AgentCPM-Explore: Realizing Long-Horizon Deep Exploration for Edge-Scale Agents

TL;DR

This work systematically investigates training agentic models at the 4B scale, identifying catastrophic forgetting, reward-noise sensitivity, and long-context contamination as key bottlenecks. It introduces AgentCPM-Explore, a three-stage framework combining parameter-space model merging, reward signal denoising, and context information refinement to enable long-horizon deep exploration in edge-scale agents. Empirical results show 4B agents achieving SOTA performance among peers and rivaling larger models on multiple benchmarks, with GAIA pass@64 reaching 97.09% under extended inference. The study demonstrates that with a carefully designed training framework, edge-scale models can realize substantial problem-solving capabilities previously attributed mainly to larger models, offering practical impact for privacy-preserving, low-resource intelligent agents.

Abstract

While Large Language Model (LLM)-based agents have shown remarkable potential for solving complex tasks, existing systems remain heavily reliant on large-scale models, leaving the capabilities of edge-scale models largely underexplored. In this paper, we present the first systematic study on training agentic models at the 4B-parameter scale. We identify three primary bottlenecks hindering the performance of edge-scale models: catastrophic forgetting during Supervised Fine-Tuning (SFT), sensitivity to reward signal noise during Reinforcement Learning (RL), and reasoning degradation caused by redundant information in long-context scenarios. To address the issues, we propose AgentCPM-Explore, a compact 4B agent model with high knowledge density and strong exploration capability. We introduce a holistic training framework featuring parameter-space model fusion, reward signal denoising, and contextual information refinement. Through deep exploration, AgentCPM-Explore achieves state-of-the-art (SOTA) performance among 4B-class models, matches or surpasses 8B-class SOTA models on four benchmarks, and even outperforms larger-scale models such as Claude-4.5-Sonnet or DeepSeek-v3.2 in five benchmarks. Notably, AgentCPM-Explore achieves 97.09% accuracy on GAIA text-based tasks under pass@64. These results provide compelling evidence that the bottleneck for edge-scale models is not their inherent capability ceiling, but rather their inference stability. Based on our well-established training framework, AgentCPM-Explore effectively unlocks the significant, yet previously underestimated, potential of edge-scale models.

Figures (8)

• Figure 1: Overall Training Framework of AgentCPM-Explore.
• Figure 2: Overall performance of different summary models in the same agent system on the GAIA benchmark. Abbreviations are as follows. DS: DeepSeek; Qwen3-4B-I: Qwen3-4B-Instruct-2507; Qwen3-4B-T: Qwen3-4B-Thinking-2507; FT: Fine-tuned.
• Figure 3: RL learning curves in different train settings.
• Figure 4: Pass@K performance comparison on the GAIA benchmark. The figure illustrates the capability boundary of the 4B model by comparing the "SFT-Merge" baseline with the final AgentCPM-Explore (RL after merging). The Qwen3-4B-thinking-2507 serves as the base model.
• Figure 5: Standardized agentic training and inference infrastructure.