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

AgentCPM Team · 2026

Read AgentCPM-Explore: Realizing Long-Horizon Deep Exploration for Edge-Scale Agents by AgentCPM Team (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-06485.txt.

Description

Most capable AI agents run on large models, leaving smaller edge-scale models underexplored. Haotian Chen and 18 co-authors present AgentCPM-Explore, a compact four-billion-parameter agent with unusually strong exploration skills. Small models face three blocks: forgetting during supervised training, noise in reward signals during reinforcement learning, and reasoning rot from redundant long-context material.

Their training framework answers each block with a matching trick: fusing models in parameter space, cleaning reward signals, and trimming context. The resulting agent led its size class on benchmarks, matched or beat eight-billion-parameter leaders on several, and even outperformed much larger models on a few. On a text-based general assistant benchmark, it scored very high under repeated sampling.

Catalog record

CREATORAgentCPM Team
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06485

Browse the full library · About the Tower of Babel