ARM: Role-Conditioned Neuron Transplantation for Training-Free Generalist LLM Agent Merging

TL;DR

ARM tackles the challenge of cross-environment robustness in interactive LLM agents by providing a training-free merging workflow. It constructs a pool of merged backbones, selects a strong initialization using Activation-Overlap Score based on role-conditioned activations, and then performs conflict-aware, localized neuron transplantation to repair remaining gaps. The method achieves superior generalist performance and improved worst-case robustness across multiple Qwen-based expert pools and benchmarks, while maintaining efficient, gradient-free operation. This approach enables practical deployment of generalist agents without expensive retraining or extensive task coverage, by explicitly preserving and repairing role-critical circuits.

Abstract

Interactive large language model agents have advanced rapidly, but most remain specialized to a single environment and fail to adapt robustly to other environments. Model merging offers a training-free alternative by integrating multiple experts into a single model. In this paper, we propose Agent-Role Merging (ARM), an activation-guided, role-conditioned neuron transplantation method for model merging in LLM agents. ARM improves existing merging methods from static natural language tasks to multi-turn agent scenarios, and over the generalization ability across various interactive environments. This is achieved with a well designed 3-step framework: 1) constructing merged backbones, 2) selection based on its role-conditioned activation analysis, and 3) neuron transplantation for fine-grained refinements. Without gradient-based optimization, ARM improves cross-benchmark generalization while enjoying efficiency. Across diverse domains, the model obtained via ARM merging outperforms prior model merging methods and domain-specific expert models, while demonstrating strong out-of-domain generalization.

Figures (12)

• Figure 1: Performance variability of common training-free merge heuristics across interactive agent benchmarks.
• Figure 2: Overview of Agent-Role Merging (ARM).Step 1: Backbone pool construction. We apply multiple training-free weight-space merge operators to benchmark-specialized experts to obtain a pool of candidate merged backbones. Step 2: Backbone selection. A selector computes the Activation-Overlap Score (AOS) using role-conditioned MLP activations on a lightweight calibration set, and chooses the candidate backbone that maximizes mean AOS across benchmarks. Step 3: Neuron transplantation. For benchmarks where the selected backbone remains weak, we transplant a small top-$k\%$ subset of donor (expert) MLP neurons into the backbone while strictly protecting neurons salient for other benchmarks to avoid negative transfer. The resulting single model consolidates expert capabilities across benchmarks without end-to-end retraining.
• Figure 3: AOS correlates positively with overall performance (Avg) across candidate merge backbones on Qwen3-8B, enabling lightweight initialization selection.
• Figure 4: Role-conditioned tracing reduces cross-benchmark overlap of salient neurons. We visualize top-$10\%$ salient neurons and highlight neurons shared across benchmark-specific sets. Compared to full-response tracing, role-conditioned tracing yields lower overlap, indicating reduced cross-environment entanglement of the traced circuits.
• Figure 5: Top-$k$ sensitivity analysis. Conflict-aware protection remains stronger as $k$ increases, indicating improved robustness to the saliency threshold.