HumanEval
Harbor adapterHumanEval: 164个函数生成
- Samples
- 164
- Released
- 2021
- Scores
- 4
NP-Bench
An environment-grounded benchmark verifying integration outcomes of parallel coding agents via real git merges across three coordination arms.
How to run
ltb · np-bench
$ npm i -g lettokenburn-cli$ ltb bench info np-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
3 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 4.8 | 100.0 | np-bench | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5 Codex | 100.0 | np-bench | Independent | medium | Oct 5, 2026 | arxiv.org |
| Claude Haiku 4.5 | 60.0 | np-bench | Independent | medium | Oct 5, 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
NP-Bench evaluates multi-agent coding coordination using a three-arm design (no coordination, reactive detection, proactive planning) grounded in real git merge environments. It measures clean-integration success and scope compliance rather than single-agent task resolution.
Scoring relies on deterministic simulations and live agent tests to verify if merged code builds and passes. The planner arm achieved 9/9 clean integrations deterministically, significantly outperforming baselines which suffered from merge conflicts.