Multipolar Task Processing Graph

• Multipolar Task Processing Graph is a directed acyclic graph that defines hierarchical task decomposition and convergent result fusion in multi-agent systems.
• It enables dynamic task planning by integrating IF-THEN rules to manage domain-specific agent behaviors and mitigate uncertainty effectively.
• XAgents leverages MTPG to surpass state-of-the-art methods in question-answering tasks, demonstrating improved efficiency and adaptable coordination.

Searching arXiv for the specified papers to ground the article in the cited sources. A multipolar task processing graph (MTPG) is a directed acyclic graph introduced within XAgents, a unified multi-agent cooperative framework built on a multipolar task processing graph and IF-THEN rules for complex tasks with uncertainty. In this formulation, task planning is organized as a decomposition phase from an original task into subtasks and a fusion phase from subtask outputs into a final result. XAgents uses the multipolar task processing graph to enable dynamic task planning and handle task uncertainty, while domain-specific IF-THEN rules constrain agent behaviors during subtask processing and global rules enhance inter-agent collaboration. Across three distinct datasets, XAgents is reported to consistently surpass state-of-the-art single-agent and multi-agent approaches in both knowledge-typed and logic-typed question-answering tasks (Yang et al., 12 Sep 2025).

1. Formal definition and graph structure

The formal definition of MTPG is a directed acyclic graph

$G=(V,E),$

with

$$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$

Here, $T$ is the original (root) task node, $T_i$ are subtask nodes, and $F$ is the fusion (leaf) node. The edge set is unweighted and dependency-only. Because edges carry no additional weights or labels, one may take $P=\varnothing$; all information about uncertainty and domain labels is carried in the IF-THEN rules, not in $E$ itself (Yang et al., 12 Sep 2025).

This definition fixes an important boundary around the construct. MTPG is not a weighted planning graph, not a probabilistic graphical model, and not a generic agent-interaction topology. Its graph-theoretic role is narrower: it expresses decomposition dependencies from the root task to subtasks and convergence dependencies from subtasks to the fusion node. Uncertainty is externalized into rule processing rather than encoded in edges.

The graph is explicitly a DAG, and this property is used to justify two theoretical claims in XAgents: the absence of cyclic dependencies and guaranteed eventual termination. The paper also states that dynamic reprocessing and path reconstruction provide a form of completeness: if a subtask remains misaligned, it is either refined or replaced. Within the framework, these guarantees are therefore attached to the combination of DAG structure and rule-driven update procedures, not to the static graph alone.

2. “Multipolar” organization: divergence and convergence

MTPG is described as being inspired by two biological “poles” of information flow. The first is SIMO (Single-Input, Multiple-Output), where a single parent node diverges into multiple children; in MTPG this is

$T\rightarrow\{T_1,\dots,T_n\}.$

The second is MISO (Multiple-Input, Single-Output), where multiple children converge to a single parent; in MTPG this is

$\{T_1,\dots,T_n\}\rightarrow F.$

Because MTPG composes divergence and convergence in a hierarchical, possibly multi-layer fashion, it is called “multipolar” (Yang et al., 12 Sep 2025).

The corresponding functional view is given by

$$\text{decompose}:T\mapsto\{T_{1},\dots,T_{n}\},\qquad \text{fuse}:\{T_{1},\dots,T_{n}\}\mapsto F.$$

This decomposition–fusion pairing is the core architectural principle. It separates the generation of subproblems from the synthesis of their outcomes, which allows the framework to interpose domain-specific rule systems and global-goal checks between the two phases.

A common misunderstanding is to treat “multipolar” as a synonym for arbitrary branching or for a fully general multi-agent graph. In the XAgents formulation, the term is more specific. It denotes the composition of SIMO-style divergence and MISO-style convergence in a hierarchical DAG. The graph can be multi-layer, but the essential idea is the alternation between branching task decomposition and convergent result fusion.

3. Processing pipeline within XAgents

The high-level procedural outline begins with an uncertain task $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$0 as input. The Planner Agent constructs the initial MTPG:

$$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$1

For each subtask node $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$2, the Domain Analyst Agent generates domain rules, the Domain Expert Agents produce outputs under those rules, and the Fusion Expert Agent performs subtask fusion. A global-goal check then evaluates the fused subtask outcome against a threshold, and the process repeats until all subtasks meet the global-goal threshold. The final output is produced by the final fusion stage:

$$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$3

The named processing units are PA(·), DAA(·), DEA(·), and FEA(·), corresponding respectively to Planner Agent, Domain Analyst Agent, Domain Expert Agents, and Fusion Expert Agent (Yang et al., 12 Sep 2025).

The operational logic is therefore not a one-pass decomposition tree. It is an iterative graph-guided workflow in which subtask handling, rule regeneration, threshold-based validation, and graph reconstruction are all first-class operations. The distinction matters because the MTPG is not merely a static plan representation; it is the substrate for dynamic task planning under uncertainty.

At the subtask level, IF-THEN rules play two roles. Domain-specific rules constrain the local behavior of agents assigned to a subtask. Global rules evaluate whether the local outcome remains aligned with the overall task objective. This separation between local rule conditioning and global-goal checking is the mechanism by which XAgents couples specialized subtask reasoning to system-level coordination.

4. Uncertainty handling and adaptive reprocessing

The framework identifies four mechanisms for handling uncertainty and ambiguity. First, multi-input balancing allows each subtask to be solved by $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$4 domain rules in parallel and then fused, thereby mitigating single-path failure under uncertainty. Second, each IF-part computes a discrete membership in $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$5. Third, if the fused subtask outcome $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$6 fails the global-goal membership threshold (ML), domain rules are regenerated with feedback:

$$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$7

Fourth, when $$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$8 rules disagree, FEA uses (i) a majority vote and (ii) the higher membership degree to choose the final semantic (Yang et al., 12 Sep 2025).

These mechanisms jointly define how uncertainty is localized, measured, and resolved. Multi-input balancing introduces parallel interpretive paths; membership-degree labels discretize confidence or alignment; adaptive reprocessing revises the rule basis when alignment is insufficient; and semantic conflict resolution specifies how disagreement among rules is collapsed into a single subtask semantic.

The framework also includes autonomous path reconstruction. If repeated reprocessing fails, the Planner Agent generates a new subgraph for the problematic subtask,

$$V=\{T,T_{1},\dots,T_{n},F\},\quad E=\{(T,T_{i})\}_{i=1}^{n}\cup\{(T_{i},F)\}_{i=1}^{n}.$$9

This is more than local rule repair: it alters the processing graph itself. A plausible implication is that XAgents treats task uncertainty as potentially structural rather than merely parametric, since failure can trigger graph expansion instead of only rule substitution.

5. Illustrative case and empirical role

An illustrative case study is the “Reply to Editor” task. In that example,

$T$0

One subtask is specified as $T$1 “Which film gave Hepburn her 2nd Oscar?” Another subtask, $T$2, is decomposed further into $T$3 when its global membership remained $T$4. For $T$5, DAA generates three domain rules (EM, History, Biology). Their membership degrees are H, M, ML. After three feedback loops, $T$6 attains High alignment with the global goal. T3’s complexity triggers autonomous path reconstruction:

$T$7

Finally, all $T$8 are fused at $T$9 to form the completed email reply (Yang et al., 12 Sep 2025).

This case illustrates several defining properties at once: the basic root–subtask–fusion topology, the use of heterogeneous domain rules for a single subtask, threshold-triggered feedback loops, and graph growth through path reconstruction. It also shows that the framework’s “multipolar” character is not limited to a single decomposition layer; the graph can deepen when an initially generated subtask remains insufficiently aligned.

Empirically, the paper reports that XAgents scales better in memory and token use than comparable multi-agent methods, with runtime and memory results on CC tasks. In conjunction with the question-answering results across three datasets, the reported performance profile situates MTPG not only as a planning abstraction but also as a coordination structure intended to improve execution efficiency under complex and uncertain task conditions.

6. Relation to graph-theoretic multitasking and open issues

The term “multipolar” also appears in a different graph-theoretic setting in “A Graph-Theoretic Approach to Multitasking” (Alon et al., 2016). There, a “multipolar” task-processing network is modeled as a $T_i$0-partite graph

$T_i$1

where edges only connect vertices in different parts. In a layered depth-$T_i$2 network, tasks correspond to node-disjoint paths $T_i$3 with $T_i$4. The paper defines multitasking capacity through induced matchings and proves, for depth-$T_i$5, $T_i$6-regular networks of width $T_i$7, that

$T_i$8

so deeper networks necessarily incur strictly worse multitasking capacity in terms of $T_i$9 (Alon et al., 2016).

This formalization is distinct from the MTPG used in XAgents. The multitasking paper studies interference among simultaneously executed tasks in $F$0-partite or layered graphs; XAgents defines a DAG with a root task, subtask nodes, and a fusion node, while placing uncertainty and domain semantics in IF-THEN rules rather than in the graph edges. Even so, the two uses of “multipolar” are conceptually adjacent in that both treat task processing through structured graph decomposition and recombination. This suggests that the XAgents objective of optimizing multipolar graph topologies for more efficient coordination can be read alongside broader results on how depth, degree, and interference affect multitasking capacity.

The open challenges stated for XAgents are threefold: enhanced rule-based explainability, further mitigation of LLM hallucinations via richer rule structures, and optimization of multipolar graph topologies for more efficient coordination (Yang et al., 12 Sep 2025). An additional limitation identified in the description is that MTPG’s reliance on automatic rule generation may introduce overhead and domain dependence that require refinement. These issues delimit the present