69.50
Math avg pass@8
53.33
HMMT25
3×10K
training subsets
9/10
个人兴趣
01 · QUESTIONarXiv:2605.08741 · 2026-05-09
能不能把 inference harness 变成一次性训练 scaffold？
训练时运行 Plan–Solve / Draft–Verify，部署时只保留 direct model。
来源：Abstract；Tables 1-31 / 9
02 · CONTRACTarXiv:2605.08741 · 2026-05-09
Harness terminal 只给 teacher，student 仍直接作答
Direct student
原 problem → on-policy response
← KL
Harness teacher
执行 procedure → terminal context → score same response
来源：Method overview2 / 9
03 · PLAN–SOLVEarXiv:2605.08741 · 2026-05-09
Reference solution 先被压成 plan，再进入 teacher terminal
1Reference
planner 训练期可见
2Plan
抽象 strategy sketch
3Solve
不再看 reference
4Teacher
读 terminal
5Student
吸收 procedure
来源：Section 3.23 / 9
04 · DRAFT–VERIFYarXiv:2605.08741 · 2026-05-09
五个支持者和五个挑战者共同复核 draft
1Retrieve
5 neighbors
2Draft
初步 label
3Confirm
5 supporting views
4Challenge
5 opposing views
5Revise
整合后 terminal
来源：Section 3.14 / 9
05 · SETTINGarXiv:2605.08741 · 2026-05-09
Qwen3-8B，三组 10K 数据
| 任务 | Steps | Harness |
|---|---|---|
| CAIL / LawBench | 300 | Draft–Verify |
| USPTO | 300 | Draft–Verify |
| DeepMath | 150 | Plan–Solve |
来源：Section 4；Appendix5 / 9
06 · RESULTarXiv:2605.08741 · 2026-05-09
OPHSD 在三类任务上超过 OPSD
LawBench OPSD
64.25
LawBench OPHSD
69.51
Math OPSD
66.68
Math OPHSD
69.50
根据 main tables 重绘6 / 9
07 · CASEarXiv:2605.08741 · 2026-05-09
检索 harness 也会把证据带偏
Harness terminal
近邻全偏向 intentional injury
Trained direct student
保留 negligent homicide 的另一种解释
来源：LawBench case study7 / 9
08 · GAME CODEarXiv:2605.08741 · 2026-05-09
内化 procedure，不假装内化未来 execution result
可内化
plan→compile→run→inspect→repair 的步骤
必须执行
当前 build 的 error、render 与 gameplay state
本文对 game query-to-code 的边界8 / 9
09 · CODEarXiv:2605.08741 · 2026-05-09
Release 质量较好，reporting protocol 要分开
公开
trainer、harness、data、4-run CSV
不可混用
v1 best checkpoint vs repo fixed checkpoint
代码审计：20eaa0c8b0674fbf81aa780a8cf4e0d7c550d4ea9 / 9