§3.4 说 Ng et al. 保证 shaping reward 相对 task reward 较小时 policy invariance 成立。但 Ng et al. 的结论是：shaping 项写成 potential 差分 F = γΦ(s′) − Φ(s) 时 optimal policy 不变，这个条件和 shaping 的相对大小无关。§4.3 又说 0.3 这个界是从 Ng 推导出来的，而 §3.4 自己说它是 empirical bound。最小的改法是删掉这两处对 Ng 的说法，不加新 claim；这样 [37] 只剩这两处引用，会从 References 里消失，后面的编号自动前移。 原文（ICML 1999）
sections/3_methodology.tex:96 · p6
原文keeps answer correctness dominant, in line with the bounded-potential shaping principle of~\citet{ng1999reward}, which guarantees policy invariance under shaping rewards small relative to the task reward.
改为keeps answer correctness dominant.
sections/4_experiments.tex:32 · p7
原文confirming the bound $\lambda_P\leq0.3$ derived in \S\ref{sec:par} from \citet{ng1999reward}.
改为confirming the empirical bound $\lambda_P\leq0.3$ of \S\ref{sec:par}.
BGE-large-en-v1.5 引错了文献可选
附录 p17、p22 写 BGE-large-en-v1.5 [44]，但 [44] 是 M3-Embedding，也就是另一个模型 BGE-M3 的论文。BGE-large-en-v1.5 的 model card 要求引用 C-Pack（Xiao et al., arXiv 2309.07597，SIGIR 2024）。改法是把 bge 这条 bib 换成 C-Pack，正文不动。 model card · arXiv
references.bib:14 · 附录
原文@misc{bge,
title={M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
year={2025},
eprint={2402.03216},
改为@misc{bge,
title={{C-Pack}: Packed Resources For General {Chinese} Embeddings},
author={Shitao Xiao and Zheng Liu and Peitian Zhang and Niklas Muennighoff and Defu Lian and Jian-Yun Nie},
year={2024},
eprint={2309.07597},
原文Lower thresholds produce more, finer
archetypes; higher thresholds produce fewer, coarser archetypes
改为Higher thresholds produce more, finer
archetypes; lower thresholds produce fewer, coarser archetypes
Table 12 里 slot 和绑定值错位可选
“the slots MISSION, EVENT, YEAR, and PERSON are bound to Voskhod-2 / Apollo / EVA / Belyayev”：按顺序 EVENT 成了 Apollo、YEAR 成了 EVA，和同表的 rollout 以及 G.8 正文的绑定（Voskhod-2、EVA、March 1965、Belyayev）不符。
table/4_exp/tab18_case_pattern_reuse.tex:36 · p29
原文bound to Voskhod-2 / Apollo / EVA / Belyayev at execution time
改为bound to Voskhod-2 / EVA / March 1965 / Belyayev at execution time
Table 13 的 caption 用了表里没有的 Δ可选
Caption 写 “Δ shows per-bucket EM gain over R1-Searcher”，表里的增益印的是绿色的 ↑32.3、↑17.0 等，全表没有 Δ。
原文Existing RL-RAG agents supervise training with answer-level signals on the predicted string~\citep{search_r1, r1_searcher} or entity-presence checks on the constructed graph~\citep{graph_r1}.
改为Existing RL-RAG agents supervise training with answer-level signals on the predicted string, whether they retrieve passages~\citep{search_r1, r1_searcher} or graph facts~\citep{graph_r1}.
sections/3_methodology.tex:59 · p4
原文closing a loophole that entity-presence rewards leave open~\citep{graph_r1}.
改为closing a loophole that answer-level rewards leave open~\citep{graph_r1}.
改为Graph-R1~\citep{graph_r1} & F$_1$ on answer + format &
sections/7_appendix.tex:1082 · p39
原文Graph-R1's entity-presence proxy is the closest prior attempt to interrogate the constructed graph (it checks whether gold-answer entities appear in the graph), but the check is binary and grants the same reward to graphs that contain the gold entity buried among $20$ other triples versus graphs that contain it among $3$ relevant triples.
改为Graph-R1 retrieves from a constructed hypergraph but rewards only format and answer F$_1$, so it grants the same reward to a graph that buries the gold entity among $20$ other triples and to one that contains it among $3$ relevant triples.
改为Search-R1~\citep{search_r1}, R1-Searcher~\citep{r1_searcher}, Graph-R1, Structure-R1, and ReSearch~\citep{chen2025research} share answer-level outcome rewards, while FrugalRAG~\citep{frugalrag2026} rewards evidence recall and stopping at the right hop;
sections/7_appendix.tex:1072 · p39
原文FrugalRAG~\citep{frugalrag2026} & EM + budget penalty &
需作者决定：(a) 写明参照对象，例如 “by +15.0 Avg-OOD EM over R1-Searcher”；(b) 改成对最强 baseline 的差。§1（1_introduction.tex 第 53 行）是同一句，要一起改。
§4.2 “rank-1 on EM and F1 across all seven datasets”建议改
Table 1 自己把 PatternBloom (7B) 的 2Wiki EM/F1 标成次佳（Structure-R1 是 69.2 / 75.9），Bamboogle F1 也是次佳（8B 的 63.7 更高）。按表，7B 在 6 个数据集上 EM 排第一。附录 F（“rank-1 on every held-out dataset”）和 G.13（“rank-1 EM on every dataset”）有同样的说法。
需作者决定，这里没有给出改动
需作者决定口径，例如 “rank-1 on EM on six of the seven datasets (second to Structure-R1 on 2Wiki)”，或只说 Avg-OOD。
GPM 是从 Stage I 最佳 checkpoint 的轨迹蒸馏出来的，Stage I 结束后才建成，不可能在 Stage I 训练时就冻结，也不是按 corpus 构建的。
sections/7_appendix.tex:546 · p22
原文they are built once per corpus and frozen for both
Stage~I and Stage~II training as well as deployment-time inference.
改为the hypergraph index and the FAISS encoder are built once per corpus and frozen for both
Stage~I and Stage~II training as well as deployment-time inference, while the GPM is distilled once between Stage~I and Stage~II and frozen for Stage~II training and deployment-time inference.