ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 120
MLOps-Bench
MLOps-Bench: Benchmarking AI Agents on Production ML Engineering in Real Repositories
A repository-grounded benchmark with 212 tasks from 31 public ML repos evaluating AI agents on production MLOps lifecycle changes using executable behavioral tests.
How to run
ltb · mlops-bench
$ npm i -g lettokenburn-cli$ ltb bench info mlops-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
Scores
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.5 | 68.7 | codex | Independent | medium | Oct 16, 2026 | arxiv.org |
| Claude Opus 4.6 | 52.9 | claude code | Independent | medium | Oct 16, 2026 | arxiv.org |
| Claude Sonnet 4.6 | 52.3 | claude code | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-5.5 | 47.9 | base agent | Independent | medium | Oct 16, 2026 | arxiv.org |
| Claude Sonnet 4.6 | 39.2 | base agent | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-5.2-chat | 31.4 | base agent | Independent | medium | Oct 16, 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
MLOps-Bench derives tasks from real ML repositories, covering data pipelines, feature engineering, training, serving, monitoring, CI/CD, and integration. Each task includes a natural-language requirement, oracle patch, buggy baseline, and provenance metadata.
Scoring uses fail-to-pass and pass-to-pass tests to measure behavioral correctness. The best agent solved 31.2% of tasks and passed 68.7% of tests, indicating significant gaps in production-ready MLOps capabilities.