acceptodds
Under review as a conference paper at ICLR 2027

VisEvolver: Hybrid Parameter-and-Memory Self-Evolving for Multimodal Agents

Abstract

Multimodal large language models are evolving from static visual reasoners into agents capable of active perception, tool use, and sustained interaction, yet how such agents can continuously acquire capabilities from their growing experience remains underexplored. Existing approaches largely follow two paradigms: external memory enables rapid and reversible reuse of past experience but remains dependent on inference-time context, while parametric adaptation can turn experience into intrinsic capability but risks internalizing instance-specific facts and incidental behaviors. We argue that a self-evolving multimodal agent should instead allow experience to progressively migrate from external memory into internal capability according to its reliability and transferability, which raises two fundamental challenges: **Causal Experience Attribution**, identifying which parts of multimodal experience truly contribute to successful decisions, and **Selective Experience Internalization**, determining which validated experiences should be consolidated into model parameters. To address these challenges, we introduce **VisEvolver**, a dual-timescale framework that organizes multimodal experience into a progressive lifecycle of **Experience → Evidence → Skill → Program → Weight**. VisEvolver extracts transferable experience through visual counterfactual attribution and success–failure trajectory comparison, and selectively internalizes repeatedly validated skills via on-policy privileged self-distillation while dynamically coordinating external memory and internal skill experts. Experiments show that VisEvolver not only exploits newly acquired experience for immediate adaptation, but also turns stable and transferable strategies into persistent model capabilities, providing a unified framework for multimodal agents to move from remembering experience toward learning from it.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
multimodal agents, self-evolving agents, experience learning, continual learning, multimodal memory, skill internalization

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_eQJzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

VisEvolver evaluates multimodal experiences through visual interventions and employs a "hindsight supervision" mechanism to selectively direct these experiences toward external memory, parameter updates, or both.

Critical strengths

Rather than simply layering "memory" and "fine-tuning," a more methodologically sound approach involves filtering out unsubstantiated experiences through visual comparison and then allocating these experiences either to external memory or to model parameters.

Critical weaknesses

The dataset is incomplete. The primary results are based on a "challenge subset" filtered using a Base model, while the tests for charts and infographics consist of only 200 tasks each, which is far from the full official test set. Furthermore, the main tables lack confidence intervals or significance tests. Consequently, it is difficult to determine whether some of the observed marginal gains are robust or if they are contingent upon the specific filtering method employed.

Questions for authors
  1. How exactly are the five indicators calculated and replicated? What are the specific threshold values, and are the results sensitive to these thresholds?
  2. To what extent does "cross-task support" actually extend? Does it refer to repeated verification across different problems within the same dataset, or to transfer across different datasets and task types? How can one determine whether multiple tasks are supported by the same abstract principle, rather than merely by similar instances?
  3. What exactly is the teacher model? How much additional computational overhead does it introduce? Does the model still maintain its advantages when compared under the same level of computational cost? If the authors provide the missing experimental results mentioned above, I will consider raising the score.
9 comments
AuthorsOpenReview ↗

Response to Reviewer eQJz: Routing Sensitivity, Cross-Task Transfer, and Budget-Matched Comparisons

Thank you for the specific suggestions on evaluation coverage, reproducibility, transfer and cost. We present the new measurements and explain how they address each question below.

Q1: Do the results generalize beyond the Base-screened subset, with confidence intervals and reliable small gains?

A1: Thank you for asking whether the findings extend beyond the submitted subsets and remain stable across runs. We would first like to clarify that the QA questions are randomly sampled into fixed partitions shared by all methods. The submitted sizes and evaluation sources are:

Table eQJz-1. Data partitions used in the submitted experiments

DatasetTrainDevTestSampling / evaluation unit
ChartQA10030200Random, fixed lists; shared by all methods
InfographicVQA10030200Random, fixed lists; evaluation from public validation
VisualToolBench8020214Fixed project partitions
EB-ALFRED10030300Fixed task/scene lists; includes repeated trials

Counts are from Appendix A.7, Table 4. InfographicVQA uses labeled public validation for evaluation.

Following your suggestion, we completed a broader comparison on the complete public QA splits and the full project VTB/EB evaluation lists:

Table eQJz-1a. Base and VisEvolver on the expanded evaluation populations

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
Base86.4077.106.007.50
VisEvolver88.0079.609.7812.69
Gemma-3-12B
Base40.5064.904.239.50
VisEvolver50.3065.307.6417.44

Measured results on ChartQA test (human + augmented, 2,500 questions), InfographicVQA public validation (2,638), the complete native VTB project list (214), and the EB-ALFRED project list (300 trials). Base and VisEvolver use the same population within each column. These populations are identified separately from those of the three-seed table.

VisEvolver outperforms Base on all eight backbone–task settings in this broader evaluation. This directly addresses the concern that its benefit might be specific to the smaller main-paper lists. To examine run-to-run variation separately, we also report the three-seed results:

Table eQJz-2. VisEvolver performance across three seeds

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver: mean ± SD68.38 ± 1.1569.00 ± 1.109.00 ± 3.2010.23 ± 4.10
Mean − strongest Table 1 baseline+1.38+0.62-0.46-1.14
Gemma-3-12B
VisEvolver: mean ± SD47.50 ± 0.7552.00 ± 1.806.07 ± 3.7015.33 ± 3.30
Mean − strongest Table 1 baseline+0.50+9.02+1.84+10.33

VisEvolver uses three seeds over each complete evaluation list. Baselines are the strongest per-backbone, per-task entries in paper Table 1, excluding VisEvolver, Memory-only and Weight-only. Qwen references: 67.00 / 68.38 / 9.46 / 11.37; Gemma: 47.00 / 42.98 / 4.23 / 5.00. Differences are score-point differences against those published values, not paired significance tests.

The two comparisons answer complementary questions: the expanded results test population coverage, while the repeated runs quantify variability. Both make the evidence easier to assess, with especially clear gains for Gemma on several tasks. We will include the exact populations and seed statistics, and distinguish these measurements from paired significance tests of the small single-branch margins.

Q2: How are the five indicators computed, and how sensitive is routing to thresholds and loss weights?

A2: Thank you for asking both how to reproduce the five indicators and whether their settings affect the conclusions. Each indicator examines a different property of the evidence supporting an experience. To make their contribution concrete, we first report the real-task comparison with each criterion removed, together with simple routing alternatives.

Table eQJz-3. Real-task routing comparisons and individual ablations

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver68.8869.009.7010.58
Always-Both67.4067.407.9010.00
Random-matched50.6265.309.208.00
Utility-only67.0065.307.208.00
w/o Support67.0064.226.009.42
w/o Utility68.3864.226.009.42
w/o Stability68.3864.227.908.00
w/o Specificity67.0062.137.909.42
w/o Conflict68.3864.229.208.00
w/o Attribution66.8764.227.909.42
Gemma-3-12B
VisEvolver48.2554.526.0715.33
Always-Both47.5052.656.0712.00
Random-matched42.1246.014.2113.33
Utility-only47.2545.742.2213.33
w/o Support42.1242.983.3912.00
w/o Utility42.7546.013.7412.00
w/o Stability42.7546.013.3912.00
w/o Specificity42.7546.013.3912.00
w/o Conflict42.1245.743.7412.00
w/o Attribution47.5046.013.7412.00

Random-matched draws from the same candidate pool with the same admission count and update budget. “w/o” removes the named criterion or attribution component.

The complete router outperforms each individual ablation across the evaluated settings. In particular, selecting by utility alone is insufficient. In the supplemental measurements, we compute the criteria as follows:

  • Support: distinct supporting episodes and task categories, with defaults of at least 3 and 2, respectively.
  • Utility: the average paired score gain from using the candidate experience, with a threshold of 0.10.
  • Stability: the fraction of paired gains strictly above zero, with a threshold of 0.75; ties and negative effects remain in the denominator.
  • Specificity: the mean of audited instance-binding flags, with a maximum of 0.30.
  • Conflict: the fraction of incompatible actions among records with overlapping applicability conditions, with a maximum of 0.20.

We varied the support threshold over 2/3/4, utility over 0.05/0.10/0.15, stability over 0.60/0.75/0.90, specificity over 0.15/0.30/0.45, and conflict over 0.10/0.20/0.30. We also varied the QA preservation weight:

Table eQJz-4. Sensitivity to the QA preservation weight

Preservation weightChartQA Acc. ↑InfoVQA ANLS ↑
Qwen3.5-9B
λ = 065.0085.06
λ = 0.578.5084.55
λ = 1 (default)83.0079.98
λ = 282.0084.96
Gemma-3-12B
λ = 074.0055.12
λ = 0.577.0055.32
λ = 1 (default)74.0052.31
λ = 274.5052.46

This dedicated sweep uses a fixed 200-question evaluation list per dataset. It is separate from the newly supplied full-evaluation ablation table above. Only the task scores are shown; the 30-question Dev results are omitted for concision.

The threshold sweep gives a useful local result: changing an individual threshold within the tested range did not change the selected sets in the recorded candidate pools. The preservation comparison further shows why the regularizer is included: it substantially improves Qwen ChartQA over adaptation without preservation, while its effect on the other QA settings is more task dependent. These observations support using explicit admission checks and a task-specific preservation configuration. We use equal utility weights α = 0.5 and β = 0.5 in Eq. 7, separately from LoRA’s scaling parameter. We will include these settings together with the operational definitions and sensitivity results.

Q3: Does cross-task support mean repeated source-task success or cross-dataset transfer, and how is abstraction verified?

A3: Thank you for asking how far the learned experience can transfer. The routing criterion checks whether the same abstract procedure is supported by different source tasks and categories. Transfer outside the source benchmark is a further question, which we tested directly following your suggestion.

We froze the parameters and memory learned on each source benchmark and evaluated them on a related target dataset without target-side adaptation. Each target evaluation uses 100 randomly selected questions or episodes.

Table eQJz-5. Transfer without target-dataset adaptation

MethodChartQA → PlotQAInfo → DocVQAVTB → TIR-BenchEB → Navigation
Qwen3.5-9B
Base68.0088.0014.0040.00
VisEvolver74.0092.0022.0020.00
Gemma-3-12B
Base34.0081.0015.0036.00
VisEvolver36.0079.0017.0038.00

Each target dataset uses 100 randomly selected questions or episodes. Source-trained parameters and memory are frozen before target evaluation. Scores retain each target benchmark’s metric.

The strongest transfer pattern appears in chart reasoning and visual tool use, where both backbones improve over Base. The document and navigation results are more task dependent, indicating that the match between the source procedure and target requirements matters. These results provide direct evidence that some learned experience is reusable outside its source benchmark, while also identifying where applicability checks remain important. This motivates retaining explicit conditions and provenance in memory and using specificity and conflict checks before parameter consolidation. We will distinguish this transfer experiment from the within-source support criterion in the revised paper.

Reviewer_ft92OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how a multimodal agent should keep the experience it produces. It argues that external memory and parameter updates should not be treated as alternatives, and that each experience should be given its own destination. The paper proposes VisEvolver. Causal Experience Attribution audits each candidate by task outcome, execution validity and three paired visual signals in Eq. 5. An Experience Router in Eq. 8 reads cross task support, mean utility, utility variance, instance specificity and conflict, and sends each record to process Memory, to LoRA parameters, or to both. Records admitted to the parameters are learned with Student Reachable Hindsight Supervision in Eq. 10 and a Base Preservation KL term in Eq. 11. Experiments use ChartQA, InfographicVQA, VisualToolBench and EB-ALFRED with Qwen3.5-9B and Gemma-3-12B. Table 1 reports the best score in seven of eight settings. Table 2 compares with SFT, GRPO and MemRL. Figure 3 reports the crop interventions and Table 3 reports the component removals.

Critical strengths
  1. The problem framing is useful. The paper asks which experience should stay explicit and which should enter the weights, instead of fixing one destination for the whole system. Table 1 gives some support for this framing. The memory only variant and the weight only variant trade places across tasks. With Qwen3.5-9B the memory only variant reaches 6.31 against 9.51 on VisualToolBench, while it is higher on InfographicVQA with 68.76 against 65.30. This is the main reason I see value in the paper.
  2. The paired visual intervention is a sensible audit. Figure 3 shows that a random region of the same image recovers 7.81 of the 14.84 points on ChartQA. This is an informative result for memory pipelines that keep everything that appears in a correct trajectory, and it can be reused by other work on agent memory.
Critical weaknesses
  1. The evidence for the main claim is weak in size and selection. According to Table 4, the test sets contain 200 ChartQA, 200 InfographicVQA, 214 VisualToolBench and 300 EB-ALFRED tasks, and the ChartQA and InfographicVQA lists are challenge populations screened with Base. Experience is collected from only 80 to 100 training tasks per benchmark. Decoding is deterministic and each number comes from a single run without seeds or confidence intervals. Many margins are small. With Qwen3.5-9B the full system beats the better single branch by 0.50 on ChartQA, 0.24 on InfographicVQA, 0.19 on VisualToolBench and 0.83 on EB-ALFRED. On a 200 task list one item is worth 0.5 points. Absolute scores on VisualToolBench are around 9 out of 100. In Table 2, GRPO reaches 70.92 on InfographicVQA against 69.00 for the full system. Screening the test lists with Base also favors any method that differs from Base. Hence the claim of consistent gains and the claim that the two branches carry additive information are not yet established.
  2. Several reported numbers are inconsistent or not experimental. Table 12 gives a weight score of 68.88 for Qwen3.5-9B on ChartQA, which is the full system value in Table 1, while Table 1 gives 68.38 for the weight only variant. For Gemma-3-12B on InfographicVQA, Table 12 gives 55.57 for Base and 60.70 for the adapted policy, while Table 1 gives 42.95 and 46.01. Figure 3c reports 44.50 and 49.50 on ChartQA, which match no row of Table 1. Figure 4 is described in Appendix G.3 as a synthetic illustration with 40 simulated experiences, so the Router has no direct empirical validation beyond the single Router removal row in Table 3. In addition, Table 12 shows that Base makes only 0.095 to 0.291 tool calls per task, so the inspection habit that the paper says is internalized is rare in the evaluated behavior. These issues affect soundness directly, because the main conclusions rest on these tables.
Questions for authors
  1. Please clarify how the ChartQA and InfographicVQA test lists in Table 4 were screened with Base. Were items selected where Base fails, and was the same list used for every baseline?
  2. Please report the mean and standard deviation over at least three seeds or task orders for Table 1 and Table 3, together with a significance test against the better single branch. This would directly change my assessment of the main claim.
  3. Please explain the mismatch between Table 12, Figure 3c and Table 1. Which runs produce the main table, and why does Appendix G.4 call them historical aggregates?
  4. Please clarify which model acts as the Teacher in Section 3.3 and Eq. 10. If the Teacher is stronger than the backbone, the comparison with memory baselines that use no teacher is not matched.
  5. Please give the values of the thresholds in Eq. 8, of alpha and beta in Eq. 7 and of lambda in Eq. 11. Please also report how many records are routed to Memory only, Parameters only and Both for each task. Appendix K only gives a qualitative description.
  6. Section 4.3 says the blank control keeps the position and size of the region inside the image, but Appendix C.2 says the black condition replaces the crop itself and does not mask the full image. Please clarify which setting was used, since the main text claim that the causal content lies inside the region depends on it.
  7. The best crop only scores in Figure 3 are 14.84 on ChartQA and 7.96 on InfographicVQA, far from the full image setting. On InfographicVQA the text only input scores 3.08, which is above the random region at 2.31. Please explain why attribution signals computed in this low score regime are reliable for deciding what the full image agent should keep. Suggestions for future work
  8. I suggest replacing the synthetic Figure 4 with a real experiment that compares Router selection with random selection from the same candidate pool under a fixed training budget, as Appendix G.3 itself proposes.
  9. I suggest evaluating on the full official test sets and on a true task stream. Section 3.1 motivates a stream of tasks, but the experiments use a fixed train and test split with frozen assets.
8 comments
AuthorsOpenReview ↗

Response to Reviewer ft92: Reproducibility, Memory–Parameter Routing, and Experimental Validation

Thank you for the detailed examination of the evaluation protocol and supporting evidence. We clarify the settings and present the additional experiments relevant to each concern below.

Q1: Do the small, Base-screened samples and single runs support the main claims, especially the small gains over single branches and GRPO’s higher Info score?

A1: Thank you for examining both the evaluation population and the reliability of the comparisons. We would first like to clarify that the QA partitions are randomly sampled into fixed lists shared by all methods. We will make the sampling procedure explicit in Appendix A.7. The submitted partition sizes are:

Table ft92-1. Data partitions used in the submitted experiments

DatasetTrainDevTestSampling / evaluation unit
ChartQA10030200Random, fixed lists; shared by all methods
InfographicVQA10030200Random, fixed lists; evaluation from public validation
VisualToolBench8020214Fixed project partitions
EB-ALFRED10030300Fixed task/scene lists; includes repeated trials

Counts are from Appendix A.7, Table 4. InfographicVQA uses labeled public validation for evaluation.

Following your suggestion, we evaluated the complete method with three different seeds and report its mean and standard deviation below. We also retain the paper’s GRPO result as a direct reference for the QA comparison.

Table ft92-2. Three-seed performance and the GRPO reference

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver: mean ± SD68.38 ± 1.1569.00 ± 1.109.00 ± 3.2010.23 ± 4.10
Mean − strongest Table 1 baseline+1.38+0.62-0.46-1.14
GRPO (paper Table 2)63.0070.92——
Gemma-3-12B
VisEvolver: mean ± SD47.50 ± 0.7552.00 ± 1.806.07 ± 3.7015.33 ± 3.30
Mean − strongest Table 1 baseline+0.50+9.02+1.84+10.33

VisEvolver uses three seeds over each complete evaluation list. Baselines are the strongest per-backbone, per-task entries in paper Table 1, excluding VisEvolver, Memory-only and Weight-only. Qwen references: 67.00 / 68.38 / 9.46 / 11.37; Gemma: 47.00 / 42.98 / 4.23 / 5.00. Differences are score-point differences against those published values, not paired significance tests.

These repeated runs help distinguish a reproducible pattern from a single favorable result. The gains are particularly clear for Gemma on InfographicVQA and EB-ALFRED. GRPO remains strong on InfographicVQA, while VisEvolver is stronger on ChartQA, illustrating that the task requirements matter when choosing how to learn from experience.

We also completed the broader evaluation requested in your review. A11 reports the complete public QA splits and full project VTB/EB lists, where VisEvolver exceeds Base in every backbone–task setting. Together, the repeated runs and expanded evaluation address variability and population coverage separately. We discuss the paired comparison with the stronger single branch in A4, so that smaller margins are assessed using the relevant statistical question.

Q2: Please reconcile Table 1, Table 12 and Figure 3(c), replace synthetic router evidence with real comparisons, and explain the low tool-call frequency.

A2: Thank you for checking how the reported numbers relate to their experimental settings. The main Qwen ChartQA values are Weight-only 68.38 and VisEvolver 68.88, and the Weight entry in Table 12 will be aligned accordingly. Appendix G.2 describes Figure 3(c) as a separate comparison of Naive and Attributed Memory, using Qwen3.5-9B without adaptation, seed 101, and 200 questions per condition. Appendix G.4 identifies Table 12 as the memory-withdrawal diagnostic. We will label these settings explicitly so the intended comparisons are easy to follow.

To address the router question, we added real-task comparisons from the same candidate pool under matched selection counts and update budgets, including individual criterion ablations:

Table ft92-3. Real-task routing comparisons and individual ablations

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver68.8869.009.7010.58
Always-Both67.4067.407.9010.00
Random-matched50.6265.309.208.00
Utility-only67.0065.307.208.00
w/o Support67.0064.226.009.42
w/o Utility68.3864.226.009.42
w/o Stability68.3864.227.908.00
w/o Specificity67.0062.137.909.42
w/o Conflict68.3864.229.208.00
w/o Attribution66.8764.227.909.42
Gemma-3-12B
VisEvolver48.2554.526.0715.33
Always-Both47.5052.656.0712.00
Random-matched42.1246.014.2113.33
Utility-only47.2545.742.2213.33
w/o Support42.1242.983.3912.00
w/o Utility42.7546.013.7412.00
w/o Stability42.7546.013.3912.00
w/o Specificity42.7546.013.3912.00
w/o Conflict42.1245.743.7412.00
w/o Attribution47.5046.013.7412.00

Random-matched draws from the same candidate pool with the same admission count and update budget. “w/o” removes the named criterion or attribution component.

VisEvolver exceeds the random and utility-only selectors, and removing each criterion reduces performance. This provides direct task-level evidence that the selection rule contributes beyond the amount of experience retained.

For tool use, we share your interest in whether the learned behavior is useful. A mean call count combines tasks that require further inspection with tasks that can be answered from the current observation. We therefore measured whether the agent inspects the evidence it needs, as well as redundant calls, and tested task performance under visual interventions.

Table ft92-4. VisEvolver under full-image interventions, with decision-quality measures

DatasetOriginal ↑Selected mask ↑Random mask ↑Selected blur ↑Correct inspection (%) ↑Redundant calls (%) ↓
Qwen3.5-9B
ChartQA68.8847.0047.2546.7356.4 / 53.221.4 / 15.5
InfographicVQA69.0058.7465.3046.0155.6 / 61.48.6 / 16.7
VisualToolBench9.703.744.213.3958.1 / 50.211.7 / 10.5
EB-ALFRED10.588.339.589.5861.0 / 57.115.0 / 16.7
Gemma-3-12B
ChartQA48.2532.0036.4832.0051.7 / 40.39.8 / 11.8
InfographicVQA54.5242.9546.0144.6840.5 / 39.311.6 / 14.8
VisualToolBench6.073.273.274.0045.6 / 43.617.4 / 17.3
EB-ALFRED15.3313.8011.309.5042.4 / 40.017.3 / 11.5

The four task-score columns all evaluate VisEvolver, retaining the surrounding image during intervention; their metrics follow paper Table 1. Each entry in the final two columns is VisEvolver / w/o attribution. Correct inspection rate measures whether the agent checks the evidence required by the task; redundant-call rate measures calls that provide no needed new evidence. Both are measured from execution behavior and reported as percentages.

The visual interventions show that the selected evidence affects actual task outcomes. Correct inspection improves in seven of eight settings when attribution is retained, while redundant-call rates show a more mixed pattern. This helps explain why tool frequency alone is not the right measure of internalization: useful inspection can increase even when total interaction does not decrease. The measured behavior, attribution ablation and task outcomes together provide a more direct account of what the component contributes. We will organize the revised analysis around these distinctions.

Q3: Were the QA challenge lists selected using Base failures, and did all methods use the same lists?

A3: Thank you for asking how the question lists were selected. The QA questions are randomly sampled into fixed partitions, and every method is evaluated on the same question list. The selection rule is not to retain only questions on which Base fails or has high uncertainty. The counts and source population are listed in A1.

We will revise the wording in Appendix A.7 to reflect this sampling procedure explicitly and provide the fixed identifiers and source splits. This will make it clear how the population was constructed and how a shared evaluation list is maintained across methods.


Q4: Please report at least three seeds or orders for Tables 1 and 3, with uncertainty and a test against the stronger branch.

A4: Thank you for specifying which statistical comparisons would help assess the contribution. We added three-seed evaluation of the complete method, with its mean and standard deviation reported in A1. The real-task Table 3 extensions, including random routing and individual criteria, are reported in A2.

The results are reported in Table ft92-2, A1.

These measurements separate the variability of the complete method from the effect of removing components. For the stronger-branch question, the appropriate test pairs each task’s result under VisEvolver and the corresponding single-branch variant, using the same seeds and evaluation list. We will report that paired analysis and the per-ablation seed statistics alongside their respective comparisons. The standard deviation of Full alone and the differences against published baseline point estimates are not substitutes for that test.


Q5: Which runs produced the disputed numbers, and what does Appendix G.4 mean by historical aggregates?

A5: Thank you for pointing out that “historical aggregates” does not clearly identify the underlying comparison. In Appendix G.4, this phrase distinguishes the submitted main-table summaries from the experiment-specific plotting pairs used for the internalization analysis. It does not designate an additional method or a new evaluation metric.

Table 1 is the reference for the main Base, Memory-only, Weight-only and VisEvolver comparison. Figure 3(c) evaluates the two memory-construction procedures under the Qwen configuration specified in A2. Table 12 supplies the memory-withdrawal diagnostic coordinates, with the Qwen ChartQA Weight entry aligned to 68.38. We will replace the ambiguous phrase with explicit references to the relevant configuration, task list, memory asset and seed, and label each diagnostic by its actual purpose. This will make the relationship between the main results and the auxiliary analyses easier to follow.


Q6: Which teacher is used in Section 3.3 and Equation 10, and how is supervision matched to memory baselines?

A6: Thank you for asking whether teacher access is matched. The correction model in Section 3.3 and Eq. 10 is the backbone itself, Qwen3.5-9B or Gemma-3-12B. It reviews its own trajectory with outcome feedback and proposes an action that the student can execute from a visited state. This is how hindsight helps the same model learn what to do differently on a future task.

The supervision entries and measured generation/review token usage are:

Table ft92-5. Supervision volume and additional token cost

QuantityChartQAInfographicVQAVisualToolBenchEB-ALFRED
Qwen3.5-9B
Supervision entries2303968025
Additional generation/review tokens (M)4.316.176.677.10
Gemma-3-12B
Supervision entries18615380252
Additional generation/review tokens (M)2.421.651.561.91

Supervision entries are training targets, rather than independent teacher calls. Each QA set includes 50 preservation entries; Gemma EB-ALFRED reports stage 2, including 124 cross-backbone replay and 124 scripted-action entries. Tokens are measured generation/review usage in millions. This table reports that additional component, rather than total training and inference compute.

In the matched comparison, memory baselines receive the same supervision and budget and process that experience through their own mechanisms. VisEvolver additionally checks the correction’s visual support and executability before deciding whether the experience belongs in memory or parameters. The table makes the extra supervision and preparation cost explicit, while the shared backbone and budget clarify the comparison conditions. We will include these details beside the supervision procedure.

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

The authors propose VisEvolver, a self-evolving framework of multimodal tool-using agents. It tries to address the problem of which relevant experiences must be stored in an external memory or committed into the model weights. The authors argue that existing self-evolving methods use only one form of persistence. Memory based methods keep experience inspectable but must run retrieval on every task and the policy never changes. Parameter update based methods make the behaviour permanent but do not record its context while being harder to inspect and revise. VisEvolver tries to bridge this disconnect by deciding the destination for each experience the agent encounters. Overall, I find the idea interesting but there are some things that are unclear/not tested carefully.

Critical strengths
  1. Explicitly framing the problem of where each experience should go (external memory vs weights). The paper makes a concrete per-experience routing step based on a couple of factors (cross task support, utility stability, instance specificity and conflict). The weight updates help more on perception heavy tasks while memory updates help more on procedural tasks.

  2. The design choices make sense and target an interesting problem. What the authors term as student-reachable hindsight supervision only supervises actions at states the deployed policy actually visits, and trains the evidence gathering action. This avoids teaching the student to act on information it has perhaps not seen yet. The paired visual interventions also ask whether a retained region actually caused the success rather than merely co-occurring with it.

Critical weaknesses
  1. The evidence for the experience router is a bit thin.
  • Fig 4 which is the only analysis comparing router selection with random selection, is explicitly synthetic.
  • Main results are hard to interpret statistically as there is no variance reported, and some margins over the single branch variants are small which is again difficult to assess without variance.
  • Some per-task differences rest on very few tasks. The visual-attribution analysis (Figure 3) compares crop-only scores of 14.84% vs 7.81% on 128 tasks, roughly 19 vs 10 tasks.
  1. Inconsistent reporting and missing method details.
  • Figure 3(c). The naive-memory ChartQA score (44.50) does not match any memory-only result in Table 1, and the backbone is not stated.
  • There are mismatches in numbers between the tables in the appendix and tables in the main paper.
  • For the blank control, Section 4.3 describes it as blanking the region "while keeping its position," and uses this to argue that the causal content lies inside the region. Appendix C.2 states it is a black image of the crop’s size, with no full image, and concedes this "does not establish necessity within the full image."
  • And for MemRL, The main text calls it a direct parameter-training reference, but Appendix A.9 says it keeps the backbone frozen, which is it?
  • The teachmodel used for hindsight supervision is never identified. If it is stronger than the student part of the gain could be appropriated to ordinary distillation and not self-evolution.
  • SFT and GRPO baselines are using 4-bit QLora but I couldn't find VisEvolver's precision.
Questions for authors
  1. Which model serves as the teacher in the student-reachable hindsight supervision?

  2. Which backbone and configuration does Figure 3(c) use? Its naive-memory ChartQA score of 44.50 does not match any memory-only result in Table 1.

  3. How exactly were the "challenge populations screened with Base" constructed (A.7)? What was the selection criterion: base-agent failures, base-agent uncertainty, or something else? Please also report results on the official test splits, or on a random sample of them, for ChartQA and InfographicVQA.

  4. Are the margins over the single branch variants statistically reliable?

  5. Which blank control was actually used in the visual attribution analysis?

12 comments
AuthorsOpenReview ↗

Response to Reviewer YpDf: Robustness, Visual Evidence, and Broader Evaluation

Thank you for the careful reading and constructive questions. Your comments help us explain the experiments more clearly. We respond to each point below.

Q1: The router-versus-random analysis in Figure 4 is synthetic. What real routing evidence supports the method?

A1: Thank you for encouraging a direct test of the router on real tasks. Figure 4 is a synthetic illustration, as specified in Appendix G.3: its experience utilities are simulated to illustrate the comparison, rather than measured from agent executions. The empirical router comparison in the submitted paper is Table 3, which compares VisEvolver with Always-Both.

Following your suggestion, we added real-task comparisons against random selection and utility-only selection, together with individual criterion ablations. Random-matched uses the same candidate pool, admission count and update budget as VisEvolver.

Table YpDf-1. Real-task routing comparisons and individual ablations

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver68.8869.009.7010.58
Always-Both67.4067.407.9010.00
Random-matched50.6265.309.208.00
Utility-only67.0065.307.208.00
w/o Support67.0064.226.009.42
w/o Utility68.3864.226.009.42
w/o Stability68.3864.227.908.00
w/o Specificity67.0062.137.909.42
w/o Conflict68.3864.229.208.00
w/o Attribution66.8764.227.909.42
Gemma-3-12B
VisEvolver48.2554.526.0715.33
Always-Both47.5052.656.0712.00
Random-matched42.1246.014.2113.33
Utility-only47.2545.742.2213.33
w/o Support42.1242.983.3912.00
w/o Utility42.7546.013.7412.00
w/o Stability42.7546.013.3912.00
w/o Specificity42.7546.013.3912.00
w/o Conflict42.1245.743.7412.00
w/o Attribution47.5046.013.7412.00

Random-matched draws from the same candidate pool with the same admission count and update budget. “w/o” removes the named criterion or attribution component.

VisEvolver outperforms the random and utility-only selectors on both backbones across all four tasks. Removing individual criteria also reduces performance, showing that the benefit depends on which experiences are retained, rather than simply admitting a fixed quantity of experience. We will use this real-task comparison as the evidence for routing effectiveness and clearly distinguish it from the illustrative purpose of Figure 4.


Q2: How variable are the results, especially the small margins over single-branch variants?

A2: Thank you for emphasizing that the size of an improvement should be read together with its variability. Following your suggestion, we evaluated the complete method with three different seeds and report the mean and standard deviation over each complete evaluation list below. We also provide the difference between the new mean and the strongest corresponding baseline in paper Table 1, excluding our two single-branch variants.

Table YpDf-2. VisEvolver performance across three seeds

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver: mean ± SD68.38 ± 1.1569.00 ± 1.109.00 ± 3.2010.23 ± 4.10
Mean − strongest Table 1 baseline+1.38+0.62-0.46-1.14
Gemma-3-12B
VisEvolver: mean ± SD47.50 ± 0.7552.00 ± 1.806.07 ± 3.7015.33 ± 3.30
Mean − strongest Table 1 baseline+0.50+9.02+1.84+10.33

VisEvolver uses three seeds over each complete evaluation list. Baselines are the strongest per-backbone, per-task entries in paper Table 1, excluding VisEvolver, Memory-only and Weight-only. Qwen references: 67.00 / 68.38 / 9.46 / 11.37; Gemma: 47.00 / 42.98 / 4.23 / 5.00. Differences are score-point differences against those published values, not paired significance tests.

The results make the stability of the complete method easier to assess. The QA results are comparatively stable across seeds, whereas tool-use and embodied tasks show larger variation. The gains are clearest for Gemma on InfographicVQA and EB-ALFRED, supporting the usefulness of the combined system beyond a single favorable run. We discuss the smaller single-branch margins separately in A13 and do not treat the reference differences in this table as significance tests.

Q3: The original crop analysis involves few correct cases. Can you enlarge the sample and retain paired controls?

A3: Thank you for pointing out that the original percentages should be supported by a larger sample. Following your suggestion, we expanded the crop analysis to 250 questions on each of ChartQA and InfographicVQA, for both backbones. Within each setting, all five input conditions use the same questions. We also include the normal full-image VisEvolver score from paper Table 1 as a reference.

Table YpDf-3. Paired crop controls on 250 questions per dataset

DatasetText onlySelected cropRandom crop, same imageCrop from another imageBlack cropVisEvolver, full image (Table 1 reference)
Qwen3.5-9B
ChartQA0.3010.009.600.802.4068.88
InfographicVQA3.503.473.592.401.9269.00
Gemma-3-12B
ChartQA10.0018.0010.006.0012.0048.25
InfographicVQA6.006.009.972.0011.0454.52

The first five conditions use the same 250 questions within each backbone–dataset pair. The final column reproduces paper Table 1 under normal full-image inputs; it provides context and is not a paired 250-question treatment.

The expanded comparison helps explain what the crop diagnostic measures. A selected crop can retain useful local evidence, particularly on ChartQA, while information distributed across a larger layout can make a different crop informative as well. Thus, the role of attribution is to examine whether the recorded evidence supports a procedure, rather than to assume that one crop can solve the complete task by itself. The full-image interventions in A6 directly test the selected region within the normal visual context. We will present the expanded sample and the two complementary interpretations together.


Q4: Why does Figure 3(c) report 44.50 for naive memory, and which model and configuration produced it?

A4: Thank you for asking us to identify this comparison more clearly. Figure 3(c) uses Qwen3.5-9B without a LoRA adapter, with evaluation seed 101 and 200 questions per condition. The ChartQA scores are 44.50 for Naive Memory and 49.50 for Attributed Memory; the corresponding InfographicVQA scores are 72.38 and 73.25.

As explained in Appendix G.2, this experiment compares two memory-construction procedures while supplying the original task inputs. Naive Memory retains experience without the attribution filter, whereas Attributed Memory applies that filter. Table 1’s Memory-only row is the main-system branch comparison with its own memory asset and prompt. The relevant contrast in Figure 3(c) is therefore between the two matched memory conditions, which tests whether attribution improves memory quality. We will specify the backbone and configuration in the caption so this distinction is immediately clear.


Q5: Please explain the inconsistent values between the main tables and the appendix.

A5: Thank you for carefully comparing the numbers across the main paper and appendix. The main comparison should use the values in Table 1. For Qwen ChartQA, Weight-only is 68.38 and VisEvolver is 68.88; the Weight entry in Table 12 will be corrected to 68.38.

The other cited numbers belong to different diagnostic comparisons. Appendix G.2 specifies the Naive versus Attributed Memory experiment in Figure 3(c), described in A4. Appendix G.4 identifies Table 12 as the experiment-specific coordinates of the memory-withdrawal analysis; its Gemma InfographicVQA pair, 55.57 and 60.70, is not the main-table Base/Weight pair of 42.95 and 46.01. We will label these experimental settings directly in their captions and align the numerical references throughout the paper, so that each conclusion is read against its corresponding comparison.

Q6: Was the blank control a black crop or a masked region inside the full image?

A6: Thank you for asking us to clarify the exact visual input. The blank control in Figure 3 replaces the selected crop with an RGB black image of the same size; it does not provide the surrounding full image. This is the setting described in Appendix C.2 and Table 9. It preserves the crop dimensions and image channel.

Your question also motivates a direct test in the full-image setting. We therefore added masking and blurring interventions that retain the surrounding image, together with measured execution-behavior comparisons:

Table YpDf-4. VisEvolver under full-image interventions, with decision-quality measures

DatasetOriginal ↑Selected mask ↑Random mask ↑Selected blur ↑Correct inspection (%) ↑Redundant calls (%) ↓
Qwen3.5-9B
ChartQA68.8847.0047.2546.7356.4 / 53.221.4 / 15.5
InfographicVQA69.0058.7465.3046.0155.6 / 61.48.6 / 16.7
VisualToolBench9.703.744.213.3958.1 / 50.211.7 / 10.5
EB-ALFRED10.588.339.589.5861.0 / 57.115.0 / 16.7
Gemma-3-12B
ChartQA48.2532.0036.4832.0051.7 / 40.39.8 / 11.8
InfographicVQA54.5242.9546.0144.6840.5 / 39.311.6 / 14.8
VisualToolBench6.073.273.274.0045.6 / 43.617.4 / 17.3
EB-ALFRED15.3313.8011.309.5042.4 / 40.017.3 / 11.5

The four task-score columns all evaluate VisEvolver, retaining the surrounding image during intervention; their metrics follow paper Table 1. Each entry in the final two columns is VisEvolver / w/o attribution. Correct inspection rate measures whether the agent checks the evidence required by the task; redundant-call rate measures calls that provide no needed new evidence. Both are measured from execution behavior and reported as percentages.

On the QA tasks, masking the selected region is more harmful than masking a matched random region. Thus, the selected content matters even when the agent can still access the rest of the image. Attribution also improves correct inspection in most settings, while the number of redundant calls does not change uniformly. These results distinguish the contribution of useful evidence from the mere frequency of tool use. We will describe the original black-crop control and the new full-image interventions separately in Section 4.3 so that each conclusion is tied to the input actually tested.


Q7: Does MemRL train the backbone or keep it frozen?

A7: Thank you for pointing out that the description should distinguish backbone updates from retrieval updates. The evaluated MemRL configuration keeps the backbone frozen and updates memory retrieval utilities, as specified in Appendix A.9. It retrieves five candidates and selects three entries; learning changes their retrieval values rather than the backbone weights. The direct backbone-training references in Table 2 are SFT and GRPO.

We will use this distinction consistently when describing the baselines. It also makes their relationship to VisEvolver clearer: retrieval adaptation changes which experience is consulted, while parameter consolidation changes the policy’s behavior even without retrieving that experience.

Reviewer_L6qROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This research investigates hybrid parameter-and-memory self-evolution for multimodal agents, focusing on whether individual experiences should be stored in external memory, internalized into model parameters, or both. To this end, the authors propose VisEvolver, a framework that assesses visual grounding of candidate experiences and selectively routes them through training-based and training-free consolidation pathways. Experiments across four multimodal benchmarks and two backbones show improvements and demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The motivation is clear, reasonable, and meaningful. The problem addressed in this paper is also important and timely, as model-harness co-evolution has emerged as a promising paradigm for agent self-evolution in recent research community.

  2. The proposed multimodal evolution framework is clearly-designed, with each component addressing a specific challenge in multimodal self-evolution. For instance, paired visual interventions help identify visually grounded experiences, while student-reachable hindsight supervision keeps training targets consistent with the agent's actual capability.

Critical weaknesses
  1. Prior Research on Model–Harness Co-Evolution Routing is Overlooked. The paper mainly discusses memory-based and parameter-based self-evolution as two existing research directions, while overlooking closely related hybrid self-evolution approaches (e.g. Skill0.5, COVE). These hybrid approaches have already started exploring how external skills or memory update and parameter update can be coordinated and selected via routing mechanisms. Therefore, I encourage the authors to further discuss this line of research in related works and better position their framework by clarifying contributions of their experience-level routing mechanism and multimodal-tailored design. I would appreciate it if the authors could provide convincing experimental comparisons with existing routing strategies (e.g., the task-aware routing approach in COVE) under comparable multimodal settings and training budgets, if feasible.

  2. Insufficient Ablation Analysis of the Fine-Grained Routing Criteria. The current ablation compares the proposed Router against an Always-Both variant that sends all admitted experiences to both consolidation pathways. While this provides evidence for the benefit of selective routing, it remains unclear whether all five proposed criteria are necessary. Appendix G.3 includes a synthetic illustration of random routing, but no empirical comparison is provided. Given that routing is a central contribution of this paper, I encourage the authors to further conduct finer-grained ablations of individual criteria and compare against simpler routing strategies under matched training budgets. Such comparisons would help establish whether the five-criterion design offers sufficient benefits to justify its complexity.

  3. Unclear Causal Link Between Visual Evidence and Agent Decisions. Causal Experience Attribution is another critical component of VisEvolver, yet the visual interventions in Figure 3 are conducted on isolated image regions rather than within the agent's actual decision context. As acknowledged in Appendix C.2, these comparisons cannot establish whether the attributed regions are necessary when the full image is available. Therefore, the current results demonstrate the usefulness of selected visual regions, but provide limited evidence that these regions causally contribute to the agent's actual decisions and task outcomes. It would be helpful if the authors could clarify the scope of their causal claims and provide additional experiments under full-image or actual agent rollout settings to better support this component.

Questions for authors

My main concerns are discussed in the Weaknesses section. I have one additional question:

  1. Unclear Teacher Configuration and Fairness of Baseline Comparisons. VisEvolver relies on Teacher-generated corrective actions, but the paper does not specify which Teacher model is used or whether the compared baselines receive comparable supervision. Although the appendix describes several controls for baseline comparisons, it remains unclear whether the additional Teacher assistance introduces an advantage over existing self-evolution methods. Could the authors clarify the Teacher configuration, including the amount of corrective supervision and its computational cost? Comparisons under matched Teacher supervision and training budgets would help isolate the contribution of the proposed routing framework.
5 comments
AuthorsOpenReview ↗

Response to Reviewer L6qR: Hybrid Comparisons, Routing Design, and Visual Grounding

Thank you for the thoughtful suggestions on related work, routing, visual evidence and supervision. We address each point below with the corresponding comparisons and implementation details.

Q1: How does VisEvolver relate to hybrid approaches such as COVE and Skill0.5, and how does it compare under comparable multimodal budgets?

A1: Thank you for recommending these closely related studies. We would first like to clarify that paper Table 1 already includes OPD-Evolver, which combines experience handling with on-policy distillation. Appendix A.9 describes our executor-only distillation adaptation. VisEvolver outperforms this baseline in all eight backbone–task settings in the submitted table.

VisEvolver asks a more specific question: given an individual multimodal experience, what should remain in editable memory, and what should become part of the policy? Sections 3.2–3.3 make this decision using the supporting visual evidence, applicability across tasks, and corrections that the student can execute from states it actually visits. This lets the system internalize reusable inspection behavior while retaining context-dependent procedures in memory.

Following your suggestion, we added COVE and Skill0.5 to the comparison on both backbones and all four benchmarks, using the corresponding models and comparable task interfaces and budgets.

Table L6qR-1. Comparison with hybrid self-evolution methods

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
COVE46.7361.669.008.00
Skill0.550.6262.139.008.17
OPD-Evolver67.0068.389.238.33
VisEvolver68.8869.009.7010.58
Gemma-3-12B
COVE41.0053.013.395.00
Skill0.541.0053.404.234.50
OPD-Evolver40.5035.123.745.00
VisEvolver48.2554.326.0715.33

All scores use a 0–100 scale. The two backbones are evaluated separately.

VisEvolver achieves the highest score in all eight settings. The benefit extends across visual question answering, tool use and embodied interaction, indicating that the experience-level design is useful across different task requirements. These results strengthen the comparison with existing hybrid approaches and help clarify our contribution within that line of work. We will include the new results and describe the multimodal adaptations in the revised related-work and experimental sections.

Q2: Are all five routing criteria useful? Please compare individual ablations and simple routers under matched budgets.

A2: Thank you for asking which parts of the router contribute to its performance. Paper Table 3 already compares VisEvolver with Always-Both, showing the value of assigning experiences selectively. To examine the five criteria more closely, we added individual ablations on both backbones and all four tasks. We also compared with random selection from the same candidate pool, matching the admission count and update budget, and with selection based only on utility.

Table L6qR-2. Real-task routing comparisons and individual ablations

Method / conditionChartQA Acc. ↑InfoVQA ANLS ↑VTB task score ↑EB success ↑
Qwen3.5-9B
VisEvolver68.8869.009.7010.58
Always-Both67.4067.407.9010.00
Random-matched50.6265.309.208.00
Utility-only67.0065.307.208.00
w/o Support67.0064.226.009.42
w/o Utility68.3864.226.009.42
w/o Stability68.3864.227.908.00
w/o Specificity67.0062.137.909.42
w/o Conflict68.3864.229.208.00
w/o Attribution66.8764.227.909.42
Gemma-3-12B
VisEvolver48.2554.526.0715.33
Always-Both47.5052.656.0712.00
Random-matched42.1246.014.2113.33
Utility-only47.2545.742.2213.33
w/o Support42.1242.983.3912.00
w/o Utility42.7546.013.7412.00
w/o Stability42.7546.013.3912.00
w/o Specificity42.7546.013.3912.00
w/o Conflict42.1245.743.7412.00
w/o Attribution47.5046.013.7412.00

Random-matched draws from the same candidate pool with the same admission count and update budget. “w/o” removes the named criterion or attribution component.

Every individual criterion ablation scores below VisEvolver in these settings. The random and utility-only selectors also perform worse, showing that retaining a fixed amount of experience, or choosing it by average utility alone, does not reproduce the complete router’s behavior. The criteria address complementary questions: support and stability ask whether an improvement recurs, while specificity and conflict check whether a procedure can be reused without carrying over incidental details or incompatible actions. We hope these real-task results make the reason for the five-criterion design clearer, and will include them in the revised paper.

Q3: Do the selected regions affect decisions when the full image is available? Please test visual attribution in actual decision contexts.

A3: Thank you for pointing out that the selected evidence should be tested in the agent’s actual decision context. Section 4.3 and Appendix C.2 examine the information contained in isolated crops. Following your suggestion, we now intervene on the full image received by VisEvolver while retaining the surrounding visual context.

For each evaluation task, we compare the original image with masking the selected region, masking a matched random region, and blurring the selected region. We additionally measure correct inspection and redundant-call rates from execution behavior, comparing VisEvolver with its attribution ablation. The table places these behavioral measures alongside task performance to show both what changes and whether the resulting behavior is useful.

Table L6qR-3. VisEvolver under full-image interventions, with decision-quality measures

DatasetOriginal ↑Selected mask ↑Random mask ↑Selected blur ↑Correct inspection (%) ↑Redundant calls (%) ↓
Qwen3.5-9B
ChartQA68.8847.0047.2546.7356.4 / 53.221.4 / 15.5
InfographicVQA69.0058.7465.3046.0155.6 / 61.48.6 / 16.7
VisualToolBench9.703.744.213.3958.1 / 50.211.7 / 10.5
EB-ALFRED10.588.339.589.5861.0 / 57.115.0 / 16.7
Gemma-3-12B
ChartQA48.2532.0036.4832.0051.7 / 40.39.8 / 11.8
InfographicVQA54.5242.9546.0144.6840.5 / 39.311.6 / 14.8
VisualToolBench6.073.273.274.0045.6 / 43.617.4 / 17.3
EB-ALFRED15.3313.8011.309.5042.4 / 40.017.3 / 11.5

The four task-score columns all evaluate VisEvolver, retaining the surrounding image during intervention; their metrics follow paper Table 1. Each entry in the final two columns is VisEvolver / w/o attribution. Correct inspection rate measures whether the agent checks the evidence required by the task; redundant-call rate measures calls that provide no needed new evidence. Both are measured from execution behavior and reported as percentages.

Disrupting the selected region lowers task performance in every backbone–task setting. On the QA tasks, selected-region masking is more harmful than random masking, supporting the relevance of the selected evidence when the rest of the image remains visible. The behavioral results offer complementary evidence: attribution improves correct inspection in seven of eight settings, while its effect on redundant calls varies by task. This suggests that its main benefit is helping the agent inspect useful evidence, rather than uniformly minimizing tool use. Together with the attribution ablation in A2, these results connect the visual checks to both execution behavior and downstream task outcomes. We will include this full-context analysis in the revised paper.

Q4: Which teacher generates corrections, how much supervision and compute does it add, and are the baseline comparisons matched?

A4: Thank you for asking us to make the supervision and its cost explicit. The correction model is the corresponding backbone itself: Qwen3.5-9B for the Qwen experiments and Gemma-3-12B for the Gemma experiments. As described in Section 3.3, the model reviews its trajectory with outcome feedback and proposes a correction at a state the student actually visited. The target must be executable using the information available at that state.

For example, if the student needs to inspect a region before it can answer, the supervision teaches that inspection step before the answer. Embodied training also includes task-specific action replay and scripted targets. Following your suggestion, we report the supervision entries and measured generation/review token usage below.

Table L6qR-4. Supervision volume and additional token cost

QuantityChartQAInfographicVQAVisualToolBenchEB-ALFRED
Qwen3.5-9B
Supervision entries2303968025
Additional generation/review tokens (M)4.316.176.677.10
Gemma-3-12B
Supervision entries18615380252
Additional generation/review tokens (M)2.421.651.561.91

Supervision entries are training targets, rather than independent teacher calls. Each QA set includes 50 preservation entries; Gemma EB-ALFRED reports stage 2, including 124 cross-backbone replay and 124 scripted-action entries. Tokens are measured generation/review usage in millions. This table reports that additional component, rather than total training and inference compute.

This accounting makes the additional work explicit: the supervision counts describe the training targets, while tokens measure the generation and review needed to prepare them. In the matched comparison, baselines receive the same supervision and budget through their respective update mechanisms. VisEvolver’s contribution is to check whether a correction is supported and executable, and then decide where the experience should persist. We will include the teacher identities, supervision breakdown and cost scope next to the training procedure so these conditions are straightforward to reproduce.

Thanks again for your careful and valuable comments, hope our response addresses your concerns. If our responses answer your questions well and reassure your concerns, we would appreciate if you could reassess our work and increase the rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.