OrchMAS: Orchestrated Reasoning with Multi Collaborative Heterogeneous Scientific Expert Structured Agents

Yichao Feng; Haoran Luo; Zhenghong Lin; Yiqun Sun; Pengfei Wei; Lawrence B. Hsieh; Anh Tuan Luu

cs.AIMar 3, 2026

OrchMAS: Orchestrated Reasoning with Multi Collaborative Heterogeneous Scientific Expert Structured Agents

Authors: Yichao Feng, Haoran Luo, Zhenghong Lin, Yiqun Sun, Pengfei Wei, Lawrence B. Hsieh, Anh Tuan Luu

Organizations: Magellan Technology Research Institute (MTRI) · Nanyang Technological University, Singapore

Abstract

Multi-agent large language model frameworks are promising for complex multi step reasoning, yet existing systems remain weak for scientific and knowledge intensive domains due to static prompts and agent roles, rigid workflows, and homogeneous model reliance, leading to poor domain adaptation, limited reasoning flexibility, and high latency on heterogeneous or long-horizon scientific tasks. They also struggle to revise earlier decisions when intermediate reasoning diverges, reducing reliability in structured and calculation heavy settings. To address these limitations, we propose a scientific domain oriented interactive two tier multi model orchestration framework. A dedicated orchestration model analyzes each task, dynamically constructs a domain aware reasoning pipeline, and instantiates specialized expert agents with tailored prompts, while an execution model performs each step under generated role and instruction specifications. The orchestrator iteratively updates the pipeline based on intermediate feedback, enabling dynamic replanning, role reallocation, and prompt refinement across multi turn interactions, strengthening robustness and specialization for scientific reasoning through structured heterogeneous model collaboration. The framework is model agnostic and supports heterogeneous LLM integration with different capacities or costs, enabling flexible performance efficiency trade offs in practical scientific deployments. Experiments show consistent improvements over existing multi agent systems and strong baselines across diverse reasoning and scientific style benchmarks.

Figures & tables

Figure 1. Comparison of multi step and MAS LLM generation versus basic prompting.

Figure 2. Structural comparison of representative MAS frameworks. Predefined MAS and Selective MAS rely on static role templates and predefined agent libraries, while Auto MAS learns coordination through SFT/RL with a trained agent policy. Interactive Auto MAS performs multi turn dynamic agent and prompt generation and feedback guided orchestration.

Figure 3. Training workflow of the OrchMAS framework. The orchestration policy iteratively performs agent role assignment, agent structure construction, and agent prompt generation to dynamically instantiate task specific agents. The process repeats until a finish action is triggered. Layer wise rewards are computed from intermediate trajectories and final answers.

Category	Emergent Role $r\in\mathcal{R}$	Typical Output	Stage	Action	State Update
Information Gathering	Researcher , Domain Expert	background facts, evidence	Early	tool_call $(r,p)$	$\mathcal{H}_{t}\leftarrow\mathcal{H}_{t-1}\oplus(p,y_{r})$
Planning	Planner , Strategist	strategy, subtask decomposition	Early	tool_call $(r,p)$	$\mathcal{H}_{t}\leftarrow\mathcal{H}_{t-1}\oplus(p,y_{r})$
Domain Solving	Math Solver , Coder , Psychologist	solution, code, domain analysis	Middle	tool_call $(r,p)$	$\mathcal{H}_{t}\leftarrow\mathcal{H}_{t-1}\oplus(p,y_{r})$
Verification	Verifier , Clarifier	validation, error identification	Late	tool_call $(r,p)$	$\mathcal{H}_{t}\leftarrow\mathcal{H}_{t-1}\oplus(p,y_{r})$
Critique	Critiquer	evaluation score, improved content	Late	tool_call $(r,p,\theta_{c})$	$\mathcal{H}_{t}\leftarrow\mathcal{H}_{t-1}\oplus(p,y_{r}^{\text{crit}})$
Synthesis	Assistant (default)	final answer $y_{q}$	Terminal	stop	$y_{q}=\pi_{\phi}(\mathcal{H}_{T},q)$

Table 1. Emergent role taxonomy and action space in Our framework. Roles are not predefined; the policy model learns to generate Agent Roles strings via RL. We categorize observed emergent roles by their functional stage in multi turn reasoning.

(a) Policy Model Prompt

\pi_{\phi}

— Interaction Protocol

Question

First, provide a simple explanation of the question and give it to the large language model for a more accurate answer. Focus on explaining the question without deep reasoning in the first step. After receiving the response, think about the large language model’s response, and by interacting with the large language model again and again, arrive at the final answer. Proceed step by step with the following rules:

1. Only in the first step, provide a brief explanation of the question and give it to the large language model:

<think> (Brief thinking must not be over 80 words) </think>

<interaction_prompt> {"name": "prompt_dynamic", "arguments": {"prompt": "...", "agent_role": "{role}" }} </interaction_prompt>

2. After the first step, in each interaction with the large language model, write:

Table 2. Our two level prompt architecture. (a) The policy model

\pi_{\phi}

follows a multi turn interaction protocol: think

\to

interact (with agent_role )

\to

observe

\to

\cdots

\to

answer . (b) The external LLM

\mathcal{M}_{\text{ext}}

receives a dynamically constructed system prompt where agent_role , learned entirely through RL, determines its persona and expertise.

Methods		2Wiki		Hotpot		GSM8K		DAPO		PopQA		MusiQue
		F1	EM	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM
Baseline	Qwen3-4B	28.21	26.04	23.21	21.88	84.38	84.38	0.00	0.00	11.02	7.29	6.71	3.13
Baseline	GPT-4o-mini	35.44	33.33	41.21	32.29	86.61	83.33	19.79	19.79	32.30	25.00	23.59	14.58
SFT	Qwen3-4B	38.91	41.67	33.09	23.96	32.29	32.29	5.21	5.21	8.89	9.38	18.61	8.33
CoT Reasoning	Qwen3-4B	25.21	20.83	24.98	18.75	81.25	81.25	0.00	0.00	10.97	7.29	8.72	4.17
CoT Reasoning	GPT-4o-mini	50.89	43.75	46.67	40.63	87.02	84.38	22.92	22.92	33.92	28.13	26.33	17.71

Table 3. performance comparison across six QA benchmarks (2Wiki, HotpotQA, GSM8K, DAPO, MusiQue, and PopQA) measured by F1 and Exact Match (EM). We compare baseline models, supervised fine tuning (SFT), CoT prompting, GRPO training, and representative MAS optimization methods (OPRO, TextGrad, GEPA) against our proposed OrchMAS framework. Our method consistently achieves the best results across all datasets, with substantial gains over the strongest non OrchMAS baseline.

Methods		In Distribution		OOD Tasks
		BookSum	W.P.	XSum	TriviaQA		MathQA		SQuAD v2
		Cos	Cos	Cos	F1	EM	F1	EM	F1	EM
Baseline	Qwen3-4B	42.32	21.55	32.64	46.54	44.79	31.86	29.17	16.22	7.29
Baseline	GPT-4o-mini	57.43	33.37	61.33	76.33	63.54	55.29	46.88	25.41	13.54
SFT	Qwen3-4B	52.87	29.68	54.61	32.82	31.25	23.74	17.71	16.14	5.21
CoT Reasoning	Qwen3-4B	35.32	14.13	6.37	51.77	46.88	32.13	27.08	16.98	6.25

Table 4. Performance comparison on in distribution and OOD tasks across summarization and question answering benchmarks. In distribution results are reported on BookSum and WritingPrompts (W.P.), while OOD generalization is evaluated on XSum, TriviaQA, MathQA, and SQuAD v2.

Figure 4. Result comparison for easy and hard questions

Dataset		2Wiki		Hotpot		GSM8K		DAPO		MusiQue		PopQA
Dataset		F1	EM	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM
OrchMAS		67.25	60.42	61.99	53.13	97.92	97.92	56.64	56.64	33.33	21.88	39.87	34.38
	- Agent Roles	50.87	50.00	32.55	25.00	86.67	86.46	16.67	16.67	11.24	4.17	13.50	12.50
	- Multi Turns	38.43	37.5	53.30	44.79	78.13	78.13	42.71	42.71	9.50	5.21	39.39	35.42
	- Environment	45.00	43.75	36.36	28.13	96.88	96.88	16.67	16.67	13.59	6.25	18.51	15.63
Dataset		BookSum		W.P.		Xsum		Squad v2		TriviaQA		MathQA

Table 5. Ablation study of OrchMAS on diverse QA, reasoning, and summarization benchmarks. Results are reported using F1/EM for QA style tasks and Cos for summarization tasks. We compare the full OrchMAS framework with variants that remove one key module at a time: dynamic agent roles, multi turn reasoning, and environment guided execution. Removing any module leads to substantial and systematic performance degradation across datasets, demonstrating that each component plays a critical role and that their combination provides complementary gains in robustness, reasoning quality, and cross task generalization.

Figure 5. Result comparison for easy and hard questions

Figure 6. Training time orchestration statistics.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Turn	Role $\sigma_{t}$	Reasoning $u_{t}^{\mathrm{reason}}$	Request $u_{t}^{\mathrm{msg}}$	Feedback $v_{t}^{\mathrm{msg}}$
Question: Do both Beauty and Bullets and In the Name of the King 3: The Last Mission films have the
directors from the same country?
Turn 1	researcher	The question asks whether the directors of “Beauty and Bullets” and “In the Name of the King 3: The Last Mission” are from the same country. I need to identify each film’s director and their nationality.	Do both Beauty and Bullets and In the Name of the King 3: The Last Mission films have the directors from the same country?	Beauty and Bullets : directed by Derrick Ku (China). In the Name of the King 3 : directed by Uwe Boll (Germany). China $\neq$ Germany, so the directors do not share the same country.
Turn 2	an assistant	The response confirms that the two directors are from different countries (China vs. Germany). The answer is definitively “No”. No further verification needed.	Answer: No

Appendix

Table 6. Case study on 2WikiMultihopQA (Data ID: 3). The coordinator

\mathcal{C}

dynamically assigns agent roles

\sigma_{t}

to resolve a nationality comparison question. Ground truth: No .

Turn	Role $\sigma_{t}$	Reasoning $u_{t}^{\mathrm{reason}}$	Request $u_{t}^{\mathrm{msg}}$	Feedback $v_{t}^{\mathrm{msg}}$
Case 1: Where was the director of film A Zed & Two Noughts born? Ground truth: Newport, Wales
1	researcher	The question asks about the birthplace of the director of “A Zed & Two Noughts.” I need to identify the director first.	Who was the director of the film “A Zed & Two Noughts”?	The film was directed by Peter Greenaway .
2	clarifier	The director is Peter Greenaway. Now I need to find out where he was born.	Where was Peter Greenaway born?	Peter Greenaway was born in Newport, Monmouthshire, Wales, UK (5 April 1942).
3	an assistant	Peter Greenaway was born in Newport, Monmouthshire, Wales. The answer is confirmed.	Answer: Newport, Monmouthshire, Wales ✓
Case 2: What nationality is the performer of song You’re My One And Only Love? Ground truth: American
1	researcher	The question asks the nationality of the performer. I suspect Celine Dion (Canadian), but need to verify the actual performer.	What nationality is the performer of song “You’re My One And Only Love”?	The performer is American .

Appendix

Table 7. Case studies on 2WikiMultihopQA. The coordinator

\mathcal{C}

dynamically assigns agent roles

\sigma_{t}

at each turn. Both cases demonstrate the three-phase collaboration: researcher

\rightarrow

clarifier

\rightarrow

an assistant .

Multi-Agent LLMs

Multi-agent large language models coordinate multiple AI agents to solve tasks, share information, and make collective decisions.

See all 315 papers

Jun 14, 2026

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Jingru Guo, Xiangyuan Xue, Lian Zhang +6

1Imperial College London · 2The Chinese University of Hong Kong · University of Illinois Urbana-Champaign +4

LLM Agent Orchestration Multimodal Reasoning Scientific Reasoning in Language Models Large Language Model-Guided Scientific Discovery RL for LLM Agents

Ask PDF

Sep 25, 2025

Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG

Large language models (LLMs) have recently shown strong progress on scientific reasoning, yet two major bottlenecks remain. First, explicit retrieval fragments reasoning, imposing a hidden "tool tax" of extra tokens and steps. Second, multi-agent pipelines often dilute strong solutions by averaging across all candidates. We address these challenges with a unified framework that combines implicit retrieval and structured collaboration. At its foundation, a Monitor-based retrieval module operates at the token level, integrating external knowledge with minimal disruption to reasoning. On top of this substrate, Hierarchical Solution Refinement (HSR) iteratively designates each candidate as an anchor to be repaired by its peers, while Quality-Aware Iterative Reasoning (QAIR) adapts refinement to solution quality. On Humanity's Last Exam (HLE) Bio/Chem Gold, our framework achieves 48.3% accuracy - the highest reported to date, surpassing the strongest agent baseline by 13.4 points and leading frontier LLMs by up to 18.1 points, while simultaneously reducing token usage by 53.5% and agent steps by 43.7%. Results on SuperGPQA and TRQA confirm robustness across domains. Error analysis shows that reasoning failures and knowledge gaps co-occur in over 85% of cases, while diversity analysis reveals a clear dichotomy: retrieval tasks benefit from solution variety, whereas reasoning tasks favor consensus. Together, these findings demonstrate how implicit augmentation and structured refinement overcome the inefficiencies of explicit tool use and uniform aggregation. The code is available at https://github.com/tangxiangru/Eigen-1.

Xiangru Tang, Wanghan Xu, Yujie Wang +13

Yale University · Shanghai Jiao Tong University · Fudan University +4

Scientific Reasoning in Language Models Multi-Agent Reasoning Retrieval-Augmented Generation Iterative Refinement

Ask PDF

Apr 30, 2026

Heterogeneous Scientific Foundation Model Collaboration

Agentic large language model systems have demonstrated strong capabilities. However, their reliance on language as the universal interface fundamentally limits their applicability to many real-world problems, especially in scientific domains where domain-specific foundation models have been developed to address specialized tasks beyond natural language. In this work, we introduce Eywa, a heterogeneous agentic framework designed to extend language-centric systems to a broader class of scientific foundation models. The key idea of Eywa is to augment domain-specific foundation models with a language-model-based reasoning interface, enabling language models to guide inference over non-linguistic data modalities. This design allows predictive foundation models, which are typically optimized for specialized data and tasks, to participate in higher-level reasoning and decision-making processes within agentic systems. Eywa can serve as a drop-in replacement for a single-agent pipeline (EywaAgent) or be integrated into existing multi-agent systems by replacing traditional agents with specialized agents (EywaMAS). We further investigate a planning-based orchestration framework in which a planner dynamically coordinates traditional agents and Eywa agents to solve complex tasks across heterogeneous data modalities (EywaOrchestra). We evaluate Eywa across a diverse set of scientific domains spanning physical, life, and social sciences. Experimental results demonstrate that Eywa improves performance on tasks involving structured and domain-specific data, while reducing reliance on language-based reasoning through effective collaboration with specialized foundation models.

Zihao Li, Jiaru Zou, Feihao Fang +6

University of Illinois Urbana-Champaign

AI for Science LLM Agent Orchestration Multi-Agent Collaboration Multi-Agent Systems

Ask PDF

Abstract

Multi-Agent LLMs

Explore similar work

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG

Heterogeneous Scientific Foundation Model Collaboration