acceptodds
Under review as a conference paper at ICLR 2027

AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents

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 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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Agentic LLM, Vision-Language Model, Test-Time Scaling, Budget Awareness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Yn7vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies active visual agents operating under varying interaction budgets. It introduces FA-DAPO which improve on top of standard DAPO and adapt it towards a turn-budget aware setting. The paper also present a simple practical nudge to improve training efficiency under varied turn budgets. In experiments, proposed pipeline show descent improvement under controlled budget comparing to vanilla training pipeline.

Critical strengths
  1. This work is focusing on a practical yet oftenly overlooked problem, and the addressing the issue does carry certain practical benefits.

  2. The propose method is simple and effective. Propose method shows strong edge comparing to vanilla training method.

Critical weaknesses
  1. Reviewer hates to say so, but there is very limited technical and science innovation. The method presented in the paper is an incremental change or adaption on top of DAPO plus a engineering tricks. Also work like AnytimeReasoner [1] already trains on answers elicited from interrupted reasoning under sampled budgets. AdaTurn extends this general idea to visual tool-use trajectories, but the experiments should validate what its particular training choices add.

  2. Turn budget does not fully reflect comparable inference cost. Eq. 7 first generates an inadmissible final-turn tool request, discards it, and generates a replacement answer. A retained trajectory of T turns can therefore require T+1 model generations. Table 6 provides useful six-turn token statistics without clearly explaining whether discarded generations and repeated processing are counted.

[1] Optimizing Anytime Reasoning via Budget Relative Policy Optimization

Questions for authors
  1. Can you compare fixed versus randomized training budgets and ordinary DAPO versus FA-DAPO with matched initialization, data, and inference-time final-answer handling?

  2. How are discarded generations, replacement answers, and repeated prefill counted? Does AdaTurn retain its advantage over budget-prompted Mini-o3 when accuracy is compared against actual model invocations or measured latency?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Yn7v [Part 1]

We sincerely thank the reviewer for the careful reading and constructive feedback. We are encouraged that the reviewer finds the problem "practical yet often overlooked," the method "simple and effective," and that it "shows strong edge comparing to vanilla training method." Below we address each concern in detail.

W1: Novelty and relation to AnytimeReasoner. AdaTurn contributes FA-DAPO, which converts over-budget tool requests into trainable final-answer steps and optimizes the full trajectory with group-normalized reward, and a load-balanced rollout scheduler that reduces worker imbalance under dynamic budgets and yields a 1.34×1.34\times throughput improvement. Both designs are driven by concrete failure modes in real deployment.

We also explored more "novel" alternatives in Appendix C. Masking the prefix turns so that only the forced-answer step receives a gradient degrades performance, suggesting that budget-aware behavior is a trajectory-level property that must be learned jointly with the preceding search decisions rather than as an isolated correction at the last turn. Restricting RL to only the final forced-answer step, where all trajectories in a group share the same prefix, performs worse still: without signal flowing back through the full rollout, the policy cannot learn to allocate its turns wisely. Taken together, these ablations directly validate what each of AdaTurn's particular training choices contributes: forced-answer regeneration over masking or penalizing (FA-DAPO vs. alternatives), full-rollout optimization over final-step-only training, and FA-DAPO over ordinary DAPO with inference-time final-answer handling (Q1 below). These experiments show that although our default design may appear less "novel" and "simple," it is the most effective approach, and that adding complexity does not improve performance.

We thank the reviewer for sharing AnytimeReasoner. AnytimeReasoner trains a language model to produce good answers at any point during its reasoning chain by sampling token-level budgets, truncating the thinking trace at the sampled budget, and supervising the model to summarize a best answer from the truncated prefix. It is a token-level, text-only study evaluated on mathematical reasoning tasks, with no visual inputs, no external tool calls, and no multi-turn agent loop. AdaTurn addresses a complementary setting: turn-level budgets, multimodal inputs, and a multi-turn agent loop where each turn involves an external image-grounding tool call. The budget boundary in AdaTurn is a discrete decision point between tool use and final-answer generation, rather than a truncation inside a single reasoning trace, which requires a different training formulation (FA-DAPO) and different systems support (load-balanced scheduling). We will discuss AnytimeReasoner in the related work section in the revision.

W2 and Q2: Token cost of discarded generations. The token counts in Table 6 include all tokens generated during the rollout, including the discarded final tool request. The discarded generation is thus already accounted for in the reported output token count, and the efficiency advantage of AdaTurn holds even when these additional tokens are counted. Furthermore, the extra cost of the discarded generation is limited compared to a full extra turn. When FA-DAPO detects an over-budget tool call and rolls back to the history before the last turn, the prefix up to that point is already in the KV cache. The re-generation of the forced-answer step therefore only needs to decode from the cached prefix without re-prefilling the full context, so the additional latency and compute are limited compared to a full extra turn. We will clarify these points explicitly in the revision.

AuthorsOpenReview ↗

Response to Reviewer Yn7v [Part 2]

Q1: Fixed versus randomized budgets, and ordinary DAPO versus FA-DAPO. Given the limited rebuttal period, we cannot provide per-benchmark ablation results for a full quantitative comparison. From our preliminary experiments, training with a dynamic budget range outperforms a fixed budget because it exposes the model to diverse training conditions: different budget levels demand qualitatively different behaviors, such as aggressive search compression under a tight budget versus extended verification under a generous one. Training across this range teaches the model to maximize performance for whatever budget it receives.

Regarding ordinary DAPO versus FA-DAPO, the budget-prompted Mini-o3 evaluated in Appendix E (Table 5) can be viewed as an approximation of ordinary DAPO combined with inference-time final-answer handling, since both Mini-o3 and AdaTurn use the same backbone and training data, and the budget prompt replicates the forced-answer mechanism at inference time without RL training. The comparison below shows that AdaTurn (FA-DAPO) still outperforms this combination on six of the seven benchmarks, demonstrating that explicitly training the budget boundary via FA-DAPO provides a clear advantage over ordinary DAPO with inference-time handling.

Model (4 turns)VP-HardVP-MediumVP-EasyV*HR-4KHR-8KMME-RW
Mini-o3 w/ budget prompt36.147.758.181.973.167.759.7
Ours (FA-DAPO)39.047.661.285.275.771.164.0
Reviewer_4LRHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies active visual agents under varying interaction-turn budgets, focusing on cases where an agent continues calling tools until its budget is exhausted without producing a valid answer. AdaTurn conditions the policy on the available budget, samples different budgets during training, and introduces Forced-Answer DAPO, which replaces an over-budget tool request with a final-answer generation based on the evidence collected so far. The paper also introduces budget-based load balancing to improve rollout throughput. Experiments across visual search and high-resolution understanding benchmarks, using multiple backbone sizes, demonstrate improvements under tight budgets while largely preserving performance at larger budgets.

Critical strengths

Turning budget exhaustion into a trainable final decision directly addresses the limitations of penalizing or discarding truncated trajectories. Optimizing the complete trajectory also allows the quality of the final answer to reinforce earlier evidence-acquisition decisions.

Low-budget gains are not confined to a single benchmark or model. In particular, after introducing the budget-prompted Mini-o3 baseline in the appendix, AdaTurn still improves performance on six of seven evaluations by approximately 2.6–4.3 percentage points. This supports a contribution beyond simply informing the agent of its remaining budget.

Critical weaknesses

The abstract highlights an improvement from 36.7% to 47.6% on VisualProbe-Medium, but budget-prompted Mini-o3 already achieves 47.7%. Although AdaTurn retains gains on the other six evaluations, this comparison materially changes the interpretation of the contribution and should be included in the main results rather than relying primarily on a budget-blind baseline.

The method first generates a tool request that is discarded, then rolls back and generates a final answer. Consequently, the same nominal turn budget may involve different numbers of model calls and different amounts of generation compute. The token-cost analysis does not sufficiently clarify whether this discarded generation is included, which affects the fairness of the claimed low-budget advantage.

Questions for authors

Does Table 6 include the tokens and computation consumed by the discarded final tool request? When all model calls and generated tokens are counted, does AdaTurn retain its accuracy–cost advantage under tight budgets?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4LRH

We sincerely thank the reviewer for the careful reading and constructive feedback. We are encouraged that the reviewer recognizes that "turning budget exhaustion into a trainable final decision directly addresses the limitations of penalizing or discarding truncated trajectories," and that "low-budget gains are not confined to a single benchmark or model," with AdaTurn improving "performance on six of seven evaluations by approximately 2.6–4.3 percentage points" even after controlling for budget prompting. Below we address each concern.

W1: Budget-prompted Mini-o3 as the main baseline. We agree with the reviewer that the budget-prompted Mini-o3 is the more informative comparison and that it should appear in the main paper rather than the appendix. AdaTurn outperforms the budget-prompted Mini-o3 on six of the seven benchmarks under a four-turn budget, with gains of +2.9%+2.9\% on VisualProbe-Hard (39.0 vs. 36.1), +3.1%+3.1\% on VisualProbe-Easy (61.2 vs. 58.1), +3.3%+3.3\% on V* (85.2 vs. 81.9), +2.6%+2.6\% on HR-Bench 4K (75.7 vs. 73.1), +3.4%+3.4\% on HR-Bench 8K (71.1 vs. 67.7), and +4.3%+4.3\% on MME-RealWorld (64.0 vs. 59.7). The only benchmark where the two models are on par is VisualProbe-Medium (47.6 vs. 47.7). As the reviewer notes, these consistent margins across six evaluations support a contribution that goes beyond simply informing the agent of its remaining budget. We will move this comparison into the main results table in the revision.

W2 and Q1: Token cost of the discarded generation. The token counts in Table 6 include all tokens generated during the rollout, including the discarded tool request. The discarded generation is thus already accounted for in the reported output token count, and the efficiency advantage of AdaTurn holds even when this additional generation is included. Furthermore, the extra cost of the discarded generation, in practice is limited. When FA-DAPO detects an over-budget tool call and rolls back to the history before the last turn, the prefix up to that point is already in the KV cache. The re-generation of the forced-answer step therefore only needs to decode from the cached prefix without re-prefilling the full context, so the additional latency and compute are limited compared to a full extra turn. We will clarify these two points explicitly in the revision.

Thank you for the clarification. My main concerns are addressed. Please clarify in the revision that the reported efficiency advantage concerns output-token usage, and qualify any claims about end-to-end latency or total inference compute accordingly. I will keep my rating,

Reviewer_3phyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AdaTurn, a budget-aware framework for active visual agents that addresses catastrophic truncation under limited interaction turns. It introduces Forced-Answer DAPO (FA-DAPO) to train agents to produce valid answers from partial evidence when the turn budget is exhausted, alongside budget randomization and load-balanced scheduling. Experiments demonstrate substantial improvements in low-budget accuracy while maintaining strong performance under larger budgets and generalizing across different backbones and multimodal benchmarks.

Critical strengths
  1. The problem that this paper tries to address is important. By developing a budget-aware framework, it will improve the reasoning efficiency.

  2. The AdaTurn, a budget-conditioned visual agent, is proposed to explicitly train the final-answer decision at the budget boundary.

  3. The proposed AdaTurn delivers gains in experiments, showing the effectiveness of the proposed method.

Critical weaknesses
  1. Although the problem is important and interesting, the proposed method is not novel. It is hard to identify the novel design from the proposed framework as shown in Figure 2 (a). It seems it is just a workflow.

  2. In Table 1, the proposed method with 32 turns only achieves 48.1 (hard) and 66.7 (easy) in VisualProbe benchmark, which are with marginal gains compared to the results of 48.0 (hard) and 67.0 (easy) achieved by Mini-o3. Suggest adding the efficiency results to reflect the advantage of the proposed method.

  3. In Figure 6, do all the methods use the same backbones for fair comparison?

  4. What is the significance compared to the set of "think with images" [a, b]? The significance compared to these works should be discussed in the paper.

[a] Deepeyes: Incentivizing" thinking with images" via reinforcement learning @ ICLR 2026

[b] VisionLeaf: Entropy-Guided Leaf-First Reasoning for Efficient and Accurate Think-with-Image @ CVPR 2026

Questions for authors

See above

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3phy [Part 1]

We sincerely thank the reviewer for the careful reading and constructive feedback. We are encouraged that the reviewer finds the problem "important and interesting," recognizes AdaTurn as "a budget-conditioned visual agent that explicitly trains the final-answer decision at the budget boundary," and acknowledges that "the proposed AdaTurn delivers gains in experiments, showing the effectiveness of the proposed method." Below we address each concern in detail.

W1: Novelty. We appreciate the reviewer's observation, and we would like to clarify that Figure 2 (a) is the agent loop workflow showing how AdaTurn interacts with the zoom-in tool during problem solving. The novelty and technical contributions of AdaTurn are captured in Figure 3 and Figure 4. Figure 3 presents Forced-Answer DAPO (FA-DAPO), our algorithmic contribution: when the rollout calls the tool at the last permitted turn, FA-DAPO discards the tool call, appends a forced-answer prompt, and regenerates the final turn, converting the over-budget event into a trainable decision with group-normalized reward. This is a principled departure from prior strategies, which either penalize or mask truncated rollouts, and it is the primary source of AdaTurn's low-budget gains. Figure 4 presents the system-side contribution: a load-balanced rollout scheduler that sorts requests by turn budget and greedily assigns each to the least-loaded engine, reducing rollout imbalance and yielding a measured 1.34×1.34\times throughput improvement.

We further note that in the ablation studies (Appendix C), we also investigated more complex alternatives that might appear "more novel," including masking prefix turns so that only the forced-answer turn is trained, and running GRPO only at the final forced-answer step when all trajectories in a group share the same prefix. Both of these variants underperform the default full-rollout FA-DAPO design, indicating that although our approach is "simple," it is the most "effective" one, and that forcing complexity does not improve performance.

W2: Performance at 32 turns and efficiency. Our work is specifically designed for the budget-limited deployment setting, and the core claim is that AdaTurn delivers large improvements when the available budget is tight. At four turns, AdaTurn outperforms Mini-o3 by +12.5%+12.5\% on VisualProbe-Hard (39.0 vs. 26.5), +10.9%+10.9\% on VisualProbe-Medium (47.6 vs. 36.7), +13.9%+13.9\% on MME-RealWorld (64.0 vs. 50.1), +6.0%+6.0\% on V* (85.2 vs. 79.2), +6.8%+6.8\% on HR-Bench 4K (75.7 vs. 68.9), and +7.9%+7.9\% on HR-Bench 8K (71.1 vs. 63.2). These are the gains that matter most when deployment constraints bind. When the budget is sufficient, the performance gap shrinks because AdaTurn uses the same backbone and training data as Mini-o3 and each agent can call enough tools to solve a task.

Regarding the reviewer's suggestion to add efficiency results, we can provide them from Appendix F (Table 6). Under a six-turn budget, AdaTurn and Mini-o3 have a similar total token cost, while AdaTurn generates considerably fewer output tokens both in total and per turn, so its low-budget gains are not obtained by spending more tokens.

Token usage (6 turns)Mini-o3Ours
Total tokens5526.64 ±\pm 293.255713.39 ±\pm 478.61
Output tokens698.96 ±\pm 67.62504.24 ±\pm 35.60
Output tokens per turn243.88 ±\pm 15.62163.62 ±\pm 11.67

We will add these efficiency results to the main paper in the revision.

W3: Backbone in Figure 6. Yes, all methods shown in Figure 6 use the same Qwen2.5-VL-7B-Instruct backbone, including DeepEyes, Mini-o3, and AdaTurn. The performance gain therefore reflects our new agent design and training algorithm FA-DAPO rather than any difference in backbone scale or architecture. We will make this explicit in the caption in the revision.

AuthorsOpenReview ↗

Response to Reviewer 3phy [Part 2]

W4: Significance compared to DeepEyes and VisionLeaf. We thank the reviewer for sharing these two references. We discuss DeepEyes throughout the paper and use it as one of our main baselines; it also appears in Table 1. We thank the reviewer for pointing out VisionLeaf. We will add a discussion of VisionLeaf in the related work section.

VisionLeaf designs an entropy-guided, tree-structured rollout strategy for think-with-image agents. The key idea is to organize rollouts as a tree with shared prefixes rather than independent root-to-leaf chains, and to concentrate exploration at the leaf node with the highest uncertainty by cloning its prefix and sampling new one-turn continuations. This is an interesting training-time technique for improving RL sample efficiency through principled exploration.

There are, however, two important differences from AdaTurn. First, VisionLeaf is designed for a fixed training horizon (five turns) and does not model or train for a variable turn budget. It has no budget-conditioning mechanism, no forced-answer step at the budget boundary, and no treatment of the over-budget case. Its efficiency reward discourages unnecessary tool calls, but this is not equivalent to teaching the policy how to behave when the budget is strictly smaller than the rollout it would naturally prefer. The problem AdaTurn addresses, catastrophic truncation under a deployment-time budget that the model was not trained for, is therefore orthogonal to VisionLeaf's contribution and not solved by it. Second, even with no budget restriction, AdaTurn compares favorably with VisionLeaf on the shared benchmarks, as shown below.

ModelV*MME-RealWorldHR-Bench 4KHR-Bench 8K
DeepEyes (6 turns)83.364.073.269.5
Mini-o3 (32 turns)88.265.577.573.3
VisionLeaf84.854.076.672.1
Ours (32 turns)86.966.377.573.4

AdaTurn outperforms VisionLeaf on all four shared benchmarks, with a particularly large gap on MME-RealWorld (+12.3%). AdaTurn therefore both outperforms VisionLeaf on standard benchmarks and addresses the deployment-time budget problem that VisionLeaf does not.

Thanks for the response. I confirm that I maintain my original rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.