VHop
Learning to Route in Visual Space via Multi-Step Embedding Retrieval: VHOP Benchmark
A flexible data generation framework and benchmark for visual multi-hop image retrieval testing matching, search planning, and symbolic verification across five difficulty levels.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info vhop
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3 Max (2025-09-23) | 49.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GE2 | 32.7 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Content & Construction: The document introduces VHop, a controllable setting for studying multi-hop visual retrieval in latent space. It connects images via shared objects to create trajectories with varying hop counts, object distractors, and task difficulties (Table 1). A flexible data generation framework provides verified ground-truth labels satisfying detailed query instructions.
Scoring & Findings: The benchmark evaluates retrievers on success rates across different difficulty levels defined by hops, corpus size, etc. Experiments show VHop-Router boosts retrieval performance from under 5% to 76.3%, improving agentic search task success rates by 52.7% and reducing token length significantly compared to baselines.