Agent-Role Merging (ARM) Techniques

• Agent-Role Merging (ARM) is a framework that dynamically integrates specialized agent roles to form adaptable generalist agents.
• It employs clustering, neuron transplantation, and masking techniques to mitigate destructive interference and facilitate efficient role transfer.
• Empirical results demonstrate ARM's robust performance in multi-agent, LLM, and VLA domains with enhanced scalability and task alignment.

Agent-Role Merging (ARM) refers to a class of methodologies and frameworks that enable the integration, composition, or dynamic adjustment of agent roles—either in multi-agent learning systems, collaborative LLM-based agents, or via training-free model merging protocols. ARM addresses both the theoretical and practical challenges of combining distinct agent capabilities, role specializations, or expert modules into unified or rapidly adaptive generalist agents without destructive interference or retraining.

1. Conceptual Foundations and Motivations

The ARM paradigm seeks to facilitate scalable multi-agent intelligence and robust generalization by strategically combining or adapting agent roles. In multi-agent systems (MAS), this entails decomposing complex joint action spaces into modular subspaces aligned with functional roles, then enabling agents to traverse, merge, or split these roles as task demands shift. For LLM- and vision-language-action (VLA) agents, ARM methods aim to merge expert models—each trained for a specific benchmark, capability, or embodiment—into a single generalist agent that retains the distinctive skills of all parents while minimizing cross-role interference (Feng et al., 12 Jan 2026, Fu et al., 24 Nov 2025).

The main motivators for ARM include:

• Alleviation of catastrophic forgetting and destructive interference in model merging.
• Efficient transfer to novel domains or enlarged agent populations.
• Decentralized and metric-guided role evolution for resilience and adaptability.
• Avoidance of expensive multi-task joint training when integrating modular expertise.

2. Role Discovery and Clustering Strategies

Role discovery is a foundational step in many ARM systems. In the RODE framework, primitive actions are embedded by their measurable effects—namely environmental transitions and rewards—and then clustered via k-means to define disjoint role action spaces. Specifically, each action $a \in A$ is mapped to a vector $\phi_a = f_e(a; \theta_e)$ using a forward model. The clustering process partitions the set of all action embeddings, yielding sets $A_j$ corresponding to roles $\rho_j$:

• The action effect representation is optimized with a joint loss $L_e$ combining observation-prediction and reward-prediction objectives.
• Fixed clusters $C_j$ and corresponding encoders facilitate transferability and modular learning (Wang et al., 2020).

In frameworks such as MorphAgent (Lu et al., 2024), roles emerge as “agent profiles” in natural language, with embeddings $R_i := e(p_i)$ evolving through LLM-based updating guided by metric scores—without explicit cluster definitions but producing comparable role separation and merger effects.

3. Model Merging, Neuron Transplantation, and Masking Approaches

ARM in the model-merging context is instantiated by procedures that merge the parameters or internal structures of expert agents, commonly using training-free operations. The ARM protocol in (Feng et al., 12 Jan 2026) is structured in three phases:

1. Backbone Construction: Multiple merged candidate models are computed via operators such as uniform averaging, task arithmetic, or TIES.
2. Role-Conditioned Activation Analysis: Role-salient neurons are identified per benchmark and role by computing absolute-mean activations. The Activation-Overlap Score (AOS) is used to select the backbone best preserving expert circuits.
3. Conflict-Aware Neuron Transplantation: For “weak” benchmarks, donor-specific role-salient neurons not overlapping with any other benchmark are transplanted into the merged backbone, ensuring minimal destructive interference.

In the vision-language-action domain, MergeVLA (Fu et al., 24 Nov 2025) addresses merge conflicts by:

• Masking LoRA adapter weights with per-task binary masks $S_m$ to retain only non-conflicting, task-significant parameters.
• Redesigning action experts to exclude self-attention, which empirically propagates task-specific signals and amplifies incompatibility.

The table below summarizes key ARM mechanisms across representative works.

Framework	Role Representation	Merging Mechanism
RODE	Action embeddings	Action-effect clustering, hierarchy
MorphAgent	NL profile embeddings	Decentralized, metric-guided profile evolution
ARM (LLM)	Transformer neurons	Backbone merging + neuron transplantation
MergeVLA	LoRA, action blocks	Task-masked LoRA + cross-attention expert

4. Dynamic Role Evolution and Metric-Guided Adaptation

MorphAgent (Lu et al., 2024) exemplifies decentralized, continuous ARM by representing agent roles as mutable natural-language profiles $p_i$. These profiles evolve through iterative prompting and revision, guided by three scalar optimization metrics:

• Role Clarity Score (RCS): Measures profile specificity and skill definition.
• Role Differentiation Score (RDS): Encourages separation in embedding space among agent roles.
• Task-Role Alignment Score (TRAS): Quantifies alignment of team profiles with the current task’s semantic and capability requirements.

Role merging and splitting are induced algorithmically when metric feedback prompts agents to specialize or blend sub-roles, implemented as semantic profile updates via LLM calls rather than explicit module recombination. This process supports continual adaptation and robustness in highly variable environments.

5. Bi-Level Control and Hierarchical Architectures

Hierarchical ARM architectures, as realized in RODE (Wang et al., 2020), split control into a high-level role selector—assigning agents to roles at coarse temporal resolutions—and low-level policies—enforcing action restrictions dictated by selected roles. The role selector operates with role embeddings defined by mean action vectors of clusters, and decomposed Q-value estimation for both selection (global $Q_{\text{tot}}^{\beta}$) and policy execution (global $Q_{\text{tot}}$). Both selector and policy levels utilize direct embedding inner products to score choices, bridging abstraction levels while maintaining sample efficiency and modularity.

This decomposition has demonstrated rapid policy transfer to domains with increased agent populations and altered action spaces via simple re-embedding and role expansion procedures.

6. Empirical Performance and Limitations

ARM protocols yield robust performance across diverse empirical settings:

• RODE: Outperforms state-of-the-art MARL algorithms on 10/14 StarCraft II SMAC scenarios, achieving top scores on all hard and super-hard maps; supports zero-shot transfer with tripled agent counts (Wang et al., 2020).
• MorphAgent: Achieves statistically significant improvements in code generation and mathematical reasoning benchmarks relative to static and centralized baselines, and maintains resilience under node failure (Lu et al., 2024).
• ARM (LLM merging): Surpasses the best-of-oracle expert selector for both in-domain and out-of-domain benchmarks, improves worst-case reliability, and does so without gradient-based retraining. Edits are localized to role-salient neurons (2–3% of MLP neurons per weak suite) (Feng et al., 12 Jan 2026).
• MergeVLA: Restores cross-task mergeability in VLA agents, achieving approximately 90% of single-task accuracy across robot manipulation, multi-embodiment, and mixed-task challenges. Failures of naïve model merging are circumvented by task masking and architectural specialization (Fu et al., 24 Nov 2025).

ARM is subject to practical limitations: dependency on architectural homogeneity and explicit checkpoint access (LLM merging), interpretability of activation traces, and the necessity for carefully designed metric or mask parameter regimes. Theoretical analysis of neuron-level interference and ensemble dynamics remains an open direction.

7. Prospects and Future Directions

Emerging ARM methodologies point to several future advances:

• Generalization to cross-family, parameter-efficient, or mixed-modality agent families (Feng et al., 12 Jan 2026, Fu et al., 24 Nov 2025).
• Automatic, unsupervised identification of roles or sub-tasks via improved role-span parsing and semantic analysis.
• Enhanced modularity through adaptive profile segmentation and dynamic sub-role recombination in large agent societies (Lu et al., 2024).
• Broader adoption in real-world collaborative robotics, language-guided planning, and adaptive decision-support systems.

The ARM paradigm unifies cluster-based, metric-guided, and neuron-level approaches to agent integration and role modularity, supporting scalable, adaptive, and highly generalist agent design.