SyncPlan: Long-Horizon LLM Coordination with Explicit Synchronization and Adaptive Correction

Imagine a group of friends trying to solve a massive, chaotic puzzle together in a video game. They have a shared goal, like defeating a boss or cooking a complex meal, but the game world is unpredictable: enemies pop up, items disappear, and teammates might get stuck. To succeed, they need to coordinate perfectly. This is the world of Multi-Agent Coordination, a field where computer scientists teach artificial intelligence (AI) agents to work as a team.

Traditionally, these teams rely on two main strategies. The first is Reinforcement Learning (RL), where agents learn through trial and error, like a dog learning tricks with treats. It's fast and efficient but often requires massive amounts of specific training for just one game, making it hard to adapt to new situations. The second strategy uses Large Language Models (LLMs), the same "brainy" AI that can write poems or chat with you. These models are great at understanding complex instructions and planning, but they are slow. Asking a super-smart AI to think through every single move in real-time is like asking a genius chef to write a new recipe for every single chop of a knife—it takes too long, and by the time the recipe is written, the ingredients have already burned.

The big question researchers are trying to answer is: How can we get the smart, flexible planning of a language model without the slow, clunky speed that makes it useless in fast-paced games?

Enter SyncPlan, a new framework proposed by researchers from City University of Hong Kong, Tencent, and others. Think of SyncPlan as a "Plan-Execute-Correct" system designed to be the ultimate team captain for AI agents. Instead of asking the AI to think constantly, SyncPlan asks it to write a long, detailed script (a "joint plan") for the whole team in one go. This script tells each agent exactly what to do, when to move, and crucially, when to wait.

Here's where it gets clever. In the past, if an AI team planned to "Attack the boss together," one agent might rush in while the other was still walking, causing a disaster. SyncPlan solves this with Explicit Synchronization. It uses special "wait" commands, like a traffic light, that force an agent to pause until a teammate arrives or an enemy is in the right spot. This turns vague ideas like "work together" into strict, machine-readable rules that prevent the team from tripping over each other.

But what if the game changes? What if the boss suddenly runs away? If the team blindly follows the old script, they fail. SyncPlan has a secret weapon: a lightweight Plan Staleness Detector (PSD). Imagine a vigilant spotter standing on the sidelines. This spotter constantly checks the game state against the current plan. If the spotter sees that the plan is no longer valid (e.g., the boss is gone), it instantly signals the "captain" (the LLM) to write a new script. If the plan is still good, the spotter stays silent, and the team keeps running without interruption. This means the system only calls the slow, thinking AI when it absolutely has to, saving massive amounts of time.

The researchers tested this system in two very different worlds: Overcooked, a chaotic cooking game where two agents must prepare soup, and Honor of Kings, a complex 5-vs-5 battle arena game. The results were impressive. In the cooking game, SyncPlan succeeded in tasks more often than previous methods while using less than 0.05% of the time those other methods took. In the battle arena, it achieved an 86.3% success rate, beating the next-best method by a wide margin and running 26 times faster.

The paper suggests that this approach works because it separates the heavy thinking (planning) from the fast action (executing). By using Supervised Fine-Tuning (SFT) to teach the AI how to write these structured scripts and Reinforcement Learning (RL) to refine them based on real-game feedback, the system learns to be both smart and speedy. The authors found that without the "spotter" (the PSD), the team often got stuck or followed bad plans, and without the "wait" commands, they collided and failed.

In short, SyncPlan doesn't try to make the AI think faster; it makes the AI think smarter by planning ahead, waiting for the right moment, and only re-thinking when the world changes. It's a shift from constantly asking "What should I do?" to "Here is the plan, and here is exactly when to change it," turning a group of slow, smart thinkers into a fast, synchronized team.