ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 118
Open-Endedness Bench
Open-Endedness Bench: Measuring Epistemic Process from Agent Records
Evaluates epistemic process and research style of LLM agents by analyzing execution logs into an event graph, independent of task outcomes or reference answers.
How to run
ltb · open-endedness-bench
$ npm i -g lettokenburn-cli$ ltb bench info open-endedness-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 | 1.0 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Fable 5 | 0.9 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| GPT-5.6 | 0.9 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Claude Opus 5 | 0.8 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Kimi K3 | 0.8 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| GLM-5.2 | 0.8 | N/A | Independent | medium | Oct 1, 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
OEB compiles agent execution records into a unified epistemic event graph linking propositions to actions. It scores competence axes (evidence, experiment, revision, no reward hacking) and persona traits without ground-truth answers.
Scoring uses fixed rules on graph counts, validated on 119 runs across 12 tasks. Findings show rescoring changes scores slightly, and only 16–29% of claimed improvements are real, with model explaining more variance than task.