acceptodds
Under review as a conference paper at ICLR 2027

SAVE: Sufficiency-Aware Visual Evidence Acquisition for Multimodal Reasoning

Abstract

Visual tools allow multimodal large language models to acquire additional visual evidence during reasoning, but effective tool use requires judging whether the available evidence is sufficient. Existing models may miss useful inspection or continue interacting after sufficient evidence has already been obtained. We therefore formulate adaptive visual tool use as an evidence sufficiency problem with two decisions: whether to acquire additional evidence and when to stop. To learn these decisions from complete-trajectory feedback, we introduce SAVE, a Sufficiency-aware framework for Adaptive Visual Evidence acquisition. For each question, SAVE samples multiple on-policy trajectories and constructs a hierarchical reward at two comparison scopes. Across direct-answer and tool-assisted strategies, it strengthens successful trajectories when observed success occurs in only one strategy; within successful tool-assisted trajectories, it favors the minimum observed successful acquisition depth. This design learns invocation and stopping preferences without estimating the marginal utility of individual tool calls or imposing a uniform per-call cost. Across three vision-language backbones and five multimodal reasoning benchmarks, SAVE improves Avg@3 over supervised fine-tuning by 8.33 percentage points on average while reducing the macro-average number of calls per tool-using trajectory by 8.6%.

Primary area:
reinforcement learning
Keywords:
Multimodal Reasoning, Visual Tool Use, Evidence Sufficiency, Reinforcement Learning, Adaptive Evidence Acquisition

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_u4DJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The authors propose the SAVE framework, which samples multiple on-policy trajectories per question, partitions them into direct-answer and tool-assisted groups, and constructs a hierarchical trajectory reward comprising three binary components: task success, a cross-strategy contrastive signal (activated only when one strategy succeeds), and an intra-tool minimum-depth preference (favoring the shallowest successful tool-assisted trajectory). The policy is optimized via REINFORCE Leave-One-Out (RLOO). The training pipeline includes difficulty-aware RL instance selection from ChartQA and SFT initialization based on a teacher model (Qwen3.5-397B). The primary claim is that across all 15 backbone-benchmark pairs, SAVE improves Avg@3 by 8.33 percentage points on average over SFT, while reducing the macro-average number of tool calls per tool-using trajectory by 8.6%.

Critical strengths
  1. Separating the invocation decision (cross-strategy comparison) from the stopping decision (intra-tool minimum-depth preference) constitutes the core conceptual contribution of the paper. The paper clearly demonstrates why a unified per-call penalty invariably favors direct trajectories over tool-assisted ones, providing a principled motivation for SAVE's alternative. The binary non-redundancy metric (Equation 11) elegantly avoids the need to tune a per-call cost coefficient. The ablation study (Table 3) confirms that both components are necessary. The equal-weight design requires no reward hyperparameter tuning, which is favorable for practical adoption.
  2. The paper reports paired bootstrap confidence intervals for all 15 backbone-benchmark comparisons, honestly disclosing that 5 of these intervals contain zero. Controlled tool-access sensitivity analysis, one-step stopping diagnostics, and training-time reward diagnostics provide multiple complementary perspectives. This level of statistical rigor and honest reporting exceeds the norm in the visual tool-use RL literature.
Critical weaknesses
  1. The applicability of the method is limited. SAVE is evaluated exclusively with local cropping operations applied to static images, and all training data comes from ChartQA. Four of the six evaluation benchmarks involve chart or document understanding (CharXiv, ChartQAPro, HR-Bench-4K/8K), which are closely related to the training domain. The paper positions itself as a general framework for "adaptive visual evidence acquisition," but this overstates what is actually demonstrated.
  2. A controlled comparison with the most closely related concurrent methods is missing. Beacon (Wang et al., 2026c) conditions its adaptive reward on whether the rollout group contains a correct tool-free answer, addressing nearly the same problem as SAVE's cross-strategy contrast. It is discussed in related work but absent from Table 2. BEE and AdaTooler-V are similarly cited but not experimentally compared. A controlled experiment training competitive alternative rewards on the same SFT-initialized model and the same ChartQA data would more convincingly isolate the contribution of SAVE's hierarchical reward design.
Questions for authors
  1. How does SAVE compare experimentally with Beacon (Wang et al., 2026c), which also conditions rewards on observed strategy success within rollout groups?
  2. Training uses only ChartQA, yet the largest gains appear on HR-Bench-8K (+27.97 pp for Qwen2.5-VL) and RealWorldQA (+24.26 pp), which are far outside the chart domain. What explains this strong positive transfer?
  3. Given that 56.52% of RL training instances are "hard" (Table 5, n+(xi)=0), and all trajectories receive zero reward when no strategy succeeds, does this mean over half of training batches contribute no gradient signal from the hierarchical components? How does this affect learning efficiency?
5 comments
AuthorsOpenReview ↗

On the Applicability of SAVE (W1) & On Comparisons with Related Methods (W2 / Q1)

We sincerely thank the reviewer for the constructive comments. We have updated the latest version of the manuscript to clarify the scope of SAVE and added comparisons with related methods, cross-domain transfer analyses, and training-time diagnostics.

1. On the Applicability of SAVE (W1)

We have revised the manuscript to clarify that SAVE focuses on adaptive visual evidence acquisition through local cropping for static-image question answering. Our experiments use ChartQA for training and evaluate the method on six image-question answering benchmarks. We have refined the positioning of SAVE to align with the evaluated setting.

2. On Comparisons with Related Methods (W2 / Q1)

We added experimental comparisons with Beacon, BEE, and AdaTooler-V in Table 2. For Beacon, we used its officially released Qwen3-VL-8B-Instruct checkpoint. For BEE and AdaTooler-V, we implemented the corresponding methods following their original papers. All results were obtained under a unified visual cropping interface and evaluation protocol.

Table 2. Comparison with BEE, AdaTooler-V, and Beacon. Acc / Rate / Calls represent Avg@3 (%), tool-use rate (%), and Calls/Used, respectively.

BackboneModelCharXivHallusionBenchRealWorldQAHR-Bench-8KHR-Bench-4KChartQAPro
Qwen2.5-VL-7BBEE-7B38.21 / 76.31 / 1.87258.92 / 74.12 / 1.72364.93 / 65.47 / 1.12254.88 / 70.21 / 1.28360.23 / 71.98 / 1.78330.90 / 80.22 / 1.293
Qwen2.5-VL-7BAdaTooler-V-7B33.75 / 32.22 / 1.37562.32 / 34.31 / 1.49260.99 / 29.30 / 1.89651.21 / 35.32 / 1.90460.46 / 40.35 / 1.56532.89 / 38.21 / 1.443
Qwen2.5-VL-7BSAVE-7B45.51 / 30.93 / 1.17072.53 / 22.40 / 1.12065.18 / 48.24 / 1.29059.26 / 20.25 / 1.60967.00 / 16.67 / 1.56034.00 / 31.04 / 1.281
Qwen3-VL-8B-InstructBeacon42.72 / 12.15 / 1.14769.93 / 17.41 / 1.26957.25 / 19.31 / 1.18364.92 / 13.92 / 1.16169.15 / 10.27 / 1.23040.93 / 12.39 / 1.231
Qwen3-VL-8B-InstructSAVE-8B43.17 / 51.17 / 1.22071.92 / 21.91 / 1.14063.13 / 51.20 / 1.12069.08 / 55.67 / 1.09072.50 / 58.21 / 1.21042.78 / 44.30 / 1.160

Although Beacon and SAVE both leverage observed strategy success within rollout groups, their reward preferences differ. When both Direct and Tool-assisted strategies succeed, Beacon downweights successful tool-assisted trajectories, whereas SAVE assigns equal rewards to successful Direct trajectories and minimum-depth successful Tool-assisted trajectories, while assigning lower rewards to longer successful Tool-assisted trajectories. This explicitly encourages adaptive stopping within the tool-assisted strategy.

As shown in Table 2, under the same Qwen3-VL-8B-Instruct backbone and cropping interface, SAVE outperforms Beacon on all six benchmarks, achieving a macro-average Avg@3 of 60.43% versus 57.48%, an improvement of 2.95 percentage points. SAVE also outperforms BEE and AdaTooler-V on all six benchmarks in the Qwen2.5-VL-7B comparison group.

AuthorsOpenReview ↗

On Strong Cross-Domain Transfer (Q2)

3. On Strong Cross-Domain Transfer (Q2)

We attribute the positive transfer primarily to improved visual evidence acquisition behavior. Although training uses ChartQA, the learned behaviors of when to invoke tools, how to use local visual observations, and when to stop can transfer to other visual question-answering tasks.

Table R1. Cross-domain performance and tool-use behavior of Qwen2.5-VL-7B. Acc denotes Avg@3 (%), Rate denotes tool-use rate (%), and Calls denotes Calls/Used. Improvements are reported in percentage points.

BenchmarkMethodAcc (%)Rate (%)Calls
HR-Bench-8KSFT31.298.381.776
HR-Bench-8KSAVE59.2620.251.609
HR-Bench-8KImprovement+27.97+11.87-0.167
RealWorldQASFT40.9231.981.546
RealWorldQASAVE65.1848.241.290
RealWorldQAImprovement+24.26+16.26-0.256

As shown in Table R1, both benchmarks exhibit substantial accuracy improvements, accompanied by higher tool-use rates and lower Calls/Used, indicating more frequent evidence acquisition with shallower tool-assisted trajectories.

We further conducted controlled tool-access experiments under matched evaluation settings.

Table R2. Controlled tool-access sensitivity analysis on Qwen2.5-VL-7B (Appendix G.2.1, Table 12). Gain denotes Auto-tool minus No-tool Avg@3 in percentage points, with paired 95% bootstrap confidence intervals.

BenchmarkCheckpointNo-tool Avg@3 (%)Auto-tool Avg@3 (%)Gain (pp) [95% CI]
HallusionBenchSFT51.40864.884+13.476 [11.479, 15.437]
HallusionBenchSAVE57.07767.201+10.125 [8.306, 11.907]
HR-Bench-8KSFT26.58329.417+2.833 [0.583, 5.042]
HR-Bench-8KSAVE28.33336.958+8.625 [6.542, 10.708]

Table R2 shows that enabling autonomous tool access improves SAVE's performance on HR-Bench-8K by 8.625 percentage points, providing additional evidence that local visual evidence acquisition is beneficial on this benchmark.

We also examined whether an additional crop is beneficial after the model decides to stop or continue.

Table R3. One-step Stop/Continue diagnostic on HR-Bench-8K (Appendix G.2.2, Table 13). Stop evaluates immediate answering, while Continue-one forces one additional model-selected crop. The difference is Continue-one minus Stop.

CheckpointDecision GroupStop Avg@3 (%)Continue-one Avg@3 (%)Difference (pp) [95% CI]
SFTStop51.85244.444-7.407 [-12.346, -2.469]
SFTContinue46.03246.561+0.529 [-7.407, 8.466]
SAVEStop45.52837.398-8.130 [-15.447, -1.626]
SAVEContinue43.47847.826+4.348 [-5.797, 14.493]

For histories where SAVE decides to stop, an additional crop reduces Avg@3 by 8.130 percentage points. Together, these findings suggest that the evidence acquisition and stopping behaviors learned from ChartQA transfer to other visual tasks, improving the model's ability to use local visual information effectively.

AuthorsOpenReview ↗

On Hard Training Instances and Learning Efficiency (Q3)

The 56.52% Hard instances are defined by the teacher model failing in all three sampled trajectories. This teacher-based difficulty classification is distinct from the actor's success during online reinforcement learning.

Table R4. Difficulty distribution of the curated RL training dataset (Appendix D, Table 6). Difficulty is determined by the number of correct trajectories among three teacher-generated samples.

DifficultyCriterionInstancesProportion (%)
MediumAt least one but fewer than three correct teacher trajectories3,39543.48
HardNo correct teacher trajectories4,41456.52
Total RL dataMedium + Hard7,809100.00

Appendix G.4 (Figure 8) provides training-time evidence of active hierarchical reward signals**. **Figures 8(a) and 8(d) show that rollout groups with successful strategies and cross-strategy contrasts occur throughout training, providing opportunities for strategy-level learning. Figure 8(b) shows the continued presence of successful tool trajectories at the minimum observed depth, while Figure 8(c) shows an overall reduction in redundant correct tool use, consistent with the intended minimum-depth preference.

Moreover, Figure 3 demonstrates improvements in validation accuracy and tool-use behavior, indicating effective policy learning during training.

For groups in which all trajectories fail or eligible rewards are identical, Appendix E.1.1 explains that rejection sampling filters these groups, and additional groups are sampled to construct optimization batches with informative reward differences. The associated efficiency cost mainly comes from additional rollout generation and answer evaluation.

We thank the reviewer again for these valuable suggestions, which have helped us strengthen the experimental analysis and clarify the scope of our method.

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

Addressing the difficulty Vision-Language Models (VLMs) face in determining whether to invoke visual tools and when to stop during reasoning, this paper formulates the challenge as an evidence sufficiency decision problem and proposes SAVE, an adaptive framework for visual evidence acquisition. The approach first equips the model with both direct-answering and tool-assisted strategies through data curation and supervised fine-tuning (SFT). In the subsequent on-policy reinforcement learning stage, it constructs a hierarchical trajectory reward from multiple rollouts of the same question. Experiments show that across three mainstream VLM backbones and multiple multimodal reasoning benchmarks, SAVE improves accuracy by an average of 8.33 percentage points while reducing average tool-call steps by 8.6%, effectively balancing reasoning accuracy and interaction efficiency.

Critical strengths
  1. The paper addresses a practical problem in visual tool use: deciding when to use tools and when to stop. The hierarchical reward design is simple and avoids the need for a fixed penalty for every tool call.

  2. The experiments cover multiple VLM backbones and benchmarks. The improvements over SFT are consistent, and the ablation studies help show the contribution of different components.

Critical weaknesses
  1. Limited generalizability. The experiments mainly use 7B–8B VLMs (mostly out of date), and the training data comes from ChartQA (some problems may be too simple). It remains unclear whether the method generalizes to newer, larger models and more diverse tasks.

  2. Limited tool diversity and shallow interactions. The only visual tool is crop, and the average number of calls per tool-using trajectory is mostly between 1.1 and 1.6. This makes stopping largely a choice between one or two calls, rather than a real test of long-horizon tool use. It is unclear whether SAVE can handle more complex scenarios involving 4–8 sequential calls to different tools.

Questions for authors
  1. The current implementation only supports crop, and the average number of calls per tool-using trajectory is relatively low. Can the authors provide evidence that SAVE also works with multiple types of tools and longer interaction sequences?
  2. What is the main reason for the relatively modest gains of Qwen3-VL-8B-Instruct? Is it because the base model is already capable of direct inference and adaptation, or is it because reinforcement learning exploration on ChartQA (a single chart domain) has hit a bottleneck? Has the Qwen3.5 series model been used as a backbone for related experiments?
  3. If a question can be answered correctly entirely by direct perception, then tool calls are not inherently necessary, and the intuition is that the model should prefer the direct strategy of zero calls; SAVE, on the other hand, treats them as ties, giving the same maximum reward. What is the theoretical motivation of the authors for such a design? Some of the stronger models were able to successfully answer most questions without tools. Did this result in unnecessary tool calls after training?
9 comments
AuthorsOpenReview ↗

On Model Scale and Task Generalizability (W1)

We sincerely thank the reviewer for the constructive comments. We have updated the latest version of the manuscript and added experiments with the Qwen3.5 series, MathVista/MathVerse multi-tool evaluations, and further analyses of backbone performance and reward design.

1. On Model Scale and Task Generalizability (W1)

In the revised manuscript, we included experiments with the newer Qwen3.5-2B model and extended our evaluation to more diverse visual reasoning tasks. We evaluated SAVE on Qwen3.5-2B, as shown in Table R1.

Table R1. Performance comparison between SFT and SAVE on Qwen3.5-2B. This table evaluates the applicability of SAVE to a newer model backbone. Acc denotes Avg@3 (%), Rate denotes tool-use rate (%), and Calls denotes Calls/Used.

BenchmarkSFT AccSFT RateSFT CallsSAVE AccSAVE RateSAVE Calls
CharXiv41.2940.311.82042.9115.491.250
HallusionBench70.0917.971.54470.904.291.300
RealWorldQA58.5455.271.54959.3829.391.190
HR-Bench-8K36.4764.211.98840.5435.901.493
ChartQAPro39.8935.731.80041.1510.411.198

SAVE outperforms SFT on all five benchmarks while reducing both tool-use rates and the average number of calls per tool-using trajectory. These results further support the effectiveness of SAVE across different model architectures and parameter scales.

AuthorsOpenReview ↗

On Tool Diversity and Longer Interaction Sequences (W2 / Q1)

(1) Motivation for using a single Crop tool in the main experiments.

SAVE focuses on two coupled decisions: when to acquire additional visual evidence and when to stop acquiring evidence. We therefore use a fixed Crop operation in the main experiments to control the additional variables introduced by tool selection and tool composition, allowing us to examine invocation and stopping decisions more directly.

Furthermore, the motivation analysis in Figure 1(b) examines first-time corrections and negative regressions across successive visual inspections, illustrating how the benefits of additional evidence vary with interaction depth. This motivates our initial study of adaptive invocation and stopping under a controlled tool interface.

(2) Additional experiments with multiple tools and longer interactions.

To examine SAVE under an expanded visual tool space, we fixed the trained SAVE-Qwen3-VL-8B-Instruct checkpoint and conducted additional experiments on MathVista testmini and the Text Dominant subset of MathVerse testmini.

For each question, we constructed three image conditions: the original image, a 90° clockwise rotation, and contrast reduced to 0.25. We evaluated the same checkpoint under Crop-only and Expanded Tools settings. The expanded tool set contains five operations, as detailed below.

Table R2. Definitions of the expanded visual tool set. This table describes the tool types, operations, and parameter ranges used in the multi-tool evaluation.

ToolFunctionParameters
CropExtract a specified region from the current canvasNormalized bounding box [0,1000]
RotateRotate the full canvas or current view90°, 180°, 270°
ZoomCrop and enlarge the central region of the current view1.5×, 2×, 3×, 4×
SharpenAdjust the sharpness of the current viewSharpening factor 1–3
ContrastAdjust the contrast of the current viewContrast factor 0.5–2

Zoom, Sharpen, and Contrast can be composed sequentially. Zoom does not employ super-resolution or generate missing visual details.

Table R3. Performance of SAVE under different visual tool settings. This table evaluates the effects of tool-set expansion on accuracy, tool-use frequency, interaction depth, and multi-type tool usage.

Tool SettingAcc (%)Rate (%)Calls/UsedCalls/AllNew Tool Usage (%)Multi-type Usage (%)
No-tool45.210.000.000.000.000.00
Crop-only51.6778.334.963.880.000.00
Expanded Tools58.3371.674.843.4726.678.33

Calls/Used denotes the average number of calls among tool-using trajectories, while Calls/All denotes the average number of calls across all trajectories. New Tool Usage measures the percentage of trajectories invoking at least one newly introduced tool, and Multi-type Usage measures the percentage invoking at least two different tool types.

The results demonstrate the following:

· Accuracy improvement: Expanded Tools achieves 58.33% accuracy, outperforming Crop-only by 6.66 percentage points and No-tool by 13.12 percentage points.

· Longer interactions: Tool-using trajectories under Expanded Tools contain 4.84 calls on average, demonstrating SAVE's behavior over longer interaction sequences.

· Multi-tool capability: 26.67% of trajectories invoke at least one newly introduced tool, and 8.33% invoke at least two different tool types.

· Tool-use efficiency: Calls/All decreases from 3.88 to 3.47 while accuracy improves.

These results provide additional evidence that the visual evidence acquisition policy learned by SAVE can be extended to multiple visual operations while maintaining effective invocation and stopping behavior over longer interaction sequences.

AuthorsOpenReview ↗

On the Modest Improvements of Qwen3-VL-8B-Instruct (Q2)

We appreciate the reviewer's question regarding the performance differences across model backbones. We believe the relatively modest improvements of Qwen3-VL-8B-Instruct are associated with its stronger SFT initialization and the remaining room for improvement.

Table 1 reports the performance of Qwen3-VL-8B-Instruct before and after SAVE reinforcement learning.

Table R4. SFT–SAVE accuracy comparison on Qwen3-VL-8B-Instruct. This table analyzes the magnitude and distribution of reinforcement learning improvements across benchmarks.

BenchmarkSFT Acc (%)SAVE Acc (%)Improvement (pp)
CharXiv40.7743.17+2.40
HallusionBench71.1471.92+0.78
RealWorldQA60.8363.13+2.30
HR-Bench-8K67.1769.08+1.91
ChartQAPro41.6642.78+1.12
Macro Average56.3158.02+1.70

SAVE achieves positive improvements on all five benchmarks. The Qwen3-VL-8B-Instruct SFT checkpoint already exhibits relatively strong visual reasoning and tool-use capabilities, leaving less room for further gains than models starting from weaker initialization.

We added experiments on Qwen3-VL-2B with different rollout group sizes (Figure 5 and Table 4), using (N=4,8,16,32).

Table R5. Complete results for different rollout group sizes on Qwen3-VL-2B (Table 4 in the revised manuscript). This table evaluates SAVE's effectiveness on a smaller backbone and its sensitivity to rollout group size. Each cell reports Acc / Rate / Calls, corresponding to Avg@3 (%), tool-use rate (%), and Calls/Used.

VariantCharXivHallusionBenchRealWorldQAHR-Bench-8KHR-Bench-4KChartQAPro
SFT8.41 / 64.86 / 1.3157.49 / 87.24 / 1.84711.81 / 90.34 / 1.43610.08 / 89.08 / 1.47413.11 / 85.32 / 1.4724.90 / 77.34 / 1.670
SAVE (N=4N=4)9.78 / 39.38 / 1.4137.69 / 38.62 / 1.78512.95 / 32.34 / 1.63510.92 / 32.23 / 1.61317.17 / 36.62 / 1.6125.75 / 31.30 / 1.648
SAVE (N=8N=8)10.51 / 31.09 / 1.41211.54 / 33.46 / 1.83616.59 / 38.64 / 1.51518.69 / 37.84 / 1.50019.32 / 35.98 / 1.5247.30 / 35.30 / 1.726
SAVE (N=16N=16)11.22 / 30.15 / 1.47612.22 / 30.66 / 1.79317.90 / 37.66 / 1.42017.67 / 36.55 / 1.51720.00 / 36.89 / 1.4658.84 / 30.76 / 1.680
SAVE (N=32N=32)10.09 / 26.83 / 1.46011.51 / 27.01 / 1.83916.50 / 29.82 / 1.49717.22 / 28.44 / 1.44220.29 / 29.33 / 1.4697.97 / 27.66 / 1.771
Cost penalty (N=16N=16)7.75 / — / —8.14 / — / —15.83 / — / —13.15 / — / —15.01 / — / —6.28 / — / —

On Qwen3-VL-2B, SAVE with (N=16) improves macro-average Avg@3 from 9.30% to 14.64%, a gain of 5.34 percentage points across six benchmarks. The (N=8,16,32) settings also achieve broadly comparable overall performance, further supporting SAVE's effectiveness on this backbone.

Together with the Qwen3.5-2B results in Table R1, these experiments show backbone-dependent differences in the magnitude of improvement. We consider initial model capability and remaining optimization headroom important factors contributing to these differences.

In addition, Figure 3 provides training-time evidence through validation accuracy, tool-use rates, conditional tool-call depth, and the fraction of correct tool-free trajectories across backbones.

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

This paper formulates adaptive visual tool use as two coupled decisions: whether additional evidence should be acquired and when acquisition should stop. SAVE samples multiple on-policy trajectories, partitions them into direct and tool-assisted strategies, rewards the uniquely successful strategy when only one demonstrates success, and favors the minimum observed successful depth among tool-assisted trajectories. The resulting hierarchical reward is optimized using RLOO. Across three 7B–8B backbones and several multimodal benchmarks, SAVE improves accuracy over supervised initialization while generally increasing invocation frequency but reducing calls per tool-using trajectory.

Critical strengths
  1. The distinction between invocation and stopping is well motivated and practically important. SAVE offers a simple trajectory-level construction that requires neither call-level necessity labels nor a separately learned utility model.

  2. The empirical study is unusually extensive: three backbones, multiple benchmarks, paired bootstrap intervals for the main SFT–SAVE comparison, component ablations, controlled tool-access and stopping diagnostics, and detailed prompts and implementation information. The large gains for Qwen2.5-VL and InternVL are substantial.

Critical weaknesses
  1. The central “sufficiency” signals are confounded by stochastic sample counts. Equation 7 defines strategy success as the maximum correctness over each strategy’s sampled trajectories. Table 13 shows roughly 10 direct versus 5 tool-assisted trajectories per group; even with equal per-trajectory success probabilities, the more frequently sampled strategy is therefore much more likely to exhibit at least one success. Likewise, the minimum observed successful depth in Equation 9 is an order statistic that depends on how many trajectories are sampled at each depth and may select a lucky short success. These quantities need not represent strategy necessity or evidence sufficiency.

  2. The ablations do not cleanly identify the effects of the reward components. As acknowledged in Appendices B.3 and E.1.1, variants have different raw reward scales and reward-dependent rejection sampling can retain different task groups. Thus, their optimization data and effective gradient scales differ in addition to the component being removed. The paper also lacks a matched, tuned uniform-call-cost baseline and repeated training runs.

Questions for authors
  1. Do the gains persist when direct and tool-assisted strategies receive equal rollout counts, or when Equation 7 is replaced by a count-adjusted estimate of per-strategy success probability? Please also quantify sensitivity to rollout group size. This is necessary to distinguish evidence sufficiency from unequal sampling frequency.

  2. Can the authors rerun the key reward ablations with matched reward scale and a fixed common set of retained task groups, and compare against a tuned reward of the form correctness - λ × calls under the same data and compute? These controls would materially affect my confidence in attributing the gains to the proposed hierarchy.

3 comments
AuthorsOpenReview ↗

Unequal Sampling and the Sufficiency Signals (W1 / Q1) & Reward Ablations and Uniform Call-Cost Baseline (W2 / Q2)

We sincerely thank the reviewer for the constructive comments. We have updated the latest version of the manuscript and added supplementary experiments and analyses to address the concerns regarding rollout sampling, reward design, and the uniform call-cost baseline.

1. W1 / Q1

(1) On unequal sampling between Direct and Tool-assisted strategies

SAVE adopts natural on-policy sampling without a predefined allocation between Direct and Tool-assisted strategies. Specifically, the actor generates (N=16) trajectories for each question and partitions them according to whether a visual tool is invoked.

To clarify the actual sampling distribution, we showed rollout composition statistics in Appendix G.5 (Figure 9 and Table 14). The average numbers of eligible Direct and Tool-assisted trajectories are:

BackboneDirectTool-assisted
Qwen3-VL-8B9.965.97
Qwen2.5-VL-7B10.385.48
InternVL3.5-8B10.794.78

Equation 7 uses the observed success of each strategy within the sampled rollout group to construct the cross-strategy preference signal.

(2) On rollout group size sensitivity

We added experiments with (N=4,8,16,) and (32) rollouts per question (Figure 5 and Table 4). The results show that (N=4) performs relatively poorly, whereas (N=8,16,) and (32) achieve broadly comparable performance across the six benchmarks.

(3) On the minimum observed successful depth

Equation 9 defines the minimum successful tool depth within the current sampled set, providing an empirical reference for learning when to stop acquiring additional evidence. As clarified in Section 3.3 and Appendix B.3, a shorter successful trajectory is not necessarily a prefix of a longer one.

Additionally, we added a one-step Stop/Continue diagnostic in Appendix G.2.2 (Table 13). For histories assigned to the Stop group by SAVE, forcing one additional tool call reduces Avg@3 by 8.130 percentage points (95% CI: ([-15.447,-1.626])). This result provides further evidence supporting the stopping behavior learned by SAVE.

2. W2 / Q2

(1) On reward scale differences

We have clarified the reward processing details in Appendix B.3. Different reward variants have different raw reward ranges. The RLOO leave-one-out baseline removes additive shifts while preserving multiplicative differences in reward scale. We further discuss how these scale differences may influence optimization dynamics.

(2) On reward-dependent task-group filtering

Appendix E.1.1 clarifies that all reward variants use the same candidate data and rejection-sampling rule. Since rejection sampling depends on the assigned rewards, different variants may retain different task groups.

We also added training-time reward diagnostics in Appendix G.4 (Figure 8), comparing the success patterns, minimum successful tool depth, redundant correct tool rate, and cross-strategy contrast rate across reward variants.

(3) On the uniform call-cost baseline

We added a uniform call-cost baseline in the main paper (Figure 5 and Table 4), using:

r=z−0.1cr = z - 0.1c

where (z) denotes answer correctness and (c) denotes the number of tool calls.

SAVE outperforms this baseline on all six benchmarks. Furthermore, Figure 5 shows that the tool-use rate of the cost-penalty baseline gradually approaches zero, whereas SAVE retains non-zero tool usage.

AuthorsOpenReview ↗

Additional experiments

Table 4. Comparison of different rollout group sizes and the uniform call-cost baseline on Qwen3-VL-2B. Acc denotes Avg@3 (%), Rate denotes tool-use rate (%), and Calls denotes Calls/Used. The cost-penalty baseline uses r=z−0.1cr=z-0.1c.

VariantCharXiv AccCharXiv RateCharXiv CallsHallusionBench AccHallusionBench RateHallusionBench CallsRealWorldQA AccRealWorldQA RateRealWorldQA Calls
SFT8.4164.861.3157.4987.241.84711.8190.341.436
SAVE (N=4N=4)9.7839.381.4137.6938.621.78512.9532.341.635
SAVE (N=8N=8)10.5131.091.41211.5433.461.83616.5938.641.515
SAVE (N=16N=16)11.2230.151.47612.2230.661.79317.9037.661.420
SAVE (N=32N=32)10.0926.831.46011.5127.011.83916.5029.821.497
Cost penalty (N=16N=16)7.75——8.14——15.83——
VariantHR-Bench-8K AccHR-Bench-8K RateHR-Bench-8K CallsHR-Bench-4K AccHR-Bench-4K RateHR-Bench-4K CallsChartQAPro AccChartQAPro RateChartQAPro Calls
SFT10.0889.081.47413.1185.321.4724.9077.341.670
SAVE (N=4N=4)10.9232.231.61317.1736.621.6125.7531.301.648
SAVE (N=8N=8)18.6937.841.50019.3235.981.5247.3035.301.726
SAVE (N=16N=16)17.6736.551.51720.0036.891.4658.8430.761.680
SAVE (N=32N=32)17.2228.441.44220.2929.331.4697.9727.661.771
Cost penalty (N=16N=16)13.15——15.01——6.28——

We sincerely hope that these additional experiments, analyses, and clarifications have adequately addressed your concerns.

Dear Reviewer mtcg,

As the discussion period is drawing to a close in about one day, we wanted to kindly follow up on our response and additional experiments. We sincerely thank you for raising these important questions, which have helped us strengthen our paper.

For W1/Q1, we added rollout group-size experiments (N=4,8,16,32N=4,8,16,32) and reported the actual sampling distribution. We also conducted a Stop/Continue experiment, showing that an additional tool call after SAVE decides to stop reduces Avg@3 by 8.13 percentage points, providing further evidence for the learned stopping behavior.

For W2/Q2, we clarified the effects of reward scaling and rejection sampling, and added training-time diagnostics. We also evaluated a uniform call-cost baseline (r=z−0.1cr = z - 0.1c). SAVE outperforms this baseline on all six benchmarks, while the baseline's tool-use rate gradually approaches zero.

We hope these additional experiments and clarifications have addressed your concerns. Please let us know if there are any remaining points we can clarify before the discussion window closes. Thank you again for your time and constructive feedback!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.