acceptodds
Withdrawn from ICLR 2027

SearchEyes: Towards Frontier Multimodal Deep Search Intelligence via Search World Simulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal search agents must reason over long horizons, yet existing pipelines design training data, search environments, and rewards in isolation. This fragmentation discards synthesis metadata, relies on irreproducible external engines,and leaves RL with sparse trajectory-level supervision. We introduce SearchEyes,a knowledge-graph-based simulated search world that unifies these components.Perception-Knowledge Chains (PKC) sample constrained multi-hop paths over visual and knowledge entities while retaining hop-level metadata that defines both a self-contained environment and reward anchors. Hop-Anchored Policy Optimization (HaPO) uses these anchors for step-level credit assignment without a learned process reward model. Across six multimodal knowledge-intensive benchmarks, SearchEyes achieves state-of-the-art performance among opensource multimodal search agents; SearchEyes-27B improves over the strongest open-source baseline by 6.2 points on average.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Search Agents, Data Synthesis, Long-horizon Reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qutPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Authors propose a framework for training multimodal search agents to solve long-horizon, knowledge-intensive tasks more effectively. It unifies three normally separate component: training-data synthesis, the search environment, and reinforcement-learning rewards through a shared knowledge graph. Its Perception-Knowledge Chains generate constrained multi-hop visual and textual questions while preserving entity-level metadata. The same graph creates a deterministic, self-contained search world. Hop-Anchored Policy Optimization then uses retained entities as step-level reward anchors, improving credit assignment.

Critical strengths
  1. Using intermediate entity for training signal is an interesting idea.
Critical weaknesses

Major Weakness:

  1. The authors conveniently skipped the high performing baselines which they cited in the introduction like Vision-DeepResearch [1], REDSearcher [2], HyperEyes [3], MM-DeepResearch [4]
  2. The paper claimed SOTA performance but they don't have SOTA performance. The papers indicated above have better performance. The question the comes : What is the point of having a data generation pipeline that in the end doesn't lead to improved performance.
  3. The data generation pipeline only samples chains of lengths 3,4,5 and other constraint types like triangle, hub, etc are not done. There is no novelty in data generation pipeline and is a mixture of existing works. All parts of the pipeline are pretty standard.
  4. HAPO gains are not properly justified. What happens if a certain rollout doesn't use evidence from the web but is able to identify the entity with parametric knowledge. In that case, HAPO wont consider that rollout in the same group even though it leads to correct answer in the end. How do you ensure, identifying all entities is necessary to come to final answer. I know, the anti-shortcut verification tries to address this but it might not be accurate.
  5. GRPO rollouts might implicitly act as an HAPO group. Can you let me know in how many cases that happen ?
  6. HAPO tries to ensure the model searches for all intermediate entities , but is that the right signal needed for training ? Does searching for all intermediate entities ensure correct answer ? There are cases when the retrieved evidence has the golden entity but the model focuses on some other entity in the retrieved evidence and anchors on that for the next searches. How will you tackle that ? How does HAPO take care of that ?
  7. The ablation study doesn't give any insights. It is just tables of numbers.
  8. The HAPO comparison with Dr.GRPO and GSPO is missing. Can you also report the number of GRPO without the KL loss.

Minor Weakness:

  1. Line 39 is gibberish . Please fix the formatting.
  2. Related work should be in main paper.
  3. Line 790 says 4.X .

[1] Huang, W., Zeng, Y., Wang, Q., Fang, Z., Cao, S., Chu, Z., Yin, Q., Chen, S., Yin, Z., Chen, L. and Chen, Z., 2026. Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models. [2] Chu, Z., Wang, X., Hong, J., Fan, H., Huang, Y., Yang, Y., Xu, G., Zhao, C., Xiang, C., Hu, S. and Kuang, D., 2026. Redsearcher: A scalable and cost-efficient framework for long-horizon search agents. [3] Li, G., Chen, J., Xu, Y., Zhang, X. and Lu, Y., 2026. HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents. [4] Yao, H., Yin, Q., Yang, M., Zhao, Z., Wang, Y., Luo, H., Zhang, J. and Huang, J., 2026. Mm-deepresearch: A simple and effective multimodal agentic search baseline.

Questions for authors

Look at the weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer qutP (Part 1 of 2)

We thank the reviewer for the comments and for recognizing the value of intermediate entities as training signals. We clarify the comparisons, design choices, and optimization mechanism below.

Q1: About Baseline Coverage and Comparison Settings. Table 1 already includes OpenSearch-VL, ODE, SenseNova-MARS, DeepEyes-V2, and WebWatcher. We supplement the comparison with the published results of MM-DeepResearch and Vision-DeepResearch. SearchEyes-27B exceeds MM-DeepResearch-32B on all five shared benchmarks and Vision-DeepResearch-30B-A3B on three of five shared benchmarks.

ModelSimpleVQAVDRMMSearchLiveVQABrowseComp-VLFVQA
MM-DeepResearch-32B67.6-69.068.043.070.1
Vision-DeepResearch-30B-A3B-37.869.677.653.774.2
SearchEyes-27B80.939.482.477.349.379.1

These are results reported in the respective papers; '-' denotes an unreported result. HyperEyes changes the inference action space through joint visual grounding and retrieval, with parallel multi-entity search within a single turn. Its 30B training also uses token-level distillation from a 235B teacher. Consequently, its reported accuracy combines changes in training and inference execution; it is not a matched comparison of data synthesis or policy optimization alone. REDSearcher likewise reports performance from its own end-to-end tool environment and evaluation configuration. Its report specifies a different multimodal backbone and tool-horizon setup, while our controlled comparisons use a shared checkpoint, search world, and rollout budget. The published configurations do not establish a common-tool, common-budget evaluation with SearchEyes. Both systems are relevant system-level references, but their raw rankings alone cannot isolate the effectiveness of our proposed components. We therefore distinguish comparisons between complete agent systems from the controlled component comparisons in Tables 2 and 3, which answer the methodological question directly.

Q2: About Performance Gains and the SOTA Statement. The conclusion that our data-generation pipeline does not improve performance is inconsistent with the controlled results already reported. Table 2 directly compares construction strategies: full PKC achieves a four-benchmark average of 61.8, compared with 49.2 for free-form synthesis, 53.5 for Wikipedia path sampling, and 55.4 for HopChain-style augmentation, improving by 12.6, 8.3, and 6.4 points, respectively. Table 1 further shows that SearchEyes improves the matched agentic backbone from 44.7 to 59.3 at 9B and from 51.1 to 68.1 at 27B under the same tool interface and evaluation loop. These results directly establish the effectiveness of the training framework. We will replace the broad SOTA wording with explicit comparisons to named baselines; the reported 6.2-point average gain specifically refers to OpenSearch-VL-32B.

Q3: About Graph Structures and the Contribution of PKC. Describing PKC as sampling chains of different lengths overlooks its construction constraints and its role in the framework. Section 2.3 explicitly introduces a disambiguation edge that creates a second evidence branch for distinguishing candidate answers. The construction also incorporates multiple visual checkpoints, semantic-domain constraints, information concealment, and anti-shortcut filtering. These structures are deliberately chosen to produce verifiable multimodal search tasks with identifiable intermediate entities. The central contribution is their unified use: the same graph supports task synthesis and the executable search environment, while retained entity metadata supplies anchors for outcome-based credit assignment. Table 2 validates these choices: removing multi-perception interleaving, disambiguation, or anti-shortcut filtering reduces average accuracy by 4.2, 2.5, and 7.7 points, respectively.

Q4: About Parametric Knowledge and Reference Chains. Final-answer correctness remains the reward signal at both levels of HaPO. Every rollout participates in its question-level trajectory group. Entity anchors additionally enable comparisons among rollouts that encounter the same intermediate entity in their retrieved observations. A rollout answering correctly through parametric knowledge receives the correct-answer reward and retains trajectory-level credit. Under the default 0.3 mixing coefficient, positions without a valid anchor retain the trajectory-level component while their hop-level term is zero. During data construction, we verify the reference chain and apply structural and behavioral shortcut checks, including bounded searches for shorter paths and checks for answers obtainable without search or visual input. The chain supplies construction metadata and alignment anchors; matching it is not a requirement for receiving the correct-answer reward.

AuthorsOpenReview ↗

Response to Reviewer qutP (Part 2 of 2)

Q5: About the Relationship Between GRPO and HaPO Groups. The groups can coincide when every rollout for a question reaches the same valid entity anchor. That anchor group then contains the full question-level rollout group, so its outcome mean and standard deviation match those of the trajectory-level group. The corresponding advantage values coincide at positions assigned to that anchor.

More generally, different anchors can select different subsets of rollouts, and the same anchor can be reached at different interaction turns. HaPO explicitly uses these entity-defined groups to compute and assign local outcome-relative advantages. GRPO's question-level grouping alone does not perform this additional computation. Table 3 tests its value: entity-anchored HaPO achieves 61.8, compared with 59.0 for trajectory-level credit alone and 59.1 for turn-index grouping. Group coincidence in particular cases is therefore compatible with the additional benefit of entity-based credit assignment.

Q6: About Entity Retrieval and Correct Reasoning. HaPO uses entity retrieval as an alignment event, while final-answer correctness supplies the reward. Retrieving a reference entity does not itself produce an additional reward. The anchor marks the first point at which the entity appears in a structured tool observation, allowing rollouts with shared evidence availability to be compared. The criterion uses explicit entity identifiers in retrieval records, rather than an arbitrary mention in generated reasoning text.

Consider the reviewer's example: two rollouts retrieve the same target entity, but one follows the relevant evidence and answers correctly, while the other follows a distractor and ultimately answers incorrectly. Both enter the corresponding anchor group. In a group containing successful and unsuccessful outcomes, the successful rollout receives a positive outcome-relative advantage and the unsuccessful rollout receives a negative one. Merely encountering the target entity therefore does not make the failed continuation favorable. The observed retrieval event supplies alignment; the final outcome determines the relative training signal. HaPO does not assign a correctness label to an intermediate reasoning step simply because a reference entity is present in the observation.

Q7: About Insights from the Ablation Studies. We thank the reviewer for this suggestion. We will expand the discussion to explain more concretely what each comparison reveals about the design. The existing results support three specific observations. First, removing anti-shortcut filtering reduces accuracy from 61.8 to 54.1, the largest decrease among individual PKC ablations, supporting explicit verification of the intended evidence requirements. Second, removing multi-perception interleaving or cross-modal grounding reduces accuracy to 57.6 or 58.5, respectively, supporting visual evidence requirements beyond the initial image anchor. Third, replacing entity-anchor grouping with turn-index grouping reduces accuracy to 59.1, supporting alignment by shared evidence rather than interaction position.

The credit-level ablation further clarifies their complementary roles: trajectory-only credit achieves 59.0, hop-only credit achieves 61.1, and their combination achieves 61.8. These comparisons distinguish global outcome supervision from the additional information introduced by anchor-conditioned comparisons. Table 3 also separately examines smooth gating, fatal-aware masking, one-sided clamping, and observation masking. We will connect the quantitative changes more explicitly to task construction, semantic alignment, and stable optimization, and use a concrete rollout example to illustrate how those mechanisms enter training.

Q8: About Additional Optimization Baselines. We thank the reviewer for suggesting Dr. GRPO, GSPO, and GRPO without KL regularization. Table 3 already compares HaPO with GRPO, GiGPO, and DAPO under aligned training conditions, obtaining 61.8 versus 57.8, 58.5, and 58.1, respectively. These optimizer comparisons share the SFT checkpoint, RL question pool, search environment, rollout budget, and update count. The trajectory-only and turn-index variants further isolate the credit-assignment design within our framework. We will extend the comparison to the suggested variants under the same backbone and training conditions. The GRPO comparison with and without KL regularization will specifically examine the effect of that term, while the additional optimizers will broaden the assessment beyond the existing baselines.

Q9: About Formatting and Presentation. We apologize for the typesetting errors and unresolved cross-reference, and thank the reviewer for identifying them. We will correct the affected text and the '4.X' placeholder, check the compiled figures and references, and ensure that the relevant related-work discussion is clearly presented in the main text.

Reviewer_qutPOpenReview ↗

Not Convinced

I thank the reviewer for the rebuttal. I am still not convinced. We can have the same argument that Search eyes uses a different backbone which is Qwen-3.5 compared to other baselines and also different data compared to previous baselines. According to your comments and explanations, your comparison with all the baselines in the paper is not fair, because every paper uses a different harness, tools, train data and base model, and training procedure.

I still feel using intermediate entity is an interesting idea but the performance gains are not there.

Reviewer_j6VMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a framework called SearchEyes, which constructs a simulated search environment that unifying data synthesis, evaluation and RL rollout environment.

Critical strengths
  • This paper clearly presents its unified framework called SearchEyes, which constructs a simulated search environment that unifying data synthesis, evaluation and RL rollout environment. Along with its motivation to build such a simulated world to mitigate the effects of unreliable search engines.
Critical weaknesses

Figure 1: “A knowledge graph serves as the unified backbone for three tightly-coupled stages: PKC question synthesis (left), a self-contained search world for agent interaction (center), and hop-anchored reward signals for policy optimization (right).”

  • positional labels including (left), (center), and (right) in the caption do not match the figure. PKC question synthesis appears in the upper-right corner but is labeled (left). World interaction, also in the upper-right area, is labeled (right), while rewards and policy update appear in the middle-left area but are likewise labeled (right). Even in a fast-paced era, I would expect the authors to exercise the care and rigor that academic work demands.

  • perception-Knowledge Chain Synthesis is essentially multi-hop question construction based on a knowledge graph, a topic extensively explored in prior work [1,2,3]. do not see significant novelty in this component, so presenting it as the paper’s 2nd contribution seems insufficiently convincing.

[1] RedSearcher: REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents

[2] CRIT: Graph-Based Automatic Data Synthesis to Enhance Cross-Modal Multi-Hop Reasoning

[3] WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning

Questions for authors

please address the critical weaknesses first.

1 comment
AuthorsOpenReview ↗

Response to Reviewer j6VM

We thank the reviewer for recognizing the unified search-world formulation and its motivation. We address the figure inconsistency and clarify how the proposed contribution relates to prior graph-based synthesis and simulated search environments.

Q1: About Figure 1 and Presentation. We sincerely apologize for the mismatch between Figure 1 and its caption. The figure layout was adjusted, but the positional descriptions were not updated consistently. This was an oversight in our final presentation check. We will correct the caption to match the figure, remove ambiguous positional references, and carefully check the remaining figures, labels, and cross-references in the next revision.

Q2: About the Contribution of PKC and the Unified Search World. The contribution should be evaluated as a unified multimodal training design. Graph-based multi-hop question construction is an established ingredient. SearchEyes uses a shared graph and retained entity identities to connect question construction, an executable multimodal search environment, and outcome-based credit assignment. These connections specify how the same structural information is reused throughout training.

Concretely, PKC constructs questions with multiple visual checkpoints, a disambiguating evidence branch, domain constraints, information concealment, and anti-shortcut checks. The graph's linked entity documents and images then define the self-contained search world. Text retrieval combines dense and lexical rankings; visual search operates on the input image or images exposed during interaction. Structured tool observations retain canonical entity identifiers. HaPO uses these identifiers to align rollouts when the same reference entity first becomes available, even if their queries, retrieved snippets, or interaction turns differ. It then computes relative advantages from final-answer outcomes within the corresponding groups.

The distinctions from the cited works are therefore specific. REDSearcher develops scalable search-agent training with synthetic data and simulated search infrastructure, including multimodal search. WebSailor-V2 combines graph-based synthesis with a simulated search environment for text-based agents. CRIT studies graph-based synthesis for cross-modal multi-hop reasoning. These are relevant precedents for the individual ingredients. The shared presence of graph synthesis or simulation does not establish equivalence to the coupling in SearchEyes: synthesized entity metadata is directly connected to structured multimodal observations and reused as the grouping variable for policy optimization.

This coupling has two practical consequences. First, generated questions, reference entities, and executable evidence come from a common world, allowing us to replay teacher trajectories and verify evidence availability before exporting SFT examples. The student is trained on the raw observations from that replay. Second, retained entity identities provide an observable alignment signal during RL, so semantically corresponding retrieval progress can be compared without requiring identical queries or a separately trained process reward model. Final-answer correctness remains the reward; metadata determines how that outcome signal is assigned.

The ablations evaluate these connections directly. Table 2 compares full PKC with free-form synthesis, Wikipedia path sampling, and HopChain-style augmentation, obtaining 61.8 versus 49.2, 53.5, and 55.4 on the four-benchmark average. Removing multi-perception interleaving lowers accuracy by 4.2 points, and removing the disambiguation constraint lowers it by 2.5 points. Table 3 isolates the optimization design: entity-anchor grouping reaches 61.8, compared with 59.1 for turn-index grouping and 59.0 for trajectory-level credit alone. These comparisons demonstrate the utility of both the task constraints and the reuse of entity metadata.

The simulation is used for training, while the reported benchmark evaluation uses external search tools. The gains therefore measure transfer from the constructed world to the benchmark search setting. Under the matched agentic interface, SearchEyes improves the 9B backbone from 44.7 to 59.3 and the 27B backbone from 51.1 to 68.1. This provides an additional check that the constructed curriculum and optimization signals support useful search behavior beyond interaction with the training environment itself.

We will merge the first two contribution statements to present PKC and the simulated search world as one integrated contribution, and make this relationship to prior work explicit. This will focus the presentation on the unified construction-environment-optimization design, with HaPO explaining how the retained structure becomes an effective training signal.

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

The paper introduces a unified framework for training multi-hop visual search agents. It achieves SOTA results among open multimodal search agents on six knowledge-intensive benchmarks. It first constructs a knowledge graph linking Wikidata entities, textual descriptions, and images. Then, it builds a self-contained search environment and samples multi-hop paths over the visual-knowledge intersection graph. Its post-training pipeline comprises SFT followed by HaPO, which performs step-level credit assignment.

Critical strengths
  1. The paper presents a unified framework that integrates knowledge-graph-based training data synthesis, search environment simulation, and a post-training recipe, providing a systematic approach to training multimodal search agents.

  2. The proposed PKC pipeline generates high-quality multi-hop training data, while the self-contained search environment enables more efficient and reproducible multimodal search agent training.

  3. The experiments are comprehensive, covering six multimodal knowledge-intensive benchmarks and two model scales (9B and 27B). The results demonstrate strong improvements over existing open-source multimodal search agents, with detailed ablation studies validating the effectiveness of the proposed components.

Critical weaknesses
  1. The paper has significant formatting issues, particularly with citations. Many citations are incorrectly formatted and disrupt the readability of the text.

  2. Several benchmarks have relatively small evaluation sets, and the experiments do not report results across multiple runs. Given that some of the reported improvements are relatively small, reporting variance would be helpful.

  3. The proposed PKC synthesis constraints do not necessarily guarantee multi-step visual reasoning. Although the sampled paths contain multiple perception nodes, the agent may rely entirely on textual knowledge after identifying the image at the initial node, without using the visual information associated with the perception nodes.

Questions for authors
  1. HaPO relies on a predefined gold entity path. However, a search task may have multiple valid reasoning paths leading to the correct answer. How does HaPO handle successful trajectories that follow alternative paths without matching the predefined gold entity anchors?

  2. Since the gold entity paths are already available during training, why not directly use them to construct explicit step-wise rewards, such as rewarding the agent for retrieving correct intermediate entities? Have the authors compared this approach with the proposed HaPO method?

1 comment
AuthorsOpenReview ↗

Response to Reviewer FFLT

We thank the reviewer for recognizing the unified framework, the quality of the synthesized training data, and the breadth of the experiments. We address the presentation, robustness, and methodological questions below.

Q1: About Citation Formatting. We sincerely apologize for the citation-formatting problems. They arose from errors in the LaTeX compilation and typesetting of the submission. We will correct the citation rendering and check the compiled text, figures, and references throughout the manuscript.

Q2: About Repeated Evaluations. We thank the reviewer for this suggestion. We have conducted three repeated evaluations of SearchEyes-27B under the same evaluation protocol. The per-benchmark mean accuracy and standard deviation are reported below, in percentage points.

BenchmarkMean accuracy ± standard deviation
SimpleVQA80.9 ± 0.1
VDR39.4 ± 0.2
MMSearch82.4 ± 0.3
LiveVQA77.3 ± 0.3
BrowseComp-VL49.3 ± 0.5
FVQA79.1 ± 0.2

These repeated evaluations quantify the variability of the reported benchmark accuracy. We will include the run count and per-benchmark statistics alongside the evaluation protocol in the next manuscript revision.

Q3: About Multiple Perception Nodes. Multiple perception nodes are deliberately introduced to distribute visual evidence along the search task. The initial image identifies the starting entity; downstream image-backed entities expose additional visual evidence only after the corresponding records have been retrieved or opened. Information concealment removes entity names and exact relation labels, while visual-grounding and anti-shortcut checks favor questions that require the intended evidence gathering. This design trains the agent to return to visual evidence as the search progresses, preserving interleaved perception and knowledge retrieval rather than reducing the task to initial image recognition followed by purely textual lookup.

The P/K labels describe the evidence structure of the synthesized task, not a fixed sequence of tool calls. During SFT trajectory synthesis, we retain correct expert trajectories and replay their calls in the unmodified environment, checking that the evidence used in later reasoning is available in the preceding observations. This supplies demonstrations consistent with the intended search structure while preserving flexibility in query wording and interaction order. Table 2 directly supports the design: removing multi-perception interleaving lowers the four-benchmark average from 61.8 to 57.6; removing cross-modal grounding lowers it to 58.5.

Q4: About Alternative Successful Paths. Our training questions are constructed around a verified reference entity chain, with answer disambiguation and complementary anti-shortcut checks. The reference chain provides auditable construction metadata and anchors for comparing retrieval progress. It does not prescribe an exact sequence of queries or tool calls: different queries, rankings, and observations can expose the same entity at different turns.

Final-answer correctness remains the reward signal. Every rollout participates in the question-level trajectory group, including successful rollouts that use another route or identify an entity through parametric knowledge. HaPO adds an anchor-conditioned term to this trajectory-level credit. With the default mixing coefficient of 0.3, positions without a valid anchor retain the trajectory-level component while the hop-level component is zero. Such a successful rollout therefore receives the correct-answer reward and remains part of the learning objective; matching every reference entity is not a condition for receiving that reward.

Q5: About Explicit Intermediate-Entity Rewards. Entity anchors and correctness rewards play different roles in our design. An entity's first appearance in a structured retrieval record identifies a shared evidence-availability point across rollouts. The advantage within that group is computed from final-answer outcomes. Simply retrieving the entity produces no additional reward. For example, two rollouts may retrieve the same entity, but one reasons correctly from it while the other follows a distractor and fails. In a mixed-outcome anchor group, their outcome-relative advantages have opposite signs. An entity-hit bonus alone would reward both retrieval events without making this distinction.

We use the reference entities to refine outcome-based credit assignment while preserving final-answer correctness as the objective. Table 3 evaluates trajectory-only, hop-only, and combined credit, obtaining 59.0, 61.1, and 61.8, respectively; turn-index grouping obtains 59.1. These comparisons support combining global outcome credit with semantic alignment. We appreciate the suggestion to compare against an explicit entity-hit reward and will include that additional comparison to examine the distinction directly.

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

Related papers

Discussion (0)

Sign in to comment.