LetTokenBurnGraphDecide: Benchmarking System One Models on Graph Tasks
A model-independent benchmark evaluating System One models on graph cognition, graph-text decision-making, and sequential graph optimization tasks using controlled inputs and candidate pools.
$ npm i -g lettokenburn-cli$ ltb bench info graphdecide
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 | |||
|---|---|---|---|---|---|---|
| Qwen3.8 27B | 100.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen2.5-72B-Instruct | 100.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| qwen3.5-27B | 99.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| qwen3.5-9B | 99.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.5 4B | 99.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.5 2B | 99.3 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Laya | 84.2 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Decider-2B v11 | 79.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Kev 4B | 77.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.5-0.8B | 70.6 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Jev | 59.7 | N/A | 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.
GraphDecide comprises three task families: structural queries on small graphs, node classification/retrieval with matched text-graph inputs, and sequential optimization (TSP, MaxCut, etc.) on public datasets. It uses choice-based interfaces to diagnose decision quality independent of reasoning traces.
Scoring relies on algorithmically computed ground truth or reference optima for accuracy and solution quality. Findings indicate that adjacency recognition doesn't ensure structural correctness, and joint inputs don't consistently improve prediction over textual or relational alone.