CoACT: Action-Preserving Observation Compression for Coding Agents

This repository contains the pretrained observation compressor released with CoACT. The model is a merged Qwen3.5-4B checkpoint trained from trajectories collected with Qwen3.5-35B-A3B.

CoACT compresses each new environment observation before it enters a coding agent's trajectory. It is trained with reward-selected supervision that favors compact observations while preserving the agent's next action.

Cross-Agent Generalization

Our cross-agent generalization experiments show that compressors trained from different agentic models achieve similar performance when transferred across agents. When evaluated with Deepseek-v4-Pro, the compressor trained from Qwen3.5-35B-A3B trajectories achieves 74.5% pass@1 with 0.863M total tokens per instance, close to 75.0% pass@1 and 0.868M total tokens for the compressor trained from Deepseek-v4-Pro trajectories. These results suggest that this checkpoint can be used across agentic models without separately training a compressor for each one, while agent-specific training may still provide a small performance advantage. We therefore release it as the default CoACT compressor for use across agentic models.

Download

hf download Kndy666/CoACT --local-dir checkpoints/CoACT

For deployment and evaluation instructions, see the CoACT repository.

For paper, see the Paper.