开发与测试
libero-inference-time-scaling
Run the LIBERO inference-time scaling experiment: debug Long-Pro tasks at frozen library snapshots, evaluate token-budget checkpoints, and plot Pareto curves.
NVlabs · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add NVlabs/ASPIRE --skill libero-inference-time-scaling技能说明
name: libero-inference-time-scaling description: "Run the LIBERO inference-time scaling experiment: debug Long-Pro tasks at frozen library snapshots, evaluate token-budget checkpoints, and plot Pareto curves."
LIBERO Inference-Time Scaling
Use this skill for Experiment 5. Start with INSTRUCTIONS.md.
Run Order
- Start with INSTRUCTIONS.md.
- Set up snapshot worktrees for the target library sizes.
- Follow main-agent-prompt.md for Stage 1 debug + Stage 2 held-out eval.
- Use subagent-prompt.md for each debug subagent.
- Follow token-scaling-coordinator.md to select token-budget checkpoints and run Stage 2 variants.
- Use clean-task-slate.md before rerunning a snapshot or task.
Files
| File | Purpose |
|---|---|
| INSTRUCTIONS.md | End-to-end human/agent entrypoint |
| main-agent-prompt.md | Coordinator recipe for debug+held-out eval |
| subagent-prompt.md | Per-task Stage 1 debug subagent prompt |
| clean-task-slate.md | Reset checklist before reruns |
| token-scaling-coordinator.md | Token-budget checkpoint eval recipe |
| ../library-size-scaling/main-agent-prompt.md | Related zero-shot snapshot eval flow |