MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks¶

Conference: ICML 2026
arXiv: 2601.14652
Code: https://github.com/SalesforceAIResearch/MAS-Orchestra (Available)
Area: LLM Agent / Multi-Agent Systems / Reinforcement Learning
Keywords: Multi-Agent Systems, Holistic Orchestration, Function Calling, GRPO, MASBench

TL;DR¶

This work reformulates "automated multi-agent system design" as a one-shot function-calling RL problem that outputs a complete MAS description. It also introduces MASBench to quantify "when MAS truly outperforms single agents" across five dimensions: Depth, Horizon, Breadth, Parallel, and Robustness.

Background & Motivation¶

Background: Automated multi-agent system (MAS) design has evolved from manual wiring (e.g., fixed topologies like Debate or CoT-SC) toward training-time orchestration. This involves an orchestrator LLM automatically generating sub-agent roles, connectivity, and execution sequences based on the task.

Limitations of Prior Work: The authors categorize existing approaches into three issues. First, at the formalization level, most works use "executable code" to describe orchestration (MAS-Zero, AFlow, W4S). The orchestrator must understand or even replicate the internal code of sub-agents, leading to high orchestration costs for complex sub-agents (e.g., multi-turn search agents) and forcing sub-agents to degrade into simple forms like CoT. Second, at the training level, methods either rely on unstable inference-time heuristic searches or use multi-step RL to assemble components incrementally, which suffers from poor long-range credit assignment and error accumulation. Third, the decision of "when to use MAS" remains purely empirical without a quantitative framework.

Key Challenge: Sequential multi-step orchestration forces the orchestrator to make locally optimal decisions at each step, which inherently conflicts with the global coordination benefits of MAS. Meanwhile, defining sub-agents at the level of "changing a prompt line" or "switching a backbone" ignores the critical dimensions of tools and workflows that truly distinguish sub-agent capabilities.

Goal: (1) Propose a MAS formalization that enables global reasoning for the orchestrator while accommodating complex sub-agents. (2) Provide a controlled benchmark to disentangle the effects of task structure, verification protocols, and orchestrator/sub-agent capabilities on MAS gains.

Key Insight: The authors observe that the orchestrator's essential capability is "high-level system design" rather than "replicating sub-agent internal behaviors." By abstracting sub-agents as black-box callable functions (exposing only signatures), the orchestrator can generate a complete system structure in one shot, bypassing the long-range credit assignment issues of sequential RL.

Core Idea: MAS is represented as a "one-shot function-calling program" using two primitives: create_agent and create_flow. The orchestrator is trained via GRPO on end-to-end task rewards to generate the entire system, with an explicit Degree of MAS (DoM) introduced as a user-controllable complexity knob.

Method¶

Overall Architecture¶

Given a dataset \(\mathcal{D}=\{(x_i,y_i)\}\) and a user-specified DoM level \(m\in\{\text{LOW},\text{HIGH}\}\), the orchestrator policy samples a complete orchestration \(a\sim\pi_\theta(\cdot\mid x,m)\). A deterministic rule parser \(f\) سپس translates \(a\) into an executable sub-agent calling graph, outputting the final prediction \(\hat{y}=f(x,a)\). The key difference from sequential approaches is that the orchestrator only observes task \(x\) at step 0 and makes no further incremental decisions; the quality of orchestration is backpropagated solely through the final answer's correctness.

Key Designs¶

Holistic Orchestration + Function-Calling Formalization:
- Function: Enables the orchestrator to output a complete MAS description, including multiple sub-agents and their connectivity topology, within a single forward pass.
- Mechanism: The orchestration space is constrained to two primitives: create_agent(role, goal, tools, workflow) to instantiate a goal-oriented sub-agent, and create_flow(from, to, payload) to describe information flow. Sub-agents are black-box functions where the orchestrator only sees signatures.
- Design Motivation: Sequential RL suffers from error accumulation and poor credit assignment. Holistic generation aligns RL signals directly with "system-level final returns," leading to stable training and allowing sub-agents to be arbitrarily complex (e.g., DeepResearch).
DoM (Degree of MAS) Explicit Constraints:
- Function: Users can specify \(m\) to enforce constraints on the orchestration space.
- Mechanism: LOW level allows at most one instantiated sub-agent and prohibits explicit inter-agent topologies; HIGH level imposes no constraints.
- Design Motivation: Empirical results show MAS does not benefit all tasks (e.g., sequential math problems). Leaving the "to MAS or not to MAS" decision to user/task priors saves significant ineffective orchestration overhead.
GRPO + Task-level Sparse Rewards:
- Function: Uses final answer correctness \(R(x,y,\hat{y})=\mathbb{1}[\hat{y}=y]\) as the sole signal for end-to-end MAS optimization.
- Mechanism: For each \(x\), \(K\) candidate orchestrations \(\{a_i\}_{i=1}^K\sim\pi_\theta(\cdot\mid x,m)\) are sampled. Group Relative Policy Optimization (GRPO, Shao et al. 2024) is used to construct a clipped policy gradient based on relative advantages within the group.
- Design Motivation: Holistic orchestration only receives feedback at the end. GRPO replaces value models with group comparisons, which naturally fits the "one prompt, multiple candidate systems" training paradigm.

Loss & Training¶

The training phase utilizes both controllable synthetic data from MASBench (generated by iGSM across Depth/Horizon/Breadth/Parallel/Robustness axes) and training sets from public benchmarks (DeepScaleR for AIME/GPQA, HotpotQA, and BrowseComp+). The sub-agent pool is fixed to five types: CoT, CoT-SC, Debate, Self-refine, and DeepResearch, all using the same LLM backend but differing in tools and workflows.

Key Experimental Results¶

Main Results¶

Using Qwen2.5-7B-Instruct as the orchestrator and GPT-OSS-120B (low) as the sub-agent backend, Ours is compared against standard standalone agents, SoTA inference-time orchestrators (AFlow, MaAS), and training-time orchestrators (ToolOrchestra):

Benchmark	Task Type	Best Standalone Agent	SoTA Orchestration Baseline	MAS-Orchestra	Notes
AIME24	Math (IID)	DebateAgent 62.08	AFlow 62.50	66.25	Low DoM
AIME25	Math (IID)	DebateAgent 57.50	AFlow 53.33	61.25	Low DoM
HotpotQA	Multi-hop QA	DeepResearch 46.44	ToolOrchestra 37.44	49.00	High DoM
BrowseComp+	Search QA	DeepResearch 8.56	ToolOrchestra 1.38	11.00	High DoM
GPQA	Reasoning (OOD)	DebateAgent 64.14	AFlow 65.43	65.21	Low DoM

In terms of efficiency, MAS-Orchestra lies on the Pareto frontier, achieving over 10× inference cost savings compared to strong baselines.

Ablation Study: MASBench Five-Axis Analysis¶

Configuration	Conclusion	Explanation
Sub-agent = Qwen-7B (Weak)	MAS beats SAS on Breadth/Parallel/Robustness; SAS beats MAS on Depth	Sequential CoT saves coordination overhead on strong dependency chains.
Sub-agent = GPT-120B (Strong)	MAS gain near zero on Depth/Horizon/Breadth/Parallel; MAS leads on Robustness	Strong sub-agents internalize structure; coordination costs outweigh benefits.
Orchestrator = RLM (e.g., GPT-OSS-20B)	Underperforms Instruction-tuned LLM	RLMs tend to solve tasks themselves instead of designing systems.
Robustness (Adversarial notes)	SAS accuracy near 0; MAS significantly leads	MAS actively adds final answer/moderator agents for cross-verification.

Key Findings¶

"Edge Capability" Hypothesis: Maximum MAS gains occur when sub-agents are "competent but not strong enough to internalize the entire structure." If too weak, sub-agents fail sub-tasks; if too strong, coordination costs negate benefits.
Holistic vs. Sequential: Ours outperforms ToolOrchestra across all benchmarks. On BrowseComp+, the jump from 1.38 to 11.00 demonstrates that sequential RL is particularly disadvantaged when sub-agents are complex.
Counter-intuitive Orchestrator Choice: Zero-shot GPT-5/Claude-4.5 as orchestrators are outperformed by a trained 7B Qwen, indicating that orchestration capability must be explicitly shaped via RL rather than emerging from general reasoning.
DoM Configuration Strategy: Tasks with strong sequential dependencies (Math/GPQA) favor LOW DoM, while tasks with parallel search (HotpotQA) favor HIGH DoM.

Highlights & Insights¶

"Function is the correct abstraction for MAS": By abstracting to signatures, the orchestrator is no longer hindered by sub-agent complexity. This allows complex agents like DeepResearch to be naturally integrated.
Falsifiable MAS Research: The five-axis decomposition of MASBench allows researchers to report exactly where MAS helps (e.g., "improves Robustness by \(X\%\)") rather than just aggregate numbers.
RLMs are not natural orchestrators: RLM's end-to-end training target makes it prefer direct task-solving over delegation. This serves as a warning for works attempting to use reasoning models like o1 as orchestrators without fine-tuning.

Limitations & Future Work¶

The sub-agent pool is still restricted to fixed workflows; the capability to "invent" entirely new sub-agent types was not verified.
Capability binding: There is an implicit coupling between the orchestrator and the sub-agent reasoning effort (context length management).
The Comparison with GPT-5/Claude-4.5 is somewhat unfair as they were not fine-tuned for this specific orchestration task.
MASBench relies heavily on synthetic iGSM data; lack of diverse use cases like code generation or long-document analysis.

vs. AFlow / MAS-Zero: They use inference-time heuristic search. MAS-Orchestra uses GRPO for explicit training, achieving 10× better efficiency and better performance via DoM parameterization.
vs. ToolOrchestra: Uses sequential RL. Sequential credit assignment is a significant bottleneck in MAS design for complex tasks.
vs. MAS-GPT: SFT-based orchestration. RL with task rewards better captures system-level coordination patterns and demonstrates more stable OOD generalization.

Rating¶

Novelty: ⭐⭐⭐⭐
Experimental Thoroughness: ⭐⭐⭐⭐⭐
Writing Quality: ⭐⭐⭐⭐
Value: ⭐⭐⭐⭐⭐