64.32
4B
65.40
8B
0.5%
train PI invocation
8/10
个人兴趣
01 · QUESTIONarXiv:2608.01589 · 2026-08-03
比 full solution 更好的 PI，会不会是 problem structure？
PS-OPSD 用 state、goal、constraints 与 transitions 替代完整推理。
来源：Abstract；Tables 1-51 / 9
02 · STRUCTUREarXiv:2608.01589 · 2026-08-03
四字段把 reference trace 改写成可行动约束
Initial
已知状态
Goal
成功条件
Constraints
规则边界
Transitions
operator + preconditions + result
来源：Method section2 / 9
03 · PROTOCOLarXiv:2608.01589 · 2026-08-03
Student rollout 不变，只替换 teacher context
1Question
student only
2Rollout
current policy
3Guidance
teacher only
4Rescore
same tokens
5Forward KL
full vocabulary
来源：Algorithm 13 / 9
04 · EXTRACTORarXiv:2608.01589 · 2026-08-03
Guidance 由 35B-A3B model 从 verified solution 提取
保留
problem dynamics 与 selected path
删除
instantiated final answer
来源：data construction4 / 9
05 · ABLATIONarXiv:2608.01589 · 2026-08-03
长度、相关性与顺序分别有贡献
| 4B condition | Avg |
|---|---|
| Wrong-task matched length | 61.76 |
| Corrupted order | 62.19 |
| Flattened fields | 63.61 |
| PS-OPSD | 64.32 |
根据 ablation table 重绘5 / 9
06 · RESULTarXiv:2608.01589 · 2026-08-03
三个 scale 都高于 OPSD
| Scale | Base | OPSD | PS-OPSD |
|---|---|---|---|
| 1.7B | 36.67 | 41.08 | 43.12 |
| 4B | 61.76 | 61.70 | 64.32 |
| 8B | 62.31 | 63.58 | 65.40 |
根据 main table 重绘6 / 9
07 · GAME MAParXiv:2608.01589 · 2026-08-03
把 game project 写成 implementation-neutral contract
Assets/State
initial
Tests/Goals
success
Engine/API
constraints
Plan
transitions
本文 game context 映射7 / 9
08 · MISSINGarXiv:2608.01589 · 2026-08-03
关键 setting 与 code 未公开
Paper 有
method、results、controls
Release 缺
repo、manifest、exact hyperparameters
2026-10-08 artifact check8 / 9
09 · VERDICTarXiv:2608.01589 · 2026-08-03
适合定义 context，不足以定义 state compatibility
来自一条 reference path 的 transition 仍可能不适配 student 当前 build。
兴趣程度 8/109 / 9