FG-BMK
Harbor: pendingFG-BMK FG-BMK is a fine-grained evaluation benchmark for large vision-language models and vision-language models.
- Samples
- N/A
- Released
- 2026
- Scores
- 0
FG-Sudoku
Fine-Grained Sudoku Benchmark
A diagnostic benchmark for evaluating MLLMs on visual Sudoku reasoning, categorized by grid size and deductive steps to measure long-horizon multimodal constraint satisfaction.
How to run
ltb · fg-sudoku
$ npm i -g lettokenburn-cli$ ltb bench info fg-sudoku
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
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 2.5 Pro | 0.9 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| o3 | 0.8 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Gemini 2.5 Flash | 0.7 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Claude Sonnet 4 (2025-05-14) | 0.6 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 0.4 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GPT-4o | 0.3 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| DeepSeek-V3 | 0.3 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 0.2 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 0.2 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Video-R1 | 0.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| InternVL3-8B | 0.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| LLaVA-NeXT-34B | 0.0 | N/A | Independent | medium | Oct 9, 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
FG-Sudoku presents visual Sudoku puzzles of varying difficulty, quantified by grid size and required deductive steps. It serves as a fine-grained diagnostic tool for assessing multi-step deductive reasoning over local and global constraints.
Scoring is based on binary outcome accuracy (correct/incorrect) without LLM-as-judge. Evaluations show frontier MLLMs fail on complex 9x9 instances due to long-range dependency failures, highlighting reward sparsity in standard RLVR.