acceptodds
Under review as a conference paper at ICLR 2027

Learning to Route in Visual Space via Multi-Step Embedding Retrieval

Abstract

LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHop, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHop-router, an end-to-end training pipeline—combining supervised fine-tuning, online imitation learning, and reinforcement learning—that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHop-router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHop-router boosts retrieval performance from under 5% to 76.3%. In agentic search, it improves task success rates by 52.7% and reduces the average token length by 61% from 1886 to 728, whereas upgrading the agent yields only a 3.7% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHop-router maintains superior performance while reducing in-context images by and cutting the cumulative API payload by . The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHop and VHop-router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Embedding, information retrieve, agentic search

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ye8bOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work examines how multi-hop visual search can be improved by moving some of the sequential decision-making responsibilities from language-model agents into the retrieval component. To investigate this problem, the authors construct VHOP, a benchmark with five progressively challenging levels that test visual correspondence, spatial reasoning, and the ability to plan retrieval sequences. Building on this benchmark, they develop VHOP-ROUTER, a retrieval model that generates navigation decisions autoregressively. Rather than requiring an LLM to issue a new textual query at every step, the proposed model operates directly over visual representations, choosing subsequent images, recovering from unsuccessful retrieval paths, and determining when enough evidence has been collected. Its training procedure consists of supervised fine-tuning, online imitation learning, and reinforcement learning.

The experimental study considers both the retriever as an independent component and its integration into an agent-based question-answering system. For the L4 difficulty level, VHOP-ROUTER obtains a retrieval success rate of 76.3%. When incorporated into the agent, it raises end-to-end task success from 32.7% with a standard retriever to 85.4%. The paper further examines token efficiency, compares the proposed approach against more capable agent configurations and retrieval alternatives, and investigates its behavior across different levels of task complexity and in settings intended to be closer to practical visual search.

Critical strengths

1. An interesting architectural approach that gives the retriever greater autonomy in multi-step visual search.

A notable contribution of this work is its reconsideration of how responsibilities should be divided between the language agent and the retrieval system. Instead of having the agent coordinate every intermediate search through natural-language instructions, VHOP-ROUTER learns to traverse visual representations through a sequence of retrieval decisions. Its ability to select candidates, reconsider earlier choices, and terminate a search allows several planning operations to be handled within the retriever. This design directly targets the limitations of expressing complex visual dependencies as textual queries and offers an alternative way to organize multi-hop retrieval systems.

2. Substantial performance gains accompanied by meaningful baseline comparisons.

The experimental findings provide strong evidence for the effectiveness of the proposed approach within the constructed evaluation environment. Most notably, the end-to-end success rate on L4 tasks increases from 32.7% to 85.4% when the baseline retrieval component is replaced by VHOP-ROUTER, alongside a considerable reduction in token usage. The authors also examine larger candidate pools, stronger agent configurations, and tasks with varying levels of difficulty. These experiments strengthen the argument that improving the retriever's ability to make sequential decisions can offer significant benefits, rather than relying exclusively on a more powerful language agent to manage the search process.

Critical weaknesses

1. The experimental results do not yet establish reliable performance in realistic search environments.

The most convincing improvements are reported on synthetic tasks where scene composition, object relationships, and sources of difficulty are explicitly controlled. The additional evaluation intended to approximate real-world search is relatively small, covering only 50 queries, and its images are re-rendered rather than directly sampled from real-world photography. The results from this setting also reveal a substantial limitation: although the correct evidence is retrieved in approximately 34–36% of cases, the complete agent answers only 20% of questions correctly. This discrepancy indicates that successfully locating relevant visual content does not necessarily translate into reliable task completion. Consequently, the paper provides insufficient evidence that its reported advantages will carry over to visually diverse, less structured, and more challenging practical applications.

2. The learned navigation strategy appears to struggle when the required retrieval sequence becomes longer.

The four-hop evaluation in Appendix F.2 reveals a particularly concerning limitation. None of the investigated training configurations exceeds 8.2% accuracy, and the standard precise mixed-hop configuration achieves just 1.0%. Admittedly, these experiments change more than retrieval depth, since the search corpus is also enlarged and additional distractors are introduced. Nevertheless, the magnitude of the performance decline suggests that the proposed policy may not transfer effectively to more demanding retrieval conditions. This is especially important because learning to make sequential retrieval decisions is central to the paper's contribution. Without stronger evidence that this capability remains effective over longer search chains, the claimed generality and scalability of the framework remain difficult to establish.

Questions for authors

1. [Clarification] Understanding the limited performance on more realistic tasks.

Appendix F.3 reports approximately 34–36% retrieval success but only 20% final-answer accuracy in the realistic evaluation. Could the authors further investigate where the system fails, particularly separating unsuccessful evidence retrieval, premature or inappropriate stopping, and mistakes made during answer generation? It would also be helpful to explain what these findings imply for the framework's applicability outside the synthetic benchmark.

2. [Clarification / Additional Evaluation] Identifying the causes of four-hop performance degradation.

According to Appendix F.2, accuracy drops to 8.2% or below across all configurations on four-hop tasks, while the default configuration achieves only 1.0%. However, these tests simultaneously involve a larger retrieval corpus and more distracting candidates. Can the authors determine whether the main difficulty comes from longer decision sequences or from the increased complexity of the search space? An experiment that varies hop count independently of corpus size and distractor difficulty would provide more direct evidence about the model's ability to handle longer retrieval chains.

3. [Additional Evaluation] Separating autonomous retrieval from continued agent intervention.

The proposed retriever handles multiple retrieval decisions internally, but the surrounding LLM agent can still reformulate its requests and initiate further searches. To better identify the source of the performance improvement, could the authors evaluate a variant in which the agent is allowed to use VHOP-ROUTER only once, without subsequent query revision or repeated retrieval calls? Keeping the final-answer generation procedure unchanged would make it easier to assess how much of the improvement comes from the retriever's learned planning behavior rather than additional reasoning and intervention by the agent.

4. [Clarification] Evaluating computational efficiency beyond token consumption.

The reported savings in generated tokens and image inputs are encouraging, but these measurements may not capture the full computational overhead of the proposed system. Unlike a conventional retrieval call, VHOP-ROUTER may execute several internal decision steps before returning its results. Could the authors compare the methods using wall-clock latency, the total number of embedding computations, and overall computational expenditure under equivalent search budgets? Such measurements would help determine whether the proposed approach reduces total inference cost or mainly reallocates computation from the language agent to the retrieval component.

5. [Additional Evaluation] Measuring the individual effects of the training stages.

VHOP-ROUTER is developed through three successive stages: supervised fine-tuning, online imitation learning, and RLVR. Although the trajectory examples offer some insight into the learned behavior, the quantitative contribution of each stage remains insufficiently clear. Could the authors evaluate checkpoints obtained after each stage using the same test conditions and report retrieval success, backtracking accuracy, stopping accuracy, and end-to-end task performance? This comparison would help establish which capabilities emerge at each stage and whether all three stages are necessary to obtain the reported results.

8 comments
AuthorsOpenReview ↗

Thanks for your review (1/n)

We sincerely thank Reviewer Ye8b for the thorough, high-confidence review and for recognizing how moving sequential navigation into the retriever reshapes the agent–retriever division of labor with substantial empirical gains. To directly address your weaknesses and five questions, we added natural Flickr photo evaluations, a 50-query realistic trajectory decomposition, fixed-gallery and 4-hop diagnostics, a strict single-call agent experiment, wall-clock GPU/API profiling, and a stage-wise checkpoint evaluation. Below, we address each item point by point.

Weakness 1

"1. The experimental results do not yet establish reliable performance in realistic search environments.

Response:

TL;DR: To directly evaluate generalization on naturally occurring photographs, we added three unmodified Flickr photo benchmarks (YFCC-VHop, DISBench-MVH, and Official DISBench [1]; Section 5.6, Figure 7, Table 8; Appendix F.5, Table 25 and Figure 14): across cluttered personal photo albums (about 750–1,900 real photos per user), VHop-Router raises multi-hop standalone Hit@1 from 11.6% → 53.6% (+42.0 pp) and Gemini 3.5 Flash agent Top-1 accuracy from 44.9% → 55.1% (+10.2 pp) on YFCC-VHop, while transferring zero-shot to Official DISBench and DISBench-MVH. Meanwhile, examining all 50 trajectories on the realistic test set (Question 1, Appendix F.3, Table 23) shows that premature stopping is rare (8%, 4/50) and all 7/50 (14%) cases separating retrieval (34%) from final accuracy (20%) occur after the target image is already returned to Flash.

Experiment Setup:

  • Constructing and Evaluating Real-Photo Benchmarks: Auditing all 122 Official DISBench queries (57 Flickr users, 109k photos) showed that 46/122 require no cross-photo visual links (relying on EXIF time/GPS metadata or single-image attributes) and the remaining 76/122 involve only single-hop visual links (no multi-hop visual chains). Along with evaluating Official DISBench, to directly test multi-hop visual link capabilities on real photos, we followed DISBench's mining workflow (recurring entity grouping → co-occurring object graph walk → shortcut rejection against earlier chain photos → bridge-dependent query composition, scored via TREC-style pooled judging) to build two real-photo benchmarks where every hop requires visual links: (1) YFCC-VHop over 78 YFCC100M Flickr users strictly disjoint from DISBench (about 750 real photos/album; 2 of 80 candidate users had zero multi-hop chains after filtering), split by user into 66 training users (590 1–3-hop queries, used to train VHop-Router following our exact same 3-stage pipeline: SFT → Online IL → RLVR) and 12 held-out test users (290 1–3-hop test queries, including n = 69 2–3-hop multi-hop queries); and (2) DISBench-MVH (Multi-Visual-Hop extension; 109 zero-shot 1–3-hop queries on Official DISBench user albums, including n = 38 multi-hop queries of which n = 25 are 2-hop; zero training on DISBench users), extending Official DISBench queries verbatim as Hop 1 with verified 2nd/3rd visual hops. We evaluated this YFCC-VHop-trained router on held-out YFCC-VHop test users and zero-shot inside a shared metadata+search agent (Gemini 3.5 Flash) on Official DISBench (n = 57 Intra-Event queries) and DISBench-MVH.
  • 50-Query Realistic Set Trajectory Analysis: Examined all 50 human-authored trajectories over 386 images (Section 5.6, Appendix F.3; full breakdown in Question 1).

Results:

  • Unmodified Real Photographs (YFCC-VHop, Official DISBench, and DISBench-MVH, Appendix F.5, Table 25): On held-out YFCC-VHop multi-hop test queries (n = 69), VHop-Router raises standalone Hit@1 over baseline Qwen3-VL-2B from 11.6% → 53.6% (8/69 → 37/69, +42.0 pp) and Gemini 3.5 Flash agent Top-1 accuracy from 44.9% → 55.1% (31/69 → 38/69, +10.2 pp). Transferred zero-shot, it lifts Official DISBench Intra-Event Any-Hit recall (75.4% → 80.7%, 43/57 → 46/57) and DISBench-MVH 2-hop standalone Hit@1 (24.0% → 36.0%, 6/25 → 9/25) / multi-hop agent Any-Hit (76.3% → 81.6%, 29/38 → 31/38).
  • 50-Query Realistic Set (Section 5.6, Table 8 and Appendix Tables 22–23): Flash + VHop-Router improves retrieval success over Flash + Qwen3 (36.0% vs. 26.0%, 18/50 vs. 13/50) and more than doubles final answer accuracy (20.0% vs. 8.0%, 10/50 vs. 4/50) in 3.3× fewer tool calls (4.2 vs. 14.0; detailed decomposition in Question 1).

Reference: [1] DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories, 2026.

AuthorsOpenReview ↗

Thanks for your review (2/n)

Weakness 2

"2. The learned navigation strategy appears to struggle when the required retrieval sequence becomes longer.

The four-hop evaluation in Appendix F.2 reveals a particularly concerning limitation. None of the investigated training configurations exceeds 8.2% accuracy, and the standard precise mixed-hop configuration achieves just 1.0%. Admittedly, these experiments change more than retrieval depth, since the search corpus is also enlarged and additional distractors are introduced. Nevertheless, the magnitude of the performance decline suggests that the proposed policy may not transfer effectively to more demanding retrieval conditions. This is especially important because learning to make sequential retrieval decisions is central to the paper's contribution. Without stronger evidence that this capability remains effective over longer search chains, the claimed generality and scalability of the framework remain difficult to establish."

Response:

TL;DR: Three findings clarify transfer to longer retrieval chains: 1) Within the trained horizon (1–3 hops), dynamic variable-hop search works reliably: holding corpus size and distractors strictly fixed (Question 2), one single VHop-Router checkpoint dynamically handles unknown 1-, 2-, and 3-hop queries (99.7% / 90.0% / 76.3%, Section 5.4) without a depth oracle (knowing the ground-truth hop count H at test time, which prefix-conditioned Visual MDR [3] requires because it has no stopping head)—including on vague prompts and real YFCC-VHop photo chains (Weakness 1). 2) On out-of-horizon 4-hop queries (beyond 3 hops), while length extrapolation is a broad open challenge across sequence models [2] and multi-hop retrievers [3, 4], our component diagnosis shows that the main bottleneck is the lightweight stopping MLP head—specifically, distribution mismatch on its 2048-dim sequence-state input at unseen 5-image chain lengths, whereas local cosine-similarity drop features remain sharp at hop 4 (Question 2). 3) VHop-Router outperforms Visual MDR on 4-hop under both non-oracle score-drop stopping (24.7% vs. 16.5%) and oracle-depth K = 4 stopping (36.1% vs. 25.8%) (Appendix F.2, Table 21).

Experiment Setup:

  • Dynamic Variable-Hop Control (1–3 Hops, Section 5.4 and Appendix E.1): Evaluated one single VHop-Router (Mixed-Hop) checkpoint across 1-, 2-, and 3-hop queries on the fixed 450-image L4 corpus (n = 300/hop, identical distractors, no depth oracle), under both precise and vague prompts (Question 2).
  • Out-of-Horizon 4-Hop Diagnostic vs. Visual MDR (Appendix F.2, Tables 20–21, n = 97, 600 images): Evaluated matched 3-Hop Visual MDR [3] and VHop-Router checkpoints zero-shot on 4-hop L4 queries under both non-oracle 1D cosine-score-drop stopping (calibrated strictly on 1–3 hop dev queries, zero knowledge of K = 4) and matched oracle-depth K = 4 stopping across Beam = 1 and Beam = 4.

Results:

  • Dynamic Hop Feasibility (1–3 Hops): Holding the 450-image corpus constant, Flash + Qwen3 drops sharply across 1 / 2 / 3 hops (75.7% → 66.7% → 32.7%, Appendix Tables 12a, 13a, and Section 5.1 Table 4a), whereas one precise Mixed-Hop VHop-Router checkpoint sustains 99.7% / 90.0% / 76.3% standalone without a depth oracle (Section 5.4, Tables 12a, 13a, 4a; and our vague-trained Mixed-Hop checkpoint sustains 97.0% / 84.7% / 50.7% vs. 21.3% / 6.7% / 2.0% with Flash on vague prompts, Tables 12d, 13d, 4b).
  • 4-Hop Extrapolation Diagnostic (n = 97, Appendix F.2, Table 21): Decoupling stopping from the sequence state via the 1–3-hop-calibrated cosine-score drop (no K = 4 oracle, Beam = 4) lifts VHop-Router from 1.0% (Table 20) to 24.7% vs. 16.5% for Visual MDR, and under oracle-depth K = 4 stopping VHop-Router doubles Visual MDR at Beam = 1 (10.3% vs. 5.2%) via learned BACKTRACK and leads by +10.3 pp at Beam = 4 (36.1% vs. 25.8%; Mixed-Hop encoder: 32.0% vs. 29.9%):
3-Hop Trained Method on 4-Hop L4 (n = 97, Table 21)Beam = 1 (K = 4 Oracle)Beam = 4 (K = 4 Oracle)Beam = 4 (No Oracle, Score-Drop Stop)
Visual MDR5.2%25.8%16.5%
VHop-Router10.3%36.1%24.7%

References: [2] Anil et al., Exploring Length Generalization in Large Language Models, NeurIPS 2022. [3] Xiong et al., Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval, ICLR 2021. [4] Trivedi et al., MuSiQue: Multihop Questions via Single-Hop Question Composition, TACL 2022.

AuthorsOpenReview ↗

Thanks for your review (3/n)

Question 1

"1. [Clarification] Understanding the limited performance on more realistic tasks.

Appendix F.3 reports approximately 34–36% retrieval success but only 20% final-answer accuracy in the realistic evaluation. Could the authors further investigate where the system fails, particularly separating unsuccessful evidence retrieval, premature or inappropriate stopping, and mistakes made during answer generation? It would also be helpful to explain what these findings imply for the framework's applicability outside the synthetic benchmark."

Response:

TL;DR: Examining all 50 trajectories on the realistic test set (Section 5.6, Appendix F.3, Table 23) reveals three main takeaways: 1) Premature stopping by VHop-Router is rare (8%, 4/50); 2) 100% of the 34.0% → 20.0% gap (7/50 queries) comes from Gemini Flash's final answer selection after VHop-Router already returned the correct target image (6/7 picking a different wide-angle multi-object room view); and 3) In-domain real-photo training (YFCC-VHop, Weakness 1) directly resolves the synthetic-to-photo entry shift, lifting real-photo multi-hop standalone Hit@1 from 11.6% → 53.6% (+42.0 pp).

Experiment Setup: Examined every trajectory across all 50 human-authored queries over the 386 images of the realistic test set (Section 5.6, Appendix F.3, Tables 22–23) for Flash + VHop-Router (Mixed-Hop) (34.0% set retrieval, 17/50; 36.0%, 18/50 for 3-Hop) vs. Flash + Qwen3 (26.0%, 13/50), partitioning outcomes into five mutually exclusive categories: (1) correct final answer, (2) target returned in cumulative chain outputs but wrong image selected by Flash, (3) premature STOP, (4) mid-chain derailment, and (5) Hop-1 entry miss.

Results:

  • 50-Query Failure Decomposition (Flash + VHop-Router Mixed-Hop): 20.0% (10/50) answered correctly (vs. 8.0%, 4/50 for Flash + Qwen3, in 3.3× fewer calls: 4.2 vs. 14.0); 14.0% (7/50; 8/50 for 3-Hop) returned the true target to the agent—accounting for the 17/50 → 10/50 (34.0% → 20.0%) retrieval-to-answer difference—where Flash picked a different cluttered multi-object room image (6/7 cases); premature stopping is only 8.0% (4/50; 2/50 = 4.0% for 3-Hop); 30.0% (15/50) derailed mid-chain; and 28.0% (14/50) missed Hop 1 (reflecting synthetic-to-everyday perceptual shift on query views).
  • Implications Outside Synthetic Benchmarks: VHop-Router lowers combined mid-chain + early-stop attrition by 32%–37% relative (56%–60% on single-step baselines vs. 38%, 19/50), and adapting on natural multi-object photographs (YFCC-VHop, Weakness 1) lifts real-photo multi-hop retrieval from 11.6% → 53.6% standalone (n = 69) and 44.9% → 55.1% end-to-end.
Reviewer_V4DtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes VHOP, a benchmark for multi-hop visual retrieval, and VHOP-ROUTER, which shifts multi-step search from the LLM agent into an autoregressive embedding retriever operating directly in visual latent space. The authors use SFT, online imitation learning, and RLVR for training and substantially improvie retrieval/agent success while reducing token and image-context costs.

Critical strengths
  1. The idea of making the retriever stronger rather than making an llm agent query a weak retriever multiple times is sound and is a step in the right direction.
  2. The experimental results are large, which shows the method is empirically sound.
Critical weaknesses
  1. The empirical gains are large but real-world validation is extremely limited. The "realistic" evaluation contains only 50 questions/386 images, where scenes are re-rendered rather than taken from naturally occurring images. Also, we see a drop to only 34-36% , so the practical generalization claim is not supported.
  2. The baseline comparison is not fair. VHOP-router is extensively trained for this multi-hop task, where the other retrieval baselines are largely standard Qwen3/GE2. Can you compare it with same training data and supervision budget ? I am not sure whether the gain comes because of the routing mechanism or because of task-specific training.
  3. The authors mention in limitations that the method is not scalable and doesn't perform well on hop counts outside those observed during training. I appreciate authors acknowledging this but then the work has really low impact because of this limitation. Real-world scenarios would have variable hops. How do you think you will address this issue ?
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Thanks for your review (1/n)

We sincerely thank Reviewer V4Dt for the constructive review and for recognizing that strengthening the retriever rather than having an LLM agent repeatedly query a weak retriever is "sound and is a step in the right direction" with large empirical gains. To directly address your concerns, we added natural Flickr photo evaluations, head-to-head trained baselines, and same-checkpoint routing ablations. Below, we address each point in detail.

Weakness 1

"The empirical gains are large but real-world validation is extremely limited. The "realistic" evaluation contains only 50 questions/386 images, where scenes are re-rendered rather than taken from naturally occurring images. Also, we see a drop to only 34-36% , so the practical generalization claim is not supported."

Response:

TL;DR: To directly evaluate generalization on naturally occurring photographs, we added three unmodified Flickr photo benchmarks (YFCC-VHop, DISBench-MVH, and Official DISBench [1]; Section 5.6, Figure 7, Table 8; Appendix F.5, Table 25 and Figure 14). Across cluttered personal photo albums (about 750–1,900 real photos per user), VHop-Router raises multi-hop standalone Hit@1 from 11.6% → 53.6% (+42.0 pp) and Gemini 3.5 Flash agent Top-1 accuracy from 44.9% → 55.1% (+10.2 pp) on YFCC-VHop, while transferring zero-shot to both Official DISBench and DISBench-MVH—complementing its 2.5× final-accuracy improvement (20.0% vs. 8.0% in 3.3× fewer calls) on the 50-query realistic set (Section 5.6).

Experiment Setup:

  • Constructing and Evaluating Real-Photo Benchmarks: Auditing all 122 Official DISBench queries (57 Flickr users, 109k photos) showed that 46/122 require no cross-photo visual links (relying on EXIF time/GPS metadata or single-image attributes) and the remaining 76/122 involve only single-hop visual links (no multi-hop visual chains). Along with evaluating Official DISBench, to directly test multi-hop visual link capabilities on real photos, we followed DISBench's mining workflow (recurring entity grouping → co-occurring object graph walk → shortcut rejection against earlier chain photos → bridge-dependent query composition, scored via TREC-style pooled judging) to construct two real-photo benchmarks where every hop requires visual links: (1) YFCC-VHop over 78 YFCC100M Flickr users strictly disjoint from DISBench (about 750 real photos/album; 2 of 80 candidate users had zero multi-hop chains after filtering), split by user into 66 training users (590 1–3-hop queries, used to train VHop-Router following our exact same 3-stage pipeline: SFT → Online IL → RLVR) and 12 held-out test users (290 1–3-hop test queries, including n = 69 2–3-hop multi-hop queries); and (2) DISBench-MVH (Multi-Visual-Hop extension; 109 zero-shot 1–3-hop queries on Official DISBench user albums, including n = 38 multi-hop queries of which n = 25 are 2-hop; zero training on DISBench users), extending Official DISBench queries verbatim as Hop 1 with verified 2nd/3rd visual hops. We evaluated this YFCC-VHop-trained router on held-out YFCC-VHop test users and zero-shot inside a shared metadata+search agent (Gemini 3.5 Flash) on Official DISBench (n = 57 Intra-Event queries) and DISBench-MVH.
  • 50-Query Realistic Set Trajectory Analysis: Examined all 50 human-authored trajectories over 386 images (Section 5.6, Appendix F.3, Tables 22–23) to separate retrieval misses, stopping behavior, and VLM answer selection.

Results:

  • Unmodified Real Photographs (YFCC-VHop, Official DISBench, and DISBench-MVH, Appendix F.5, Table 25): On held-out YFCC-VHop multi-hop test queries (n = 69), VHop-Router raises standalone Hit@1 over baseline Qwen3-VL-2B from 11.6% → 53.6% (8/69 → 37/69, +42.0 pp) and Gemini 3.5 Flash agent Top-1 accuracy from 44.9% → 55.1% (31/69 → 38/69, +10.2 pp). Transferred zero-shot, it lifts Official DISBench Intra-Event Any-Hit recall (75.4% → 80.7%, 43/57 → 46/57) and DISBench-MVH 2-hop standalone Hit@1 (24.0% → 36.0%, 6/25 → 9/25) / multi-hop agent Any-Hit (76.3% → 81.6%, 29/38 → 31/38).
  • 50-Query Realistic Set (Section 5.6, Table 8 and Appendix Tables 22–23): Examining all 50 trajectories shows that Flash + VHop-Router improves set retrieval over Flash + Qwen3 (34.0% Mixed-Hop / 36.0% 3-Hop vs. 26.0%) and final answer accuracy (20.0% vs. 8.0% Qwen3 / 12.0% GE2, 10/50 vs. 4–6/50) in fewer tool calls (4.2 vs. 14.0); premature stopping is rare (8%, 4/50 on Mixed-Hop; 4%, 2/50 on 3-Hop), while all 7/50 (14%; 8/50 on 3-Hop) cases separating set retrieval from final accuracy had the target returned in the chain before Flash's final multi-object room selection.

Reference: [1] DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories, 2026.

AuthorsOpenReview ↗

Thanks for your review (2/n)

Weakness 2

"The baseline comparison is not fair. VHOP-router is extensively trained for this multi-hop task, where the other retrieval baselines are largely standard Qwen3/GE2. Can you compare it with same training data and supervision budget ? I am not sure whether the gain comes because of the routing mechanism or because of task-specific training."

Response:

TL;DR: Our new attribution experiments isolate routing vs. task-specific training and show three clear conclusions: 1) Under the exact same training data and backbone, the routing mechanism is essential, as vanilla embedding training without routing (Single-Step Training: 8.0%) fails on multi-hop chains; 2) VHop-Router (76.3%) substantially outperforms trained Visual MDR [2] (15.0% greedy / 48.0% Beam = 4), even when MDR is given extra Beam = 4 inference compute and a test-time depth oracle (knowing the ground-truth hop count H in advance to stop, which MDR requires because it has no stopping head, whereas VHop-Router stops autonomously without knowing H); and 3) Holding our trained encoder weights 100% fixed inside the agent, enabling internal routing jumps accuracy from 38.3% → 59.7% (85.4% Mixed-Hop).

Experiment Setup: Across all 300 3-hop L4 test queries (n = 300/hop; Section 5.1, Appendix E.5, Table 17), we compared VHop-Router (SFT → Online IL → RLVR, Beam = 1, no depth oracle)—where Stage 1 uses our 5,880 certified 1–3-hop L4 training questions (1-, 2-, and 3-hop suffix union in approx. 1:1:1 ratio across the seed and training worlds in Table 2 / Appendix D.1), followed by Online IL state supervision and RLVR—against three trained comparators: (1) Single-Step Training, training the vanilla Qwen3-VL-2B embedding model and LoRA setup on the exact same 5,880 training questions to map (query image + instruction) → final target image in one step with no routing mechanism; (2) Visual MDR [2], prefix-conditioned training on the exact same 5,880 questions decoded via Beam = 1 greedy and Beam = 4 beam search (both granted a test-time depth oracle H to terminate search); and (3) 3-Hop Trained Per-Step Tool Ablation (extending Appendix Tables 12–13), exposing our fully trained 3-hop VHop-Router checkpoint to Flash with internal chaining, BACKTRACK, and STOP disabled vs. enabled.

Results:

  • Standalone Retriever Attribution (0 VLM Calls): Single-Step Training with the same training data reaches 8.0% on 3-hop L4, and Visual MDR (even given a test-time depth oracle) reaches 15.0% (Beam = 1) and 48.0% (Beam = 4; 32.3% / 49.3% on 2-hop; Section 5.1 and Appendix E.5, Table 17), whereas VHop-Router (Beam = 1, no depth oracle) reaches 76.3% on 3-hop (+61.3 pp over greedy MDR; +28.3 pp over Beam = 4 MDR; Table 4a and Table 17) and 90.0% on 2-hop (+40.7 pp over Beam = 4 MDR; Section 5.4 and Appendix Table 13a).
  • Same-Checkpoint Agent Attribution (Weights 100% Fixed, Routing Off vs. On): Using our trained 3-hop encoder as a single-step tool yields only +5.6 pp over untrained Qwen3 (32.7% → 38.3%), whereas enabling multi-step routing on that exact same 3-hop checkpoint jumps accuracy to 59.7% (85.4% with Mixed-Hop VHop-Router; and 78.7% vs. 0.3% when restricted to strictly 1 tool call, Table 17):
Method on 3-Hop L4 Test Set (n = 300, Table 17)Learned BACKTRACK + STOP?Depth Oracle Required?Accuracy
Single-Step Training (same 5,880 training questions, standalone)NoNo8.0%
Visual MDR (Beam = 1, standalone)NoYes (H = 3)15.0%
Visual MDR (Beam = 4, standalone)NoYes (H = 3)48.0%
VHop-Router (Beam = 1, standalone)YesNo76.3%
Flash + Trained Per-Step Tool (3-Hop ckpt, routing disabled)NoNo38.3%
Flash + VHop-Router (Same 3-Hop ckpt, routing enabled)YesFloor59.7%
Flash + VHop-Router (Mixed-Hop ckpt, routing enabled)YesNo85.4%

Reference: [2] Xiong et al., Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval, ICLR 2021.

AuthorsOpenReview ↗

Thanks for your review (3/n)

Weakness 3

"The authors mention in limitations that the method is not scalable and doesn't perform well on hop counts outside those observed during training. I appreciate authors acknowledging this but then the work has really low impact because of this limitation. Real-world scenarios would have variable hops. How do you think you will address this issue ?"

Response:

TL;DR: Three points directly address variable-hop deployment and deeper scaling: 1) Dynamic variable-hop search is already fully supported within the trained horizon (1–3 hops, Section 5.4): a single VHop-Router checkpoint dynamically handles unknown 1-, 2-, and 3-hop queries without a depth oracle (99.7% / 90.0% / 76.3%), including vague prompts and variable-hop real YFCC-VHop photo chains (W1). 2) On out-of-horizon 4-hop queries (beyond 3 hops), while length extrapolation is a broad open challenge across sequence models [3] and multi-hop retrievers [4], our diagnosis pinpoints the main bottleneck as the lightweight stopping MLP head—specifically, distribution mismatch on its 2048-dim sequence-state input at unseen 5-image chain lengths, whereas local cosine-similarity drop features remain sharp at hop 4. 3) VHop-Router outperforms Visual MDR on 4-hop under both non-oracle score-drop stopping (24.7% vs. 16.5%) and oracle-depth K = 4 stopping (36.1% vs. 25.8%), pointing to two clear paths for deeper horizons: decoupling the stopping head to use length-invariant local similarity-drop features (already validated zero-shot below) or including longer chains in the mixed-hop curriculum.

Experiment Setup:

  • Dynamic Variable-Hop Control (1–3 Hops, Section 5.4 and Appendix E.1): Evaluated one single VHop-Router (Mixed-Hop) checkpoint across 1-, 2-, and 3-hop queries on the fixed 450-image L4 corpus (n = 300/hop, identical distractors, no depth oracle), under both precise and vague prompts, as well as on variable 1–3 hop real YFCC-VHop albums (W1).
  • Out-of-Horizon 4-Hop Diagnostic vs. Visual MDR (Appendix F.2, Tables 20–21, n = 97, 600 images): Probed the stopping head's 2048-dim sequence state vs. 4-dim local similarity features on 4-hop chains, and evaluated 3-Hop Visual MDR vs. VHop-Router zero-shot under both non-oracle 1D cosine-score-drop stopping (calibrated strictly on 1–3 hop dev queries, zero knowledge of K = 4) and oracle-depth K = 4 stopping across Beam = 1 and Beam = 4.

Results:

  • Dynamic Hop Feasibility (1–3 Hops): Holding the 450-image corpus constant, Flash + Qwen3 drops sharply across 1 / 2 / 3 hops (75.7% → 66.7% → 32.7%, Appendix Tables 12a, 13a, and Section 5.1 Table 4a), whereas one precise Mixed-Hop VHop-Router checkpoint sustains 99.7% / 90.0% / 76.3% standalone without a depth oracle (Section 5.4, Tables 12a, 13a, 4a; our vague-trained Mixed-Hop checkpoint sustains 97.0% / 84.7% / 50.7% vs. 21.3% / 6.7% / 2.0% with Flash on vague prompts, Tables 12d, 13d, 4b; and 53.6% vs. 11.6% on real YFCC-VHop multi-hop queries).
  • 4-Hop Extrapolation Diagnostic (n = 97, Appendix F.2, Table 21): At hop 4, local retrieval similarity still drops sharply (0.64–0.67 on hops 1–3 → 0.41 at hop 4), while only the out-of-length 5-image sequence state suppresses the stopping head (stop probability 0.25 < 0.5). Decoupling stopping via the 1–3-hop-calibrated local score drop (no K = 4 oracle, Beam = 4) lifts VHop-Router from 1.0% (Table 20) to 24.7% vs. 16.5% for Visual MDR, and under oracle-depth K = 4 stopping VHop-Router leads at both Beam = 1 (10.3% vs. 5.2%) and Beam = 4 (36.1% vs. 25.8%; Mixed-Hop encoder: 32.0% vs. 29.9%):
3-Hop Trained Method on 4-Hop L4 (n = 97, Table 21)Beam = 1 (K = 4 Oracle)Beam = 4 (K = 4 Oracle)Beam = 4 (No Oracle, Score-Drop Stop)
Visual MDR5.2%25.8%16.5%
VHop-Router10.3%36.1%24.7%

References: [3] Anil et al., Exploring Length Generalization in Large Language Models, NeurIPS 2022. [4] Trivedi et al., MuSiQue: Multihop Questions via Single-Hop Question Composition, TACL 2022.


Questions For Authors

"See Weaknesses."

Response: Thank you again for your constructive feedback. All concerns above are addressed point by point under W1–W3 and our presentation clarity plan, and we welcome any further questions during the discussion period.

Reviewer_V4DtOpenReview ↗

Thanks for the rebuttal

I thank the authors for their detailed rebuttal and additional experiments, which have addressed some of my concerns. However, I will maintain my original score, mainly due to W3 that I pointed out. The significant performance drop on unseen 4-hop queries (1.0% with autonomous routing and 24.7% with modified stopping) suggests that scalability remains a major limitation. I also noticed differences in the reported agent scores between Table 4 (32.7%/85.4%) and Table 17 (36.3%/88.3%), which needs to be addressed.

Overall, I appreciate the authors efforts, but these concerns prevent me from raising my score.

Reviewer_P17hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper argues that visual agentic search is bottlenecked by single-step retrievers: the agent must describe each retrieval step in text, and the images needed for the next hop are often missing from the top results. The authors propose moving multi-hop navigation into the retriever. They introduce VHop, a synthetic benchmark in which images are linked through shared objects, with controllable difficulty levels. They also propose VHop-Router, which turns a standard multimodal embedding model into an autoregressive multi-step retriever with SELECT, BACKTRACK and STOP actions. The Router is trained with SFT, online IL and RLVR, and returns a whole image chain in a single tool call. Experiments on VHop show clear improvements over standard embedding retrievers, both standalone and as the retrieval tool of an LLM agent.

Critical strengths
  1. The paper is well motivated and clearly written. Offloading multi-hop navigation to the retrieval tool while leaving the LLM agent untouched is a simple and practical idea.
  2. The VHop generator is carefully designed. Hop count, distractors, dead ends and corpus size are all controllable. The ground-truth labels are symbolically verified, which makes VHop more reliable.
  3. The experimental results validate the effectiveness of the proposed method.
Critical weaknesses
  1. The novelty of the task and benchmark is somehow overstated. Multi-step, context-dependent image retrieval over personal visual histories has already been studied, e.g., DeepImageSearch (DISBench) [1], which formulates image retrieval as agentic multi-step exploration over realistic photo histories. That work is neither discussed nor included in Table 1. For the method, conditioning each retrieval step on previously retrieved evidence works like the visual counterpart of multi-hop dense retrieval [2].
  2. All main experiments use a single synthetic benchmark, generated by the same pipeline (linking rules, instruction templates and rendering) as the training data. The Router is trained on exactly this structure, so large gains over untrained baselines are somewhat expected.
  3. The method essentially compiles the agent's reason-then-retrieve loop into the embedding model's autoregressive retrieval. The evidence suggests it learns VHop's specific link patterns rather than general multi-step search. It fails on four-hop queries and on unseen linking rules. This raises the question of whether the approach helps beyond simple, templated visual re-identification chains.
  4. The headline comparisons rely on weak baseline configurations. The agentic baselines return only the top-1 image per call. With k = 10, Flash + GE2 already reaches 78.0% on L4 and outperforms the Router agent on L1–L2 (as shown in Table 14).
  5. There is no trained baseline, so every comparison is trained vs. untrained. The results also come from single runs with large variance across training configurations (e.g., 48.0% vs. 76.3% on L4), some vague-prompt levels have only about 50 queries, and the 61% token reduction claimed in the abstract comes from a pilot on only 10 queries.
  6. All images are generated, including the 386 re-rendered images in the "realistic" test set. On that set the advantage shrinks sharply. Whether the approach holds on real photographs with clutter, occlusion and much larger collections remains unclear; evaluation on real-image benchmarks such as DISBench is needed.

Ref:
[1] DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories, 2026.
[2] Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval, 2021.

Questions for authors

Please check my concerns in critical weaknesses.

7 comments
AuthorsOpenReview ↗

Thanks for your review (1/n)

We sincerely thank Reviewer P17h for the detailed review and for recognizing our clear motivation, practical tool design, and symbolically verified VHop benchmark. To directly address your comments, we added head-to-head trained baselines, full-set profiling, and real-world Flickr photo evaluations that strengthen our empirical conclusions. Below, we address each point in detail.

Weakness 1.1

"The novelty of the task and benchmark is somehow overstated. Multi-step, context-dependent image retrieval over personal visual histories has already been studied, e.g., DeepImageSearch (DISBench) [1], which formulates image retrieval as agentic multi-step exploration over realistic photo histories. That work is neither discussed nor included in Table 1.

  • [1] DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories, 2026."

Response:

TL;DR: Thank you for pointing out DISBench [1]—we have discussed it prominently in Section 2 and included it in Table 1 (with full real-photo experiments and visual examples in Section 5.6, Figure 7, Table 8, and Appendix F.5, Table 25 and Figure 14). VHop and DISBench address complementary problems. Official DISBench is an evaluation-only agent benchmark where 46 of 122 queries require no cross-photo visual links (relying on EXIF timestamp/GPS metadata or single-image attributes), and even among the remaining 76 visually linked queries, the visual link itself is strictly single-hop (with no multi-hop visual links). By contrast, VHop isolates multi-hop visual chaining and provides trajectory supervision to train an in-retriever visual router that works alongside metadata tools—verified on Official DISBench and two new real-photo multi-visual-hop benchmarks (YFCC-VHop and DISBench-MVH).

Experiment Setup:

  • Modality Audit of Official DISBench: Prior to running any model, we annotated all 122 Official DISBench queries (57 Flickr users, 109k total photos, about 1,900 photos/user) by required reasoning modality (cross-photo visual re-identification vs. EXIF timestamp/GPS metadata).
  • Constructing YFCC-VHop and DISBench-MVH: Along with evaluating Official DISBench, to directly test multi-hop visual link capabilities on real photography, we followed DISBench's mining workflow (recurring entity grouping → co-occurring object graph walk → shortcut rejection against earlier chain photos → bridge-dependent query composition, scored via TREC-style pooled judging) to construct two real-photo benchmarks where every hop requires visual links: (1) YFCC-VHop over 78 YFCC100M Flickr users strictly disjoint from DISBench (about 750 real photos/album; 2 of 80 candidate users had zero multi-hop chains after filtering), split by user into 66 training users (590 1–3-hop queries, used to train VHop-Router following our exact same 3-stage pipeline: SFT → Online IL → RLVR) and 12 held-out test users (290 1–3-hop test queries, including n = 69 2–3-hop multi-hop queries); and (2) DISBench-MVH (Multi-Visual-Hop extension; 109 zero-shot 1–3-hop queries on Official DISBench user albums, including n = 38 multi-hop queries of which n = 25 are 2-hop; zero training on DISBench users), keeping each original query verbatim as Hop 1 and appending verified 2nd/3rd visual hops. We evaluated this YFCC-VHop-trained router on held-out YFCC-VHop test albums and zero-shot inside a shared DISBench-style metadata+search agent (Gemini 3.5 Flash) on Official DISBench and DISBench-MVH.

Results:

  • DISBench Modality Audit: Out of 122 Official DISBench queries, 46/122 (37.7%) involve zero cross-photo visual links (40 rely purely on EXIF time/GPS metadata links across photos and 6 are single-image scene checks); among the remaining 76/122 (62.3%) that use a visual link (61 combined with EXIF time/GPS filters), the visual correspondence itself is strictly 1 hop (matching a single recurring entity across photos, with 0 multi-hop visual chains).
  • Real-Photo Multi-Hop Retrieval: On held-out YFCC-VHop multi-hop test queries (n = 69), VHop-Router lifts standalone Hit@1 over baseline Qwen3-VL-2B from 11.6% → 53.6% (8/69 → 37/69, +42.0 pp) and Gemini 3.5 Flash agent Top-1 accuracy from 44.9% → 55.1% (31/69 → 38/69, +10.2 pp). Transferred zero-shot to DISBench user albums alongside metadata tools, it lifts agent Any-Hit recall from 75.4% → 80.7% (43/57 → 46/57) on Official DISBench Intra-Event queries (n = 57) and from 76.3% → 81.6% (29/38 → 31/38) on DISBench-MVH multi-hop queries (2-hop standalone Hit@1: 24.0% → 36.0%, 6/25 → 9/25).
AuthorsOpenReview ↗

Thanks for your review (2/n)

Weakness 1.2

"For the method, conditioning each retrieval step on previously retrieved evidence works like the visual counterpart of multi-hop dense retrieval [2].

  • [2] Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval, 2021."

Response:

TL;DR: Building on our citation of Multi-Hop Dense Retrieval (MDR) [2] in Section 2 and Appendix A, we expanded Section 2 to clarify the methodological distinction and added a direct Visual MDR baseline (Section 5.1; Appendix E.5, Table 17). VHop-Router uses a visual MDR stage as its Stage 1 SFT foundation, and surpasses MDR even when MDR is granted extra beam-search compute and prior hop-count knowledge. Specifically, MDR trains a prefix-conditioned encoder via teacher forcing on gold chains and decodes via fixed-depth forward beam search—which has no stopping head and therefore requires a test-time depth oracle (knowing the ground-truth hop count H in advance to terminate search) and cannot recover from lookalike dead ends. Post-training that same MDR base with Online IL + RLVR equips VHop-Router with explicit BACKTRACK recovery and autonomous variable-depth STOP (requiring no depth oracle), driving +28.3 to +61.3 pp gains.

Experiment Setup: Using the exact same Qwen3-VL-2B backbone, LoRA rank, and 5,880 certified 1–3-hop L4 training questions (1-, 2-, and 3-hop suffix union in approx. 1:1:1 ratio across the seed and training worlds in Table 2 / Appendix D.1), we evaluated our Stage 1 prefix-conditioned SFT retriever as Visual MDR—decoded with both Beam = 1 greedy search and original MDR's Beam = 4 beam search (both granted a test-time depth oracle, H known)—head-to-head against VHop-Router (Online IL + RLVR post-trained on that same MDR base, decoded greedily at Beam = 1 without a depth oracle) on the 300-query L4 test set (n = 300/hop).

Results: Even when granted a test-time depth oracle, trained Visual MDR reaches 15.0% (Beam = 1) and 48.0% (Beam = 4) on 3-hop L4 (32.3% and 49.3% on 2-hop; Section 5.1 and Appendix E.5, Table 17), whereas post-training VHop-Router on that same MDR base boosts accuracy to 76.3% on 3-hop (+61.3 pp over greedy MDR; +28.3 pp over Beam = 4 MDR; Table 4a and Table 17) and 90.0% on 2-hop (+40.7 pp over Beam = 4 MDR; Section 5.4 and Appendix Table 13a) at Beam = 1 without a depth oracle and at comparable compute (9.8 vs. 9.0 state encodes/query).

AuthorsOpenReview ↗

Thanks for your review (3/n)

Weakness 2

"All main experiments use a single synthetic benchmark, generated by the same pipeline (linking rules, instruction templates and rendering) as the training data. The Router is trained on exactly this structure, so large gains over untrained baselines are somewhat expected."

Response:

TL;DR: Our new and existing experiments show three clear takeaways: 1) Under the exact same training data, the stateful routing mechanism (SELECT / BACKTRACK / STOP) is essential, as trained baselines without routing (Single-Step Training and Visual MDR, W1.2) fall far short of VHop-Router; 2) VHop-Router does not merely memorize VHop's generator, transferring zero-shot across unseen appearance twins, 3-object cluttered scenes, non-templated prompts, and a new 3D ray-traced rendering engine; and 3) Our training recipe transfers directly to unmodified real-world Flickr photo albums (YFCC-VHop and zero-shot DISBench-MVH, W1.1).

Experiment Setup:

  • Same-Data Trained Baselines and Out-of-Distribution Shifts: Sharing the identical Qwen3-VL-2B backbone, LoRA setup, and 5,880 1–3-hop L4 training questions (W1.2), we compared VHop-Router against Single-Step Training (training the vanilla Qwen3-VL-2B embedder on our exact same training data to map query image + instruction directly to the final target image in one step, with no routing mechanism) and Visual MDR (W1.2) on 3-hop L4 (n = 300, Appendix E.5, Table 17). We also evaluated zero-shot transfer from two-object precise L4 training to L5 appearance twins (Section 5.1, Table 4a, w0), three-object cluttered scenes (Section 5.5, Table 7, n = 299), natural-language rephrased and vague L4 prompts (Appendix E.4, Table 16, n = 100 dev / n = 150 test), and a new procedural 3D Blender Cycles rendering shift (n = 100, matched K = 3, Beam = 4 encoder evaluation, Appendix F.2), alongside our vague-prompt Mixed-Hop configuration (Section 5.1, Table 4b, n = 150).
  • Real-World Flickr Photo Albums (W1.1): Evaluated our SFT → Online IL → RLVR pipeline on unmodified personal Flickr albums in YFCC-VHop (n = 69 multi-hop test queries) and zero-shot on DISBench-MVH (W1.1, Appendix F.5, Table 25).

Results:

  • Same-Data Baselines and Distribution Shifts: Trained on the identical L4 data, Single-Step Training (8.0%) and Visual MDR (15.0% greedy / 48.0% Beam = 4 with depth oracle) trail far behind VHop-Router (76.3%, no depth oracle) on 3-hop L4 (Table 17). Trained solely on two-object precise L4, VHop-Router transfers zero-shot to L5 appearance twins (84.0% vs. 28.0% with Flash in Section 5.1, Table 4a; 72.0% vs. 45.0% standalone vs. Visual MDR on w0), three-object cluttered scenes (52.2% vs. 27.4%, Section 5.5, Table 7), natural-language rephrased L4 prompts (19.0%–31.0% standalone and 20.0% zero-shot on vague L4 in Appendix E.4, Table 16; 42.0% vs. 2.0% standalone under vague Mixed-Hop training in Table 4b), and 3D Blender Cycles renders (36.0%–42.0% vs. 13.0% Visual MDR, n = 100, Appendix F.2).
  • Real Flickr Photo Histories (W1.1): On held-out YFCC-VHop user albums (n = 69), VHop-Router boosts multi-hop standalone Hit@1 from 11.6% → 53.6% and Flash agent Top-1 from 44.9% → 55.1%, and lifts zero-shot 2-hop Hit@1 on DISBench-MVH (n = 25) from 24.0% → 36.0%.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.