ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 120
ANCHOR
Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding
A benchmark evaluating engineering constraint following in repository-grounded agentic coding across 358 multilingual tasks with hierarchical constraints.
How to run
ltb · anchor
$ npm i -g lettokenburn-cli$ ltb bench info anchor
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
13 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 4.8 | 19.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.3-codex | 12.3 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.4 | 12.0 | cline | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.4 | 11.5 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.3-codex | 9.2 | cline | Independent | medium | Oct 12, 2026 | arxiv.org |
| GLM-5.1 | 7.8 | cline | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.4-mini | 7.3 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| GLM-5.1 | 6.1 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| GLM-5 | 4.2 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| Kimi K2.5 | 4.2 | opencode | Independent | medium | Oct 12, 2026 | arxiv.org |
| GPT-5.4-mini | 4.1 | cline | Independent | medium | Oct 12, 2026 | arxiv.org |
| Kimi K2.5 | 3.9 | cline | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen3.5 Flash | 2.8 | cline | Independent | medium | Oct 12, 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
ANCHOR comprises 358 naturalistic, multilingual repository tasks spanning eight language groups. It features a taxonomy of 10 categories and 78 leaf types, with atomic constraints composed into AND/THEN/PRIORITY relations.
Evaluation uses a trajectory-grounded framework with rule-based and LLM judges to check code diffs and execution trajectories. The strongest system achieved 70.1% atomic compliance, while only 19.0% of tasks passed all items under the strict FCCR criterion.