ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
AgentHorizon
AgentHorizon: Evaluating Agentic Judges for Long-Horizon Computer-Use Tasks
A benchmark of 1,373 computer-use tasks with paired instructions to evaluate agentic judges on long-horizon trajectory success verification across three OSs.
How to run
ltb · agenthorizon
$ npm i -g lettokenburn-cli$ ltb bench info agenthorizon
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
10 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.5 | 80.9 | codex | Independent | medium | Oct 8, 2026 | arxiv.org |
| Gemini 3.1 Pro | 77.7 | gemini cli | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Opus 4.7 | 76.0 | claude code | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.6 27B | 70.4 | opencode | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.4-mini | 69.2 | codex | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.6 35B A3B | 64.6 | opencode | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Haiku 4.5 | 52.2 | claude code | Independent | medium | Oct 8, 2026 | arxiv.org |
| Gemini 3.1 Flash Lite | 52.1 | gemini cli | Independent | medium | Oct 8, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 49.4 | opencode | Independent | medium | Oct 8, 2026 | arxiv.org |
| google/gemma-4-31B-it | 46.8 | opencode | Independent | medium | Oct 8, 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
AgentHorizon comprises 1,373 instruction-trajectory pairs from 166 hours of human recordings. It uses a paired design where swapping instructions creates negative examples to test if judges distinguish successful trajectories from incompatible ones.
Scoring uses balanced accuracy to measure a judge's ability to accept valid trajectories and reject failed ones. The best agentic judge, GPT-5.5, achieved 80.9% balanced accuracy, highlighting discrepancies between frontier and open-weight models in tool-use scenarios.