AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks
Fali Wang, Hui Liu, Zhenwei Dai, Jingying Zeng, Zhiwei Zhang, Zongyu Wu, Chen Luo, Zhen Li, Xianfeng Tang, Qi He, Suhang Wang
2025-08-05
Summary
This paper talks about AgentTTS, a smart system that uses large language model agents to decide the best way to use computing power during complex tasks with many stages.
What's the problem?
The problem is that complex tasks often have many parts that use different amounts of computing resources, and managing how much computing each part gets is difficult but important for good performance.
What's the solution?
AgentTTS solves this by using an LLM agent to dynamically allocate computing resources at each stage of the task, optimizing how the system works as it proceeds instead of using fixed or manual resource assignments.
Why it matters?
This matters because it makes AI systems more efficient and reliable when handling complicated tasks, helping them perform better without wasting computing power.
Abstract
AgentTTS, an LLM-agent-based framework, optimizes compute allocation for multi-stage complex tasks, improving performance and robustness compared to traditional methods.