ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
2D-FET-Bench V2
2D-FET-Bench: From Spatial Reasoning to FET Design on Flakes
A benchmark of 128 layout tasks for evaluating language-model agents in generating geometrically valid FET designs on microscopy-derived 2D flake contours.
How to run
ltb · 2d-fet-bench-v2
$ npm i -g lettokenburn-cli$ ltb bench info 2d-fet-bench-v2
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
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Luna | 62.3 | react-3/c0 | Independent | medium | Oct 5, 2026 | arxiv.org |
| GLM-5 | 28.8 | react-3/c0 | Independent | medium | Oct 5, 2026 | arxiv.org |
| Kimi K2.5 | 23.0 | react-3/c0 | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3 235B | 3.8 | react-3/c0 | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3-32B | 1.3 | react-3/c0 | Independent | medium | Oct 5, 2026 | arxiv.org |
| Ministral-3-14B | 0.8 | react-3/c0 | Independent | medium | Oct 5, 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 benchmark comprises 128 tasks with textual specifications and contour coordinates, requiring agents to generate executable polygon/path operations for GDSII layouts. It includes hole-containing flakes and multi-flake scenarios.
Scoring uses a deterministic verifier for geometric/structural rules and input integrity. The best model (GPT5.6-Luna) achieved 80.5% task coverage and 43.8% consistency across five attempts.