01 · CORE PROBLEMarXiv:2608.05219 · 2026-08-05

Reference path 只在兼容 state 上有效

Student 改变 subgoal 顺序后，full-path teacher 可能把 recovery action 当错。

0.865

ALFWorld

0.693

WebShop success

20.2%

structured coverage

10/10

个人兴趣

来源：Abstract；Tables 1-41 / 9
02 · MISMATCHarXiv:2608.05219 · 2026-08-05

同一个任务，不代表同一个 execution state

Location

agent 在哪

Progress

哪些 subgoal 已完成

Admissibility

reference next action 还合法吗

来源：Method state definition2 / 9
03 · ROUTINGarXiv:2608.05219 · 2026-08-05

Match 才 distill；不 match 就 abstain

1Observe

student real history

2Signature

structured state

3Match

reference prefixes

4Route

latest compatible

5Distill

matched turns only

来源：Method overview3 / 9
04 · TWO SIGNALSarXiv:2608.05219 · 2026-08-05

Dense loss 可以关闭，GRPO 不关闭

All turns

GRPO outcome reward

Matched turns

privileged self-distillation

来源：Objective section4 / 9
05 · REAL CASEarXiv:2608.05219 · 2026-08-05

Lettuce 已入 fridge，下一步不该再开一次

History matcher

文本相似 → 建议重复 open fridge

Structured matcher

progress 已完成 → 路由 second-object phase

来源：ALFWorld case study5 / 9
06 · ABLATIONarXiv:2608.05219 · 2026-08-05

路由比单纯 sparse distillation 更重要

ConditionALFWorld Avg@4
FullPath-SD0.746
Random same-count0.723
Routing only0.836
Routing + local context0.865
Context on unmatched0.695
根据 ablation table 重绘6 / 9
07 · MATCH AUDITarXiv:2608.05219 · 2026-08-05

Structured matching 提高 coverage，不明显牺牲 precision

20.2%

structured coverage

15.4%

history coverage

781/781

sampled replay success

来源：matcher audit7 / 9
08 · GAME STATEarXiv:2608.05219 · 2026-08-05

Code state 需要 AST、scene、fixture 与 runtime 一起描述

Code/AST

files/functions

Scene

nodes/signals

Verifier

fixture/obligation

Runtime

state/events

本文 game code state design8 / 9
09 · CODE REALITYarXiv:2608.05219 · 2026-08-05

Core code 在，paper data 不在

公开

matcher、trainer、env、tests

缺失

4 个 claimed included JSON files

代码审计：f3e43bab78828d12ee7e2f717e041b70f43ad43f9 / 9