CROSS-Bench
Harbor: pendingA benchmark of 28,000 decision problems evaluating candidate-bound visual contribution through matched invalidation and rebinding tests on vision-language models.
- Samples
- 28,000
- Released
- 2026
- Scores
- 4
EgoTools-Bench
EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos
A diagnostic benchmark of 1,000 QA pairs across four tracks evaluating tool-use understanding from perception and geometry to procedure and causal reasoning.
How to run
ltb · egotools-bench
$ npm i -g lettokenburn-cli$ ltb bench info egotools-bench
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
9 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 69.7 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 64.5 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Gemini 3.1 Flash Lite | 58.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GLM4.1V-Thinking | 46.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 46.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 45.9 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| MiMo-VL-7B-SFT | 41.4 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 40.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| LLaVA-OneVision-7B | 37.6 | 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 runsLoading community runs…
About
Content & Construction: EgoTools-Bench is a comprehensive suite consisting of 1,000 question-answer pairs designed to isolate specific reasoning demands in egocentric video. It comprises four distinct evaluation tracks: Affordance & Causality (goal-directed tool choice), Perception & Grounding (observable facts like identity and state), Procedural Dynamics (temporal organization of actions), and Spatial Reasoning (geometric relations). The benchmark is constructed from a larger corpus, EgoTools-Data, which includes 100 hours of synchronized audio recordings with dense captions.
Scoring & Key Findings: Models are evaluated on their ability to ground tool use in visual evidence. Current models struggle significantly; for instance, Gemini-3.1-Pro achieves only 51.7% accuracy on Perception & Grounding despite a higher overall score of 66.9%. The benchmark serves as both an evaluation metric and a training resource, demonstrating that supervised fine-tuning can improve model performance from 50.0% to 60.9%.