- 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
- Using intermediate entity for training signal is an interesting idea.
Critical weaknesses
Major Weakness:
- 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]
- 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.
- 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.
- 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.
- GRPO rollouts might implicitly act as an HAPO group. Can you let me know in how many cases that happen ?
- 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 ?
- The ablation study doesn't give any insights. It is just tables of numbers.
- The HAPO comparison with Dr.GRPO and GSPO is missing. Can you also report the number of GRPO without the KL loss.
Minor Weakness:
- Line 39 is gibberish . Please fix the formatting.
- Related work should be in main paper.
- 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
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.
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.
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.
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.