parts of the LLM.

Meng: From an implementation standpoint, that’s a massive gain in reliability. We're not asking the main LLM to guess how to run a simulation or optimize dispatch; we are sending it a simple command description and letting the specialized tool handle its own execution.

Lalam: This suggests that we can evolve the AI's capabilities by improving its communication protocols, not just by making one specific component more accurate.

Tom: So, Jane, it' about building robust scaffolding around the LLMs to make them better collaborators?

Jane: Exactly. It’s about making the system intelligent enough to coordinate the details and be adaptable enough for new technologies that haven't even been invented yet.

Lu: The paper is providing a blueprint for how complex systems should be designed to achieve functional excellence, which is a huge theoretical win.

Meng: And it allows us to integrate new DSM models seamlessly without having to overhaul the entire architecture, which really simplifies deployment tremendously.

Improvements (Continued): Tom: We've seen the communication mechanism; let's talk about another major improvement in "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing." Jane, what is the focus of the paper on automated training pipelines?

Jane: The authors recognized that some tasks—like generating an ADN model or consulting grid codes—are too language-intensive for general LLMs. So, they developed an automated pipeline to create specialized small language models, or FTSLMs.

Tom: This is where the fine-tuning comes in; instead of just hoping a massive model knows all its domain knowledge, we are creating smaller ones specifically trained on relevant data sets.

Lu: The way the paper automates this process—from generating prompts to verifying the output—is genius, because it means domain experts aren't stuck manually annotating thousands of samples.

Meng: From a practical standpoint, that' automation is crucial for scaling; we can generate one thousand six hundred eighty instruction-answer pairs for model adjustments and then have a small model learn from that volume of data in just days.

Lalam: The cultural impact here is the idea that knowledge transfer can be automated, allowing us to infuse complex domain expertise into AI in a repeatable, scalable manner.

Tom: So, Jane, we' are using these FTSLMs as dedicated experts within the "Multiple Experts" framework of the ADN-Agent?

Jane: Precisely. They act as specialized DSM extensions that handle those tricky language-intensive subtasks with much higher accuracy than a general LLM would be able to achieve.

Lu: It's not just about making a smaller model; it's about making a *better* expert for the specific task, which is much more sophisticated.

Meng: That specialized knowledge ensures that we are using the right tool for the job, whether it’s an optimization solver or a fine-tuned language model designed to adjust the grid model.

Conclusion: Tom: We've covered so much ground—the architecture, the workflow, and even how to improve specialized AI tools. Jane, let's wrap up this discussion on "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing." What is the big picture here?

Jane: It’s that we' are moving away from a singular, monolithic approach to a highly coordinated network of specialized AI tools. This allows us to handle the intense complexity of modern energy systems reliably.

Lu: I find it incredibly exciting because the potential for dynamic task routing suggests that the creative possibilities for designing and optimizing future power grids are genuinely limitless now.

Meng: From an implementation standpoint, this architecture allows us to scale up by just plugging in new domain-specific models without having to redesign the entire control structure, which is a massive win.

Lalam: It feels like this technology is going to fundamentally change how we view the relationship between human oversight and automated systems, making collaboration the standard.

Tom: You’re right, Lalam; it's about enabling a system where the operator just asks what they want, and the AI intelligently figures out which experts need to work together.

Lu: I can't wait to see how this concept applied to other highly technical domains beyond energy systems.

Meng: I'm already thinking about the software integration required to make this architecture run reliably at scale across different operators.

Lalam: It's a powerful vision for the future, and it’s a great foundation for our next topic of discussion.

Xu Yang, Chenhui Lin, Haotian Liu, Qi Wang, Yue Yang, Wenchuan Wu

Tsinghua University, Beijing 100084, China. · Hong Kong Polytechnic University, Hong Kong. · Hefei University of Technology, Hefei 230009, China.

eess.SY, cs.AI, cs.SY

Submitted: 2026-08-24

Updated: 2026-08-25

Comments: Accepted by CSEE Journal of Power and Energy Systems in June 2026

Code: https://github.com/langchain-ai/langchain

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

Importance score: 94/100

The gist: The paper, titled "One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing," addresses the challenges posed by the evolution of Active Distribution Networks

Key concepts

ADN-Agent Architecture
This three-part system—comprising a Planner, specialized DSM experts, and a Summarizer—is designed to handle multi-scenario requests. The Planner breaks down the initial user request into manageable subtasks for the system to execute.
Adaptive Task Routing
This capability allows the AI to adapt its workflow based on user intent rather than forcing a request into a static model. It identifies and invokes only the specific, necessary specialized tools needed for a given task.
Fine-Tuned Small Language Models (FTSLMs)
These are smaller models trained specifically on relevant data sets to handle language-intensive subtasks. They allow complex domain expertise to be infused into AI in a scalable way, improving accuracy over general LLMs.

Terminology

Summary

The paper, titled One Request, Multiple Experts: LLM Orchestrates Domain Specific Models via Adaptive Task Routing, addresses the challenges posed by the evolution of Active Distribution Networks (ADNs) into a complex multi-scenario, multiobjective problem. This complexity arises from two trends: the integration of massive distributed energy resources (DERs) and the reduction of entry barriers in electricity markets, which together make real-time management difficult.

The primary obstacles faced by ADN operators are identified as:

  1. Interface heterogeneity: DSM interfaces are inherently diverse, making it difficult to integrate sources from different departments.

  2. Domain expertise dependence: The complexity of DSM design requires multidisciplinary expertise, making it nearly impossible for ADN operators to fully grasp all of them within a limited time.

To address these issues, the authors propose the ADN-Agent architecture, which leverages a general large language model (LLM) to coordinate multiple Domain Specific Models (DSMs). The architecture consists of three key components:

  1. The Planner: A general LLM that performs intent recognition and analysis of the request, breaking it down into a series of manageable subtasks. It is responsible for assigning these tasks to the most suitable DSM based on the operation request.

  2. A suite of DSM-specific tools: These include specialized models such as a Data Tool, Model Tool, Simulation Tool, Optimization Tool, and Result Organization Tool.

  3. The Summarizer: A general LLM that aggregates and synthesizes corresponding results from the subtasks to deliver the final answer to the operator.

To manage the inherent heterogeneity of existing DSMs, a novel communication mechanism is introduced:

  • A dedicated workspace is created for each operation request to prevent cross-contamination of information.

  • Each DSM is augmented with a customized Translator module powered by general LLMs. This Translator module handles subtask retrieval, generates the necessary executable commands, and returns the results. This design ensures that each DSM only needs to expose its functionality to the Planner, significantly reducing the cognitive burden on the upper-level Planner.

Furthermore, for language-intensive subtasks, such as grid code consultation or ADN model generation, which exceed a general LLM's knowledge, the authors propose an automated training pipeline for fine-tuning small language models (FTSLMs). This pipeline involves:

  1. Generation prompt design: Domain experts craft prompts to define the target scenario and methodology.

  2. Data generation: A general LLM is used to generate instruction-answer pairs, supplemented by traditional data augmentation techniques like parameter perturbation.

  3. Data verification: A combination of a regular expression verifier, a rule-based verifier, and an LLM verifier is employed to filter out incorrect data.

  4. LoRA fine-tuning: The verified samples are used to train an FT-SLM using Low-Rank Adaptation (LoRA).

The effectiveness of the proposed method was validated through numerical studies comparing ADN-Agent against several baseline methods: Function-Call, Multi-LLM collaboration, No-Trans, No-FT, and Zero-Shot. The results showed that the ADN-Agent achieved the best performance, attaining a result accuracy of 95.8%.

Key findings from the comparison include:

  • The Function-Call paradigm suffers from parameter mis-specification, which leads to erroneous final outputs.

  • The Multi-LLM collaboration paradigm suffers a significant degradation in performance, resulting in low completion rates because it lacks an upper level Planner for coordinated orchestration.

In conclusion, the paper demonstrates that the ADN-Agent architecture provides a more rational and scalable architectural design for managing diverse and complex operational demands within Active Distribution Networks.

Improvements for AI systems

The references provided indicate a rapidly maturing field at the intersection of advanced Large Language Models (LLMs) and critical infrastructure engineering, specifically electrical power systems (e.g., Optimal Power Flow, Distribution System Analysis). Given that errors in this domain carry immense financial and safety risks, the focus must shift from mere capability to verifiable reliability and systemic robustness.

Based on the trends observed (Agentic frameworks [35], RAG [19], Multi-Agent systems [25, 36], and integrating linguistic intent into mathematical problems [29]), I propose three highly specific architectural improvements.

The Problem: Current LLM-based systems often treat physical constraints (e.g., Kirchhoff's laws, equipment ratings, voltage limits) as implicit knowledge or mere prompts. If the LLM hallucinates a mathematical formula or violates a physical boundary, the resulting power flow calculation is invalid and dangerous.

The Improvement: Implement an agentic architecture where the primary LLM acts not only as an interpreter but also as a Constraint Validator. This requires augmenting the standard Function Calling mechanism (as per [32]) with a dedicated, symbolic constraint layer.

What the Improved AI System Can Do:

  1. Intent-to-Constraint Mapping: Accepts high-level, natural language objectives (e.g., Minimize total line losses while ensuring no bus voltage drops below 0.95 p.u.). The VCAO translates this into a formal, verifiable set of mathematical constraints (Subject To).

  2. Pre-Solve Validation: Before calling the external numerical solver (e.g., an OPF solver), the VCAO passes the generated model parameters and constraints through a Symbolic Solver Module. This module checks for mathematical consistency, feasibility, and adherence to hard physical limits without running the full optimization.

  3. Iterative Refinement: If the symbolic check fails (e.g., a proposed line tap setting creates an inherent singularity), the VCAO does not pass the error to the user; instead, it uses its internal reasoning loop to diagnose which constraint was violated and proposes a minimal, corrected set of inputs for re-solving.

  4. Guaranteed Topology Encoding: Given a textual description of a network (Bus A connects to Bus B via Line L1, which has impedance Z AB ), the HyDSR module forces the LLM to output a graph structure (Nodes = Buses; Edges = Lines) with associated attributes (Impedance, Capacity). This output is machine-readable and cannot be ambiguous.

  5. Automated Matrix Population: The system then uses this guaranteed SIR to mathematically populate the Y bus matrix programmatically, bypassing the need for the LLM to write complex indexing logic. This ensures that every component of the resulting optimization problem is computationally sound and physically traceable back to its textual source.

  6. Domain-Specific Knowledge Graph Integration: The SIR module must be pre-populated with validated domain knowledge (e.g., standard transformer models, line charging constants) retrieved via RAG [19], ensuring that the LLM cannot hallucinate physical parameters.

  7. Conflict Resolution: When the optimization goal shifts (e.g., moving from a cost-minimization mode to an emergency stability restoration mode), the Supervisor Agent intercepts the proposed control action (u proposed). It evaluates u proposed against a prioritized hierarchy of safety policies (e.g., Never allow line overloading; Always maintain critical bus voltage >0.95).

  8. Policy-Guided Action Selection: If the proposed action violates a safety policy, the Supervisor Agent overrides it and delegates the task to a specialized Safety Policy Agent that is trained exclusively on stabilizing maneuvers, ensuring that the resulting optimal dispatch (u final) is not just mathematically optimal but also **operationally safe and

Sources

Related papers