LetTokenBurnAgent Monitoring Benchmark
A software-engineering benchmark evaluating monitors' ability to identify consequential agent decisions and locate evidence for user oversight across three subsets.
$ npm i -g lettokenburn-cli$ ltb bench info agentmonbench
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 | 1.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| GLM-5.3 | 1.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Kimi K3 | 1.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5.6-terra | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 0.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Claude Sonnet 5 | 0.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5.6 Luna | 0.7 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| DeepSeek V4 Pro | 0.5 | 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.
AgentMonBench comprises three subsets (SpecGAP, SilentSwap, FeedbackTrace) derived from DeNovoSWE and SWE-chat to test alignment with requirements and awareness of autonomous decisions. It evaluates decision identification and evidence localization capabilities of monitoring systems in long-horizon coding tasks.
Scoring focuses on the accuracy of identifying consequential decisions and localizing supporting evidence. Experiments show that the proposed Evidence-Grounded Behavior Graph (EBG) method improves performance over baselines across eight models.