STRIDE
SpatioTemporal Reasoning In Driving Edge Cases
An object-centric benchmark evaluating spatiotemporal reasoning in autonomous driving edge cases via multiple-choice and open-ended VQA tasks.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info stride
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
13 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-6 | 48.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-5.6 | 46.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-5.5 | 44.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q8-I | 44.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q2.5 | 42.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Claude | 40.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| G-Fl. | 40.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q27 | 40.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| G-Pro | 38.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q8-T | 38.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q30-T | 36.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Dolphins | 32.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Q35 | 32.0 | N/A | Independent | medium | Oct 16, 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
STRIDE comprises 39 question templates across six task families (e.g., spatial perception, trajectory prediction) using multi-view temporal driving data with unusual object behaviors. It pairs multiple-choice questions with free-form VQA to assess model justification.
Scoring uses accuracy for MCQs. Evaluation of 12 models shows significant gaps, with top performance at 43.4% accuracy, indicating current models struggle with reconstructing object motion and safety-critical judgments.