AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents
AdaTurn improves low-budget active visual reasoning by explicit budget conditioning and boundary-aware reinforcement learning, boosting accuracy by over 10%.
Key Findings
Methodology
AdaTurn employs a conditional policy that takes maximum turn count as input, integrating FA-DAPO reinforcement learning to convert over-budget events into trainable boundary decisions. During training, it samples various budgets randomly, learning boundary behaviors across scales. A load-balanced scheduler manages multi-engine environments, ensuring efficient training and inference. The core algorithm extends DAPO, optimizing boundary actions with advantage-based rewards, enabling the model to adaptively trade off between searching and answering based on available resources.
Key Results
- At 4 turns, VisualProbe-Medium improves from 36.7% to 47.6%, with similar gains across benchmarks. The model maintains high performance at larger budgets, transferring well to different backbones and tasks. The load-balanced scheduler accelerates throughput by 1.34×, making training feasible with dynamic budgets.
- Compared to state-of-the-art Mini-o3, AdaTurn surpasses it by over 10% in low-budget regimes, especially on VisualProbe-Hard and MME-RealWorld datasets, demonstrating robustness and generalization.
- Experimental results highlight the effectiveness of boundary training and scheduling, with significant improvements in resource-constrained scenarios, validating the approach’s practical value.
Significance
This work addresses a critical gap in deploying multi-turn visual reasoning systems under resource constraints. By explicitly modeling budget as an input and training boundary behaviors, AdaTurn enhances robustness and accuracy in real-world applications such as interactive AI assistants and autonomous robots. It enables models to make optimal decisions when resources are limited, reducing catastrophic truncation and resource wastage, thus advancing the field towards practical, scalable multimodal systems.
Technical Contribution
The key innovation is the integration of budget conditioning with a boundary-aware reinforcement learning strategy (FA-DAPO), transforming over-budget failures into learnable decision points. The system also introduces dynamic sampling of budgets and a load-balanced scheduler, significantly improving training efficiency and system robustness. These contributions enable multi-scale, resource-adaptive reasoning, setting a new standard for deployment-aware multimodal agents.
Novelty
This is the first approach to explicitly incorporate budget as a conditioning variable in multi-turn visual reasoning, with a novel boundary training mechanism that teaches models to synthesize evidence at resource limits. Unlike prior methods that treat over-budget cases as errors or mask them, AdaTurn actively learns to handle the boundary, representing a fundamental shift in training paradigm for resource-constrained inference.
Limitations
- Despite improvements, the model may still struggle in extremely tight budgets (e.g., 1-2 turns) where information is insufficient for accurate reasoning. Further optimization of tool usage and evidence compression is needed.
- System complexity increases with dynamic scheduling and boundary training, posing challenges for real-world deployment. Future work should focus on simplifying architecture and reducing computational costs.
- Current experiments are mainly on benchmark datasets; real-world scenarios with more diverse and noisy data require additional validation.
Future Work
Future directions include integrating more efficient tool calling strategies, exploring adaptive budget prediction, and extending to more complex, real-world environments. Enhancing model robustness under extreme resource constraints and reducing system complexity will be key to practical deployment. Additionally, combining this approach with continual learning could enable models to adapt to evolving resource profiles.
AI Executive Summary
AdaTurn introduces a novel framework for active visual reasoning that explicitly incorporates deployment resource constraints into the decision-making process. Traditional models often ignore the variability in available inference budgets, leading to premature truncation or underutilization of resources. AdaTurn addresses this by conditioning the policy on the maximum allowed turns, enabling the model to adapt its behavior dynamically. The core innovation lies in FA-DAPO, a reinforcement learning strategy that transforms over-budget tool requests into trainable boundary decisions, teaching the model to synthesize partial evidence when further tool use is impossible. This boundary-aware training ensures the model can make intelligent decisions at the resource limit, avoiding catastrophic truncation and optimizing performance across a range of budgets.
To facilitate training under variable budgets, AdaTurn employs a dynamic sampling approach, randomly selecting different maximum turn counts during training. This exposes the model to diverse resource scenarios, improving its adaptability. Additionally, a load-balanced scheduler distributes rollout requests across multiple engines, reducing system bottlenecks and increasing throughput by 1.34×. Experimental results demonstrate that AdaTurn significantly outperforms existing methods like Mini-o3, especially in low-budget settings, with improvements exceeding 10% on benchmarks such as VisualProbe-Medium and MME-RealWorld.
The approach's transferability to different backbone architectures and its robustness across multiple multimodal benchmarks highlight its practical relevance. By enabling models to operate effectively under resource constraints, AdaTurn paves the way for deploying multi-turn visual reasoning systems in real-world, resource-limited environments. Future work will focus on further optimizing tool use, extending to more complex scenarios, and integrating adaptive resource prediction mechanisms, aiming to make active visual agents more efficient, reliable, and scalable.
Deep Analysis
Background
多模态视觉推理已从早期的静态单次预测发展到多轮交互式推理。代表性工作如LLaVA、Qwen-VL和InternVL，通过增强视觉编码和动态切片技术提升高分辨率理解能力。主动感知方法如SEAL、DC2和Chain-of-Focus强调多轮信息采集，显著改善复杂场景表现。强化学习模型如DeepEyes和Mini-o3引入工具调用，增强交互深度，但在实际部署中，资源限制未被充分考虑，导致模型在有限预算下表现不佳，存在灾难性截断问题，亟需引入预算感知机制以提升鲁棒性。
Core Problem
现有多轮视觉推理模型多在训练时忽略部署环境中的预算变化，导致模型在实际场景中出现性能退化。具体表现为，模型在预算不足时过度请求工具，提前终止，无法提供有效答案；在预算充裕时，未能充分利用剩余资源进行验证。这种盲目追求长短期推理能力的做法，限制了模型在实际应用中的表现。解决方案应在训练中明确预算条件，学习在不同资源限制下的最优行为，尤其是在预算边界的决策能力，避免灾难性截断，提升低预算场景的表现。
Innovation
本研究提出将最大回合数作为模型输入条件，实现预算感知的主动推理。引入FA-DAPO强化学习策略，将超预算事件转化为模型训练中的边界决策目标，增强模型在资源有限时的边界行为学习能力。系统还设计了动态采样和负载均衡调度机制，确保多引擎环境下的训练效率。与传统仅在训练或推理中硬性限制的方案不同，AdaTurn在边界行为上实现了显著改进，增强模型在不同预算场景下的适应性和鲁棒性。这一创新突破了预算盲训练的局限，为多轮主动视觉推理提供了新的思路。
Methodology
- �� 将最大回合数Tmax作为模型输入条件，使模型在每轮决策时考虑预算信息。• 采用FA-DAPO强化学习策略，将超预算的工具调用转为边界决策训练目标。• 在训练中随机采样不同预算，增强模型在多尺度场景下的适应能力。• 引入负载均衡调度器，动态分配多引擎资源，提升训练和推理效率。• 设计边界行为的训练机制，使模型在预算边界学会合理压缩推理或提前回答。• 结合动态预算采样和边界训练，确保模型在不同预算下的表现均衡。• 训练过程中利用奖励函数结合准确率和格式一致性，优化模型决策。• 在推理时，根据实际预算条件调整模型行为，确保性能最大化。
Experiments
采用VisualProbe、V* Bench、HR-Bench和MME-RealWorld等多模态推理基准，评估AdaTurn在不同预算下的性能。对比基线Mini-o3和DeepEyes，验证低预算场景的提升。实验中随机采样Tmax，模拟实际部署环境。指标包括准确率、推理速度和系统吞吐量。通过消融实验验证FA-DAPO的边界训练效果，分析调度器对训练效率的影响。多场景迁移测试表明模型具有良好的泛化能力。超参数调优确保模型在不同预算下的最优表现，验证其资源适应性。
Results
在4回合预算下，AdaTurn在VisualProbe-Medium上达47.6%，比之前最优提升10%以上。在更大预算下，性能保持或超越对比模型，验证了边界训练的有效性。引入调度器后，训练和推理速度提升1.34倍。模型在多模态基准上表现优异，验证其广泛适用性。低预算性能提升尤其明显，解决了灾难性截断问题，增强了实际部署的鲁棒性。
Applications
该方法适用于需要多轮交互的视觉问答、场景理解等应用，特别是在边缘设备或实时系统中资源有限的场景。模型可根据实际资源动态调整推理深度，提升效率和准确率。未来可结合自动预算调节机制，实现自适应资源管理，推动智能监控、机器人等行业的技术落地。
Limitations & Outlook
模型在极端预算（如1-2回合）下仍可能因信息不足导致误判。系统复杂度较高，训练和调度机制增加了实现难度。当前实验主要在基准测试中验证，泛化到更复杂、多样化场景仍需进一步研究。未来需优化工具调用策略和模型压缩技术，以降低部署成本。
Plain Language Accessible to non-experts
想象你在一个工厂工作，工厂每天都要完成不同的任务。有时候任务很简单，只需要用少量工具和时间就能完成；有时候任务很复杂，需要用很多工具、花费很多时间。以前工厂的工人总是按照固定流程工作，不管任务难易，都用一样的时间和工具。这样一来，遇到简单任务时，工人可能会用太多时间，浪费资源；遇到复杂任务时，又可能因为时间不够而提前放弃，导致任务没完成。AdaTurn就像给工厂的工人装上了智能助手，告诉他们根据任务的难度和剩余时间，合理安排工具和时间。它学会在资源有限时，快速做出最合理的决定，既不浪费，也不提前放弃。这样，无论任务多复杂或多简单，工厂都能高效完成，资源得到最优利用。
ELI14 Explained like you're 14
你可以把AdaTurn想象成一个聪明的学生在做作业。有时候作业很简单，只需要几步就能完成；有时候很难，要花很多时间查资料、写答案。以前这个学生总是按照固定的时间和步骤做，不管作业难不难。有时候，他会花太多时间在简单题上，浪费时间；有时候，时间不够，没能完成难题。AdaTurn就像给这个学生装了个智能助手，告诉他根据剩余时间和作业难度，合理安排做题策略。它学会在时间有限时，快速总结已知信息，做出最合理的答案，而不是一直追求完美。这样，无论作业难易，学生都能高效完成，得到好成绩。
Glossary
Budget-Aware Active Visual Agent (预算感知主动视觉代理)
一种能根据资源限制调整行为的多轮视觉推理模型，结合强化学习优化边界决策。
论文中提出的核心系统架构。
Forced-Answer DAPO (FA-DAPO)
一种强化学习策略，将超预算事件转化为模型训练中的边界决策目标，增强模型在预算边界的表现。
模型训练和边界行为学习的关键技术。
rollout budget (回合预算)
在多轮交互中允许的最大操作次数，用于控制推理深度和系统资源消耗。
模型输入条件之一，影响模型行为。
load-balanced scheduler (负载均衡调度器)
一种动态调度机制，合理分配多引擎资源，提升训练和推理效率。
系统优化的重要组成部分。
catastrophic truncation (灾难性截断)
模型在预算不足时提前终止，导致无法完成有效推理或回答的现象。
本文关注的主要问题之一。
Open Questions Unanswered questions from this research
- 1 如何在极端预算（如1-2回合）下保持模型性能，仍是未解难题。模型在信息不足时的决策机制需要进一步优化。未来研究应结合更高效的工具调用和信息压缩策略，以应对更复杂的实际场景。
Applications
Immediate Applications
智能问答系统
在有限时间内提供准确答案，特别适用于边缘设备或实时应用。模型根据资源动态调整推理深度，提升效率和准确性。
机器人视觉导航
机器人在复杂环境中有限的交互轮次内，快速理解场景并做出决策，增强自主能力。
Long-term Vision
自适应智能系统
未来系统能根据环境变化自动调节推理深度，实现资源最优利用，推动智能监控、自动驾驶等行业发展。
Abstract
Active visual agents solve fine-grained image tasks by interleaving reasoning with image-grounding actions across multiple turns. However, deployment-time rollout budgets are rarely fixed: some requests permit long rollouts, while others require the agent to act under a tight turn limit. Existing methods train the policy as if the rollout budget were hidden, so when the available budget is smaller than the trajectory the agent prefers, the interaction is often truncated before any valid answer is produced; we term this failure \emph{catastrophic truncation}. To overcome this challenge, we present AdaTurn, a budget-aware framework that conditions the agent on the allowed number of turns and explicitly trains the boundary behavior induced by the budget. Our key component, Forced-Answer DAPO (FA-DAPO), converts the over-budget event from a masked or penalized failure into a trainable final-decision step, teaching the model to synthesize partial evidence when further tool use is no longer possible. We further randomize rollout budgets during both training and inference and introduce a load-balanced scheduler that makes such operations practical. AdaTurn substantially improves low-budget accuracy, for example raising VisualProbe-Medium from 36.7% to 47.6% at four turns, while preserving strong scaling at larger budgets and transferring effectively to multiple backbones and general multimodal benchmarks.