Interaction as Intelligence Part II: Asynchronous Human-Agent Rollout for Long-Horizon Task Training

TL;DR

Apollo addresses the challenge of training LLM agents for long-horizon, domain-specific tasks by introducing asynchronous human guidance coupled with action-level supervision. It blends a lightweight human–AI interaction interface, ReAct-inspired trajectory handling, and long-context summarization to stabilize data collection and learning. Empirical results on InnovatorBench show substantial performance gains over untrained baselines and non-interactive variants, demonstrating improved data efficiency and transfer to new tasks. The work suggests a practical, scalable pathway for high-quality data synthesis in long-horizon AI research and lays groundwork for future multi-domain and multi-agent extensions.

Abstract

Large Language Model (LLM) agents have recently shown strong potential in domains such as automated coding, deep research, and graphical user interface manipulation. However, training them to succeed on long-horizon, domain-specialized tasks remains challenging. Current methods primarily fall into two categories. The first relies on dense human annotations through behavior cloning, which is prohibitively expensive for long-horizon tasks that can take days or months. The second depends on outcome-driven sampling, which often collapses due to the rarity of valid positive trajectories on domain-specialized tasks. We introduce Apollo, a sampling framework that integrates asynchronous human guidance with action-level data filtering. Instead of requiring annotators to shadow every step, Apollo allows them to intervene only when the agent drifts from a promising trajectory, by providing prior knowledge, strategic advice, etc. This lightweight design makes it possible to sustain interactions for over 30 hours and produces valuable trajectories at a lower cost. Apollo then applies supervision control to filter out sub-optimal actions and prevent error propagation. Together, these components enable reliable and effective data collection in long-horizon environments. To demonstrate the effectiveness of Apollo, we evaluate it using InnovatorBench. Our experiments show that when applied to train the GLM-4.5 model on InnovatorBench, Apollo achieves more than a 50% improvement over the untrained baseline and a 28% improvement over a variant trained without human interaction. These results highlight the critical role of human-in-the-loop sampling and the robustness of Apollo's design in handling long-horizon, domain-specialized tasks.

Figures (10)

• Figure 1: Apollo allows humans to instruct Agents when they make both technical errors and strategic errors asynchronously and trains the model with correct steps.
• Figure 2: The pipeline of Apollo.
• Figure 3: The infrastructure bridges humans, agents, and computers to enable scalable long-duration (10–30 h) interaction during autonomous research discovery.
• Figure 4: The display of training trajectory format. Only the green line step will be trained. In the summary trajectory, the correct step is not trained.
• Figure 5: Comparisons between Apollo and the original GLM-4.5.