acceptodds
Under review as a conference paper at ICLR 2027

SimpleSearch-VL: A Simple Recipe for Efficient Multimodal Agentic Search

Abstract

We present SimpleSearch-VL, an efficient, reliable, and practical framework for multimodal agentic search. Its core idea is to improve the agent's own search-and-verification process rather than scaling data, tools, or auxiliary model components. For efficiency, Factorized Adaptive Rollout (FAR) improves sampling efficiency by forming more informative training groups while using redundant samples to mitigate long-tail latency and expose hard samples. For reliability, SimpleSearch-VL performs evidence-verified reasoning, explicitly using chain-of-thought verification to assess the relevance of retrieved visual and textual cues to the original context. For practicality, SimpleSearch-VL keeps a lightweight tool interface and performs webpage self-summary within the agent, requiring no additional external dependencies. With only 5K supervised tool-interleaved trajectories and 2K RL data, SimpleSearch-VL improves Qwen3-VL agentic baselines by 15.8 and 16.0 average points for the 8B and 30B-A3B variants, respectively. The SimpleSearch-VL-30B-A3B model further achieves performance competitive with agentic Gemini-3-Pro.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MLLMs, Agentic Search, Reinforcement Learning

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_7uQqOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes 3 methods to improve multimodal agentic search, including Factorized Adaptive Rollout to improve RL training efficency, evidence-verified reasoning to generate training data that encourages evidence verification, and self-summary to summarize webpages without a separate model. With these methods, SimpleSearch-VL achieves better performance than many competing baselines.

Critical strengths
  1. Good performance in Table 1
Critical weaknesses
  1. Unfair experimental setup - SimpleSearch-VL uses different training data, search pipeline, tools and compute compared to other methods in Table 1. I don't see any fair comparison that isolates the improvements from the proposed methods. How do we know the gains comes from these methods rather than simply better training data or search pipeline? Without a controlled comparison, it is hard to tell what actually contributes to the performance improvement.
  2. Limited novelty - FAR is just allocating more rollouts to unsuccessful prompts. Evidence-Verified Reasoning is mainly using teacher-generated trajectories and Gemini to filter and rewrite reasoning, which have already been explored in SenseNova-MARS, OpenSearch-VL, etc. Self-summary also seems more like a pipeline change rather than a new method. I don't clearly see what is technically new here or how much each of these methods actually contributes.
Questions for authors
  1. Why are some baseline results missing in Table 1? e.g. SenseNova-MARS-32B have result for LiveVQA and SimpleVQA. Why are they not in the table? Have all baseline results been checked? If results missing but model released, they should be evaluated using the same settings as SimpleSearch-VL for better comparison.
  2. Isn't FAR partly just benefiting from more rollouts? Figure 8(b) already shows that increasing rollout size from 4 to 6 gives a large improvement. How does FAR compare with fixed rollout sizes of 6 and 8 under the same training settings and compute budget?
  3. Is the improvement from evidence-verified reasoning or just better training data? Table 2 compares Direct SFT and Evidence-Aware SFT, but appendix C also mentions filtering trajectories and rewriting final answers. Are the same training examples and final answers used? How does this ablation isolate verification reasoning itself?
  4. Why does self-summary perform better than a dedicated summarizer in Table 3? What checkpoint and prompt are used for the dedicated summarizer? Are the generation settings matched?
  5. Table 4 compares H200 with H20 GPUs and uses different training data. How is this a fair comparison of training efficiency? Can you provide a fairer comparison that accounts for hardware differences and data construction costs?
10 comments
AuthorsOpenReview ↗

Overview of our response and manuscript revisions

Thank you for the careful review and for recognizing SimpleSearch-VL's performance in Table 1. We respond to your concerns in seven comments below:

  1. Comparison fairness: The training data and tools used by existing search agents, and how SimpleSearch-VL compares in training cost, performance, and model requirements at inference.
  2. Technical contributions: The problems addressed by FAR, visual verification, and self-summary, and the experiments supporting each design.
  3. Missing baseline results: The corrected SenseNova-MARS scores, checks of the other added results, and the baselines evaluated under the same protocol as SimpleSearch-VL.
  4. FAR versus fixed rollouts: Additional fixed n=6 and n=8 experiments over 100 training steps, plus the separate ablation after SFT.
  5. Evidence-verification ablation: The paired SFT comparison, which changes verification reasoning while keeping examples, tool traces, and final answers fixed.
  6. Self-summary: Why sharing the agent's weights allows indirect task adaptation, with the generation settings and serving results.
  7. Training resources: The additional single-node H20 run and the scope of the training-cost comparison.

Changes specific to your review are marked in red. Each reply points to the relevant sections and tables. Please feel free to ask follow-up questions under each comment; we will respond promptly.

AuthorsOpenReview ↗

Response to Critical Weakness 1: Cross-system comparisons, data efficiency, and controlled evidence

Thank you for recognizing the performance in Table 1. Multimodal search agents typically construct their own training trajectories and adopt method-specific harness configurations and tool interfaces. Representative examples include:

  • SenseNova-MARS combines FVQA, Pixel-Reasoner, and curated expert data for approximately 3K cold-start examples, followed by RL on FVQA, DeepEyes-4K, and Visual-Probe [1].
  • Vision-DeepResearch constructs long-horizon trajectories from verified VQA, text-only QA, and synthesized fuzzy multi-hop questions, using 30K SFT trajectories and 15K RL examples for its 30B-A3B model [2].
  • DeepMMSearch-R1 distills Gemini trajectories from InfoSeek into 10K SFT examples, then uses FVQA for RL [3].
  • OpenSearch-VL combines Wikipedia-derived questions with LiveVQA, FVQA, and WebQA, producing 36,592 SFT trajectories and approximately 8K RL examples [4].

Our tool set is also deliberately simple: text search, reverse-image search, and webpage visiting. These functions are already covered by Vision-DeepResearch and OpenSearch-VL [2, 4]; OpenSearch-VL additionally provides OCR and image-enhancement tools [4]. These overlaps make the cited systems relevant comparators, while data construction and execution remain part of each method. Table 1 compares complete systems; matched backbone evaluations and component ablations address where our gains come from. We summarize the evidence along three dimensions:

1. Training efficiency and resource cost. Like many existing multimodal search agents, we draw our training questions from public QA sources, including FVQA, LiveVQA, WebQA, and Wikipedia-derived QA. Our data sources substantially overlap with those used by existing methods, particularly OpenSearch-VL. We use 5,193 versus 36,592 SFT trajectories, and 1,995 versus approximately 8K RL examples [4]. SFT+RL takes approximately 18 hours at 8B and 27 hours at 30B-A3B on one eight-H200 node (Table 5; configuration in Appendix A). We have also completed an 8B run on one eight-H20 node (96 GB per GPU), taking 10 hours for SFT and 27 hours for RL—37 hours in total.

2. Search performance and controlled evidence. Despite using fewer training examples, our 8B model has higher reported scores than OpenSearch-VL-30B-A3B on the five shared benchmarks in Table 1. For the within-system comparison, each Qwen3-VL agentic baseline and its trained counterpart use the same evaluation harness and protocol within that model scale. The six-benchmark average improves by 15.8 percentage points at 8B and 16.0 at 30B-A3B, reflecting the full post-training recipe. The component studies provide more targeted evidence: FAR (2×, 4–6) improves the three-benchmark average from 66.3 to 70.7, with the SFT initialization, RL data, and training steps held fixed (revised Table 20).

3. Model capacity without an auxiliary LLM. During search inference, the same 8B or 30B-A3B policy performs search planning, visual evidence verification, webpage summarization, and final-answer generation. Our self-summary design does not call a separate LLM to process retrieved evidence. Thus, the reported agent size is also the size of the model performing these reasoning and summarization tasks: an 8B result is not obtained by pairing an 8B agent with a larger auxiliary LLM. This makes the model capacity behind the results explicit while avoiding an additional summarizer deployment.

The SFT-initialized FAR control is summarized in Section 3.3, with full results in Appendix F.3 (Table 20).

References

[1] Yong Xien Chng et al. SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning. In CVPR, 2026. SenseNova-MARS technical report: arXiv:2512.24330.

[2] Wenxuan Huang et al. Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models. In ICML, 2026.

[3] Kartik Narayan et al. DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search. arXiv preprint arXiv:2510.12801, 2025.

[4] Shuang Chen et al. OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents. arXiv preprint arXiv:2605.05185, 2026.

AuthorsOpenReview ↗

Response to Critical Weakness 2: Technical contributions of FAR, visual verification, and self-summary

Our contribution is a practical recipe for making multimodal search agents efficient to train, grounded in visual evidence, and simpler to deploy. We target three concrete bottlenecks: slow and uneven tool-interleaved rollouts, plausible but visually mismatched search results, and the overhead of a separately served webpage summarizer.

1. FAR: recovering useful RL signal while controlling rollout cost. Fixed sampling can leave hard prompts with only incorrect answers and no answer-level contrast, while a few long tool-call trajectories delay the entire step. FAR coordinates prompt expansion, selective within-prompt allocation, and stopping rules: it expands all-wrong groups breadth-first, masks later expanded slots once a correct answer appears, and can end generation once enough valid groups have been collected or the near-completion criterion is met. At most n_min trajectories per retained group enter the update, while the full completed group informs the leave-one-out baseline. Thus, FAR changes both where additional exploration is spent and how long the step waits, rather than merely assigning more rollouts to unsuccessful prompts.

Simple scheduling rules can yield substantive gains. DAPO's dynamic sampling uses oversampling and group filtering [1], while APRIL addresses rollout tails through over-provisioning, early termination, and resumption [2]. What matters is resolving a concrete bottleneck and saving training time without sacrificing accuracy, not rule complexity.

Figure 4(a) shows that 4–6 and 4–8 allocation recover 11.7% and 17.7% of initially all-wrong groups. In Figure 4(b)'s fixed-step direct-RL study, FAR (2×, 4–6) improves the three-benchmark average from 56.6 to 62.0, taking 165 seconds per step versus 192 for standard sampling. Separately, the same FAR configuration improves the three-benchmark average from 66.3 to 70.7 in the SFT-initialized control, holding the initial checkpoint, RL data, and training steps fixed (revised Table 20). These results support hard-group recovery and a better accuracy–time trade-off.

2. Evidence-verified reasoning: checking visual correspondence before trusting a source. A search result can provide a convincing title and URL while depicting a different object from the queried image or region. The contribution is to make this correspondence check explicit in the agent's training: inspect returned thumbnails, assess whether they support the current hypothesis, and decide whether further evidence is needed. Our audit and rewrite pipeline constructs trajectories around these decisions. The emphasis is on how the agent learns to use visual evidence, not the identity of the teacher or auditor; thumbnail access itself is also available in SenseNova-MARS [3].

Table 2 reports 61.0 → 64.4 (+3.4 points) for the evidence-aware SFT data pipeline. A separate control on the untuned checkpoint improves the four-benchmark average from 41.4 to 44.3 when thumbnails are enabled. These support the data pipeline and the value of retrieved visual evidence. Section 3.3.2 reports the untuned-policy control in Table 3.

3. Self-summary: removing a model dependency and sharing serving resources. A dedicated summarizer requires another model deployment and separates webpage processing from the search policy that uses its output. We instead route webpage summarization through the policy's own rollout engine. This removes the separate summary service, lets search and summary generation share serving resources, and gives both tasks access to the same task-adapted weights. Reasoning, summarization, and answering thus remain within the agent, without an auxiliary LLM. Shared task-adapted weights also offer a plausible explanation for the accuracy benefit.

In Table 4's 8B setting, self-summary achieves 69.8 versus 67.9 average accuracy with one policy-serving node instead of one agent node plus one summary node. At the same two-node budget, two self-summarizing agent replicas complete the evaluation workload in 358 versus 501 seconds (28.5% less time), with 69.7 versus 67.9 average accuracy. This measures a resource-allocation benefit, not just a simpler interface.

Together, these designs improve training signal, visual verification, and deployment simplicity. Their significance lies in these practical capabilities and the experimental evidence for them, rather than in adding more components to the system.

References

[1] Qiying Yu et al. DAPO: An Open-Source LLM Reinforcement Learning System at Scale. In NeurIPS, 2025.

[2] Yuzhen Zhou et al. APRIL: Active Partial Rollouts in Reinforcement Learning to Tame Long-Tail Generation. arXiv preprint arXiv:2509.18521, 2025.

[3] Yong Xien Chng et al. SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning. In CVPR, 2026. SenseNova-MARS technical report: arXiv:2512.24330.

Reviewer_9LtrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SimpleSearch-VL, a multimodal search agent that answers visual questions by searching the web and checking retrieved evidence to reason. The main contribution is better rollout allocation for RL and evidence verification to achieve better performance with less data. The agent uses three tools: text search, reverse image search over selected image regions, and web search. It can choose regions from input images, inspect search results, gather more evidence, and then answer.

This paper introduces Factorized Adaptive Rollout (FAR), which augments group-based RL by expanding the candidate prompt pool, normally by 2x, and by making more attempts on difficult questions. The policy uses an RLOO-style clipped objective. Reverse image search can sometimes return a plausible title and URL for the wrong object. Their proposed approach returns thumbnails, titles, and URLs, allowing the agent to compare each candidate with the queried image region before trusting its associated information. It reports strong performance on benchmarks like MMSearch, MMSearch+, and BrowseComp-VL.

Critical strengths
  1. The key contribution of this paper - Factorized Adaptive Rollout (FAR), which augments group-based RL by expanding the candidate prompt pool and by making more attempts on difficult questions, makes sense intuitively.
  2. Allocating extra attempts to all-wrong groups and stopping redundant sampling is well motivated.
Critical weaknesses

This paper misses several baselines, some are not even cited:

  1. DeepMMSearch-R1: Narayan, K., et al. (2025). DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search. https://arxiv.org/pdf/2510.12801 (missing in the results table comparison)
  2. LMM-Searcher: Du, Y., et al. (2026). Towards Long-horizon Agentic Multimodal Search. https://arxiv.org/pdf/2604.12890
  3. SVRL: Sathish, V., et al. (2026). Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning. https://arxiv.org/pdf/2609.08025
  4. ProMMSearchAgent: Yan, W., et al. (2026). ProMMSearchAgent: A Generalizable Multimodal Search Agent Trained with Process-Oriented Rewards. https://arxiv.org/pdf/2604.20486
  5. Huang, S., et al. (2026). Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents. https://arxiv.org/pdf/2605.10832
Questions for authors

Could you explain the rationale for not including the above baselines? Please try to update the results table by including the above baselines.

4 comments
AuthorsOpenReview ↗

Overview of our response and manuscript revisions

Thank you for your careful review and for recognizing the motivation behind FAR: directing additional attempts to difficult prompts while stopping redundant sampling. We also appreciate your suggestions for broadening the baseline comparison. We address these related concerns together in Comment 1, so the added results and their evaluation context can be read in one place. Manuscript revisions specific to your review are marked in green, with the relevant sections and table identified in the reply.

  1. Added baselines and related work: Present results for the five suggested methods, explain the original omissions and benchmark overlap, and clarify the source protocols and connections to our work.

Please feel free to raise follow-up questions under Comment 1; we will respond promptly.

AuthorsOpenReview ↗

Response to Critical Weakness 1 and Question 1: Added baselines and expanded related-work comparisons

Thank you for drawing our attention to these relevant works. Each contributes to an important aspect of multimodal search, and we agree that their contributions should be properly represented. We have added all five suggested methods to Table 1 and expanded the related-work discussion.

The original main-table selection emphasized benchmark overlap within the available space. We clarify the original coverage and the revisions for each work below.

DeepMMSearch-R1 [1]. Its combination of synthesized search trajectories, SFT, and online RL provides a valuable recipe for learning multimodal search behavior, including adaptive image-region and text queries. We had already cited it in Appendix E.2, but only SimpleVQA overlaps with our six-benchmark suite, which contributed to its omission from the original main table. We have now included that result and expanded the discussion of its search-training approach.

LMM-Searcher [2]. Its file-based visual references and on-demand image loading address an important bottleneck in long-horizon multimodal search: managing visual evidence as the interaction history grows. This work was also cited in Appendix E.2, but its numerical comparison was missing from the original table. Its evaluation overlaps with ours on MMSearch and MMSearch+. We have added its results on the overlapping benchmarks and expanded the discussion of its context-management design.

SVRL [3]. SVRL contributes an RL-only approach to evidence verification and filtering, with feedback on search necessity and query diversity. This is particularly relevant to reliable search: an agent needs not only to retrieve evidence but also to assess whether it supports the answer. Its first arXiv version appeared on September 7, 2026, after our initial submission on September 4, 2026, explaining its absence from that version. We now cite and compare SVRL, and discuss its connection to our visual-verification component.

ProMMSearchAgent [4]. Its process-oriented rewards and transfer from a static search sandbox to live web search address a practical training challenge: making search skills learned in a controlled environment useful in real web interactions. This provides a valuable perspective on both search-process supervision and deployment. Our original discussion emphasized rollout allocation and visual evidence verification, but gave this complementary direction insufficient coverage. We have added its results and a discussion of its reward design and sandbox-to-live transfer.

On-Policy Data Evolution (ODE) [5]. ODE combines reusable visual references with a data generator refined through policy-rollout feedback. It offers a useful perspective on data-efficient training by adapting data construction to the evolving policy rather than treating it as a fixed preprocessing stage. This policy-aware data-construction direction was not adequately represented in our original comparison. We have added its results at both model scales and expanded the discussion of how its data evolution complements the rollout-allocation problem addressed by FAR.

Expanded comparison. The results are grouped by model scale below. “—” denotes an unreported result for that benchmark.

7B/8B models

MethodMMSearchMMSearch+BC-VLFVQALiveVQASimpleVQA
DeepMMSearch-R1-7B (RL) [1]—————55.9
SVRL-full-7B [3]60.3——65.357.858.3
ProMMSearchAgent-7B [4]55.2——63.132.558.6
ODE-8B [5]66.024.941.964.7†—70.3†
SimpleSearch-VL-8B (ours)77.132.552.176.875.276.6

30B models

MethodMMSearchMMSearch+BC-VLFVQALiveVQASimpleVQA
LMM-Searcher-30B, 100 turns [2]72.334.8————
ODE-30B-A3B [5]69.728.146.169.3†—71.0†
SimpleSearch-VL-30B-A3B (ours)83.634.455.979.081.179.6

† ODE uses random 300-example subsets for FVQA and SimpleVQA [5]. LMM-Searcher and ODE evaluate the single-image MMSearch+ subset [2, 5].

The added results and evaluation details are in Table 1 and Section 3.2, and the expanded related-work discussion is in Appendix E.2. Revisions specific to your review are marked in green.

References

[1] Narayan et al. DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search. arXiv:2510.12801v1.

[2] Du et al. Towards Long-horizon Agentic Multimodal Search. arXiv:2604.12890v2.

[3] Sathish et al. Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning. arXiv:2609.08025v1.

[4] Yan et al. ProMMSearchAgent: A Generalizable Multimodal Search Agent Trained with Process-Oriented Rewards. arXiv:2604.20486v1.

[5] Huang et al. Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents. arXiv:2605.10832v3.

Dear Reviewer 9Ltr,

As the discussion phase is nearing its end, we would like to briefly follow up to see if our revision has addressed your concern regarding the missing baselines.

Following your suggestion, we have added all five requested methods (DeepMMSearch-R1, LMM-Searcher, SVRL, ProMMSearchAgent, and ODE) to our main comparison in Table 1 across both 7B/8B and 30B scales, clarified their evaluation protocols in Section 3.2, and expanded the discussion of their connections to our work in Appendix E.2 (marked in green).

Could you please let us know if these additions satisfactorily resolve your concern, or if there is anything else we can clarify? We would deeply appreciate your feedback and any update to your assessment. Thank you again for your helpful suggestions!

Dear Reviewer 9Ltr,

With the discussion period ending soon, we would greatly appreciate it if you could share whether our response resolves your concerns. Thank you very much!

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

The paper proposes SimpleSearch-VL, a data-efficient recipe for training multimodal web-search agents on Qwen3-VL (8B and 30B-A3B), using 5K SFT trajectories and 2K RL samples. The recipe has three components. Factorized Adaptive Rollout (FAR) enlarges the candidate prompt pool, gives extra rollouts only to groups whose answers are all wrong, and stops near completion so that slow tool calls do not stall training. Thumbnail-based visual evidence verification returns thumbnails from reverse image search, so the model can check whether a retrieved result actually matches the queried image before trusting its title or URL. Self-summarized webpage visits let the policy model summarize webpages itself instead of relying on an external summarizer. Across six benchmarks, the method improves the corresponding Qwen3-VL agentic baselines by 15.8 (8B) and 16.0 (30B-A3B) average points, and the 30B model is competitive with agentic Gemini-3-Pro.

Critical strengths
  1. Strong end-to-end results at two scales. Gains in Table 1 are consistent across all benchmarks. Table 17 confirms that the 8B agentic baseline uses the same harness, so the improvement can be attributed to post-training.
  2. Well-motivated rollout design. this paper shows that expanding rollouts uniformly gives diminishing returns. Fig. 4a shows that extra attempts recover 11.7–17.7% of all-wrong groups.
Critical weaknesses
  1. The ablations lack statistical power. All component ablations are single runs on 100-example subsets, with no seeds or variance reported. Some gaps are only 1–2 points
  2. FAR is not ablated in the final setting. The ablations use direct RL, while the main models use SFT+RL. The final model also uses FAR (2×, 4–6) rather than the better-scoring 4–8 configuration.
  3. The thumbnail and self-summary ablations are confounded. Thumbnails are removed only at inference from a model trained with them, so the drop may partly reflect train–test mismatch. Likewise, if the dedicated summarizer is the untuned base model, the self-summary gain may come from training rather than from the design itself.
Questions for authors
  1. How many training seeds were used for each ablation result? Are the 100-example subsets fixed across all compared variants?
  2. fairness. Does the 30B-A3B agentic baseline also use the same harness as SimpleSearch-VL, with identical tool-call limits, search-date settings, and judging protocol? During evaluation, were websites hosting benchmark data (e.g., Hugging Face dataset pages) blocked, to prevent answer leakage through search?
  3. Why does the final model use FAR (2×, 4–6) rather than the 4–8 configuration, which scored higher in the ablation?
8 comments
AuthorsOpenReview ↗

Overview of our response and manuscript revisions

Thank you for your careful review and for recognizing the consistent gains at both model scales and the motivation behind FAR's rollout design. We address your concerns in five separate comments below, covering the experimental controls, final training choices, and evaluation protocol. Manuscript revisions specific to your review are marked in blue, with the relevant sections and tables identified in our replies.

  1. Evaluation reliability: Clarify the training seeds, fixed evaluation subsets, and the time and API costs underlying our evaluation choices.
  2. FAR after SFT: Present the additional SFT-initialized control and its implications for the final SFT+RL recipe.
  3. Final FAR configuration: Explain why we select 4–6 rather than 4–8 based on the accuracy–training-time trade-off.
  4. Thumbnail and self-summary controls: Present the untuned-model thumbnail control and clarify the self-summary comparison, task adaptation, and serving benefits.
  5. Matched evaluation and leakage checks: Clarify the 30B-A3B baseline protocol and the scope of our evaluation-trajectory inspection.

Each numbered item corresponds to the detailed comment with the same number below. Please feel free to raise follow-up questions under the relevant comment; we will respond promptly.

AuthorsOpenReview ↗

Response to Critical Weakness 1 and Question 1: Training seeds, fixed evaluation subsets, and statistical reliability

Thank you for the careful review and for recognizing the consistent improvements at both model scales. We would first like to explain why we use fixed subsets for the component studies. Evaluating a search agent involves multi-turn generation and online tool calls; a full evaluation takes approximately 3–4 hours in our setup and incurs substantial web-search API costs. Because we evaluate every 10 training steps to monitor learning, repeating the full evaluation at this frequency would substantially slow research iteration and generate cumulative API expenses that are difficult to accommodate within our ablation budget.

Our training-time monitoring suite contains 671 questions across six benchmarks: all 171 MMSearch questions in our image-only protocol and a fixed 100-question subset from each of the other five benchmarks. The corresponding evaluation-set sizes and coverage are listed below.

BenchmarkEvaluation-set sizeAblation-set sizeCoverage
MMSearch171171100.0%
MMSearch+31110032.2%
BrowseComp-VL39910025.1%
FVQA1,8001005.6%
LiveVQA3,6021002.8%
SimpleVQA1,0131009.9%

The sizes above refer to the evaluation versions used in our protocol. Individual ablations use the benchmarks listed in their respective tables, rather than necessarily using the entire monitoring suite.

Full-set evaluation would be preferable for the ablations as well. However, repeating it across all variants and intermediate checkpoints would be difficult to sustain within our available time and web-search API budget. We therefore use full-set evaluation for our final-model results in Table 1, while using fixed subsets for the component studies and training-time monitoring.

Sampled evaluation is also used in the main results of related search-agent work: Vision-DeepResearch uses 300 randomly sampled examples per benchmark for FVQA and LiveVQA [1], and ODE uses 300-example subsets for FVQA and SimpleVQA [2]. These examples provide context for the practical trade-off between evaluation coverage and the cost of running search agents.

The subsets are fixed across all compared variants, and each ablation holds the training and evaluation configuration constant except for the factor being studied. We used one training run per configuration with common seed settings. The released configuration specifies a data/shuffling seed of 3407 and a default Megatron base seed of 42; evaluation generates one response per question with temperature 0 and top-p 1.0. We have documented the seed settings, evaluation frequency, and decoding settings in Appendix F.3, with the subset sizes listed in Table 19.

References

[1] Huang et al. Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models. arXiv:2601.22060v1.

[2] Huang et al. Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents. arXiv:2605.10832v3.

AuthorsOpenReview ↗

Response to Critical Weakness 2: FAR remains effective after SFT initialization

Our initial experiments, inspired by SenseNova-MARS, started with direct RL because Qwen3-VL-Instruct already provides a useful starting point for tool use. Studying FAR from this unadapted checkpoint was also a deliberate choice: it allowed us to examine rollout allocation without introducing a particular agentic SFT recipe as an additional factor. The choice and extent of cold-start training can affect subsequent RL, so a common base checkpoint offers a straightforward reference for studying the sampling mechanism.

As the study progressed, however, direct RL alone did not achieve the performance we were seeking. We therefore introduced SFT to give the policy a stronger starting point before RL. This distinction is visible in Table 2 of the main paper, where the SFT-initialized RL model reaches 69.8 average accuracy compared with 64.9 for direct RL. The tool-use analysis in Appendix F.2 is also consistent with SFT providing a stronger prior for sustained search. Thus, the original ablations examined the mechanism from a common base model, while the final recipe incorporated SFT to improve overall capability.

To answer your question directly, we have now repeated the FAR ablation after SFT. All four variants begin from the same SFT checkpoint and follow the direct-RL ablation protocol: 100 training steps, batch size 64, and the same evaluation sets. Only the sampling configuration changes.

MethodAverage accuracyMMSearchBrowseComp-VLFVQA
Standard sampling (4 rollouts per prompt)66.374.953.071.0
Rollout allocation (4–6)66.773.154.073.0
Prompt expansion (2×)68.874.355.077.0
FAR (2×, 4–6)70.777.259.076.0

FAR improves average accuracy from 66.3 to 70.7, a gain of 4.4 percentage points over standard fixed-rollout sampling, with improvements on all three benchmarks. Combining prompt expansion and adaptive rollout allocation also gives higher average accuracy than either change alone. These results show that FAR remains useful after SFT, rather than only compensating for a weaker initialization. The comparison fixes training steps, not cumulative computation; we have added the results to Appendix F.3 (Table 20) and summarized this control in Section 3.3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors of SimpleSearch-VL

Dear Reviewers and Area Chair,

Thank you for your time and feedback. SimpleSearch-VL aims to make strong multimodal search practical under modest training and serving budgets. Its main contributions are:

  • Training cost and search quality. Using 5K SFT trajectories and 2K RL data, our additional 8B H20 run used 296 GPU-hours—about 1/93 of OpenSearch-VL-8B's reported 27.6K H20 GPU-hours with 36K SFT and 8K RL data. These are different training recipes, and the costs cover policy training only. Separately, our H200-trained 8B model improves the four-benchmark average over OpenSearch-VL-8B by 12.0 points (70.3 vs. 58.3).
  • Gains over untuned models. Under matched evaluation protocols within each scale, post-training improves the six-benchmark average over the Qwen3-VL-8B and Qwen3-VL-30B-A3B agentic baselines by 15.8 and 16.0 points, respectively.
  • Efficient RL exploration. FAR explores two-axis dynamic rollout allocation through Prompt Expansion and Rollout Allocation. In the fixed-step direct-RL comparison, it improves the three-benchmark average by 5.4 points over standard sampling while reducing rollout time per step by 14.1%. The gain remains 4.4 points after SFT.
  • Evidence-verified reasoning. Learning to verify visual evidence adds 3.4 average points in the paired SFT comparison, with training examples, tool traces, and final answers held fixed.
  • Self-summary. Removing the auxiliary summarizer reduces evaluation-workload time by 28.5% at the same two-node serving budget, with 69.7 versus 67.9 average accuracy.

Our rebuttal provides further evidence for these contributions and clarifies the design choices, evaluation protocols, and scope of the comparisons:

Reviewer qems

  • Added FAR ablations after SFT and a thumbnail control on an untuned model.
  • Explained the final FAR configuration, self-summary's checkpoint differences, and its serving benefits.
  • Clarified ablation seeds, fixed evaluation subsets and statistical limitations, matched 30B-A3B evaluation, and the scope of the leakage checks.

Reviewer 9Ltr

  • Added the requested baselines and missing citations.
  • Expanded related work and clarified comparison protocols.

Reviewer 7uQq

  • Clarified contributions and comparison fairness, and corrected baseline omissions.
  • Expanded rollout comparisons and clarified the verification ablation.
  • Added training-cost evidence and clarified self-summary's resource trade-offs.

Revisions are marked in blue (qems), green (9Ltr), and red (7uQq).

We would appreciate a brief follow-up on whether these responses address your main concerns, or which issues remain. We would be glad to clarify them, and hope the additional evidence and explanations will inform your updated assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.