ARC-AGI-2
Harbor nativeARC-AGI-2: 更难的抽象推理
Verified100%Abstract Reasoning
- Samples
- 167
- Released
- 2025
- Scores
- 120
LegalRewardBench
LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation
A human-verified benchmark with 1,331 preference pairs evaluating reward models on legal tasks across five capability levels.
How to run
ltb · legalrewardbench
$ npm i -g lettokenburn-cli$ ltb bench info legalrewardbench
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 4.7 | 86.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Opus 4.5 (2025-11-01) | 85.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.5 | 83.7 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 83.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3 235B A22B | 81.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini 2.5 Pro | 81.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Kimi K2.5 | 80.7 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.2-chat | 80.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GLM-4.5 | 79.9 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3-30B-A3B | 79.3 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-4.1 | 79.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3.5 397B A17B | 78.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini 2.5 Flash | 78.5 | N/A | Independent | medium | Oct 13, 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
LegalRewardBench comprises 1,331 verified preference quadruplets (query, chosen, rejected, reference) spanning knowledge, understanding, analysis, decision, and writing. Data is constructed via rubric-guided automatic scoring followed by human verification.
It evaluates generative and discriminative reward models, finding general-purpose LLMs most robust. Scoring relies on alignment with human-verified preferences using rubric-based criteria.