Reproduce “AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Susan Liang, Chao Huang, Filippos Bellos, Jing Bi, Jason J Corso, Chenliang Xu
AdaTurn introduces a framework that enables active visual perception agents to adapt their behavior to varying computational budgets. It explicitly conditions the agent's policy on the maximum allowed turns and employs a specialized reinforcement learning strategy called Forced-Answer DAPO (FA-DAPO), yielding performance improvements in low-budget scenarios (e.g., +12.5% on VisualProbe-Hard at 4 turns) while maintaining efficacy at higher budgets.
Key insights
- Explicitly conditioning an active agent's policy on the maximum available turns allows it to learn distinct behaviors optimized for specific budget constraints.
- Training the agent's behavior at the precise budget boundary, where tool use is no longer allowed, is critical for mitigating catastrophic truncation and ensuring a coherent answer is always produced.
- Efficiently managing dynamic rollout budgets during training through a load-balanced scheduler is essential for making budget-aware agent development practical and scalable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.