LetTokenBurnRoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment
An interactive benchmark of 674 tasks evaluating 3D spatial reasoning by aligning rotated object projections to target silhouettes via discrete rotations.
$ npm i -g lettokenburn-cli$ ltb bench info rota-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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 17.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini 3.6 Flash | 10.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-6 Astra | 10.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.6-terra | 8.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.6 Luna | 4.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Opus 5 | 0.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Sonnet 5 | 0.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3.7 Plus | 0.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
RoTA-Bench constructs 3D objects from 2D icons and requires models to perform sequential rotations to match a target silhouette. It features 674 canonical tasks with a stratified 100-task test set, using reversible initialization to ensure solvability.
Scoring relies on achieving a projection Intersection over Union (IoU) of at least 0.90 within an action budget. Results show a significant human-model gap, with top models succeeding on only 17% of tasks compared to 100% for humans.