AlgoPuzzleVQA
Harbor: pendingWe introduce the novel task of multimodal puzzle solving, framed within the context of visual question-answering.
- Samples
- 1,800
- Released
- 2024
- Scores
- 1
Unified Suite of 25 Visual Reasoning Benchmarks
Unified suite of 25 visual reasoning benchmarks
A collection of 25 existing visual reasoning benchmarks aggregated to evaluate VLM attention alignment and downstream performance.
How to run
ltb · unified-suite-of-25-visual-reasoning-benchmarks
$ npm i -g lettokenburn-cli$ ltb bench info unified-suite-of-25-visual-reasoning-benchmarks
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
1 sourced entry| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 61.1 | N/A | Independent | medium | Oct 4, 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 runsLoading community runs…
About
The authors aggregate 25 existing visual reasoning benchmarks to create a comprehensive evaluation suite, addressing the lack of consensus in prior work. This suite is used to reproduce results and evaluate the proposed Self-Saliency method.
Scoring involves mean score and average rank across the benchmarks. The study finds that Self-Saliency significantly outperforms prior attention-steering methods and baselines, achieving better average rank and mean score.