Repository navigation
docs(self-evolution): Trustworthy RSI（可信递归自我改进）精读与通俗拆解入库（145，I-243） - #1221
Merged
ThreeFish-AI merged 3 commits intoOct 31, 2026
Merged
ThreeFish-AI merged 3 commits into
ThreeFish-AI merged 3 commits into
Conversation
…B/U/RI 形式化 + S1–S4 改进环 + 四主题风险-保障矩阵 + 成对延续评估协议，附原型与 5 张论文原图; - 新增 145 精读（arXiv:2610.35193v1，2026-10-26，Jiayue Pu 等 12 人·人大高瓴，301 引文）：把「AI 自己改自己」拆成可审计对象——归纳偏置 B / 偏置更新函数 U / 递归算子 RI 形式化（式 1–3、表 2 三问判据：Promptbreeder 是「B 变 U 不变」、Gödel machine 是理论前辈）；S1–S4 改进环 + 三层主张（update/successor/meta 一层证据只撑一层，编码 agent 三变体归层）；四主题 × S1–S4/RI 风险-保障矩阵（四类已证失效与 Hyperagents 基准投毒 HTTPS 案、DP 自适应组合、审批权外置原则）；成对延续 + 总成本核算 + 联合判据评估协议（26 篇 RSI 标签研究仅 4 篇做过严格比较）；五规律三争议五批判边界。 - 信源核验：webReader 首取为幽灵全文（正文含不存在引用、作者块幻影为「单作者」，实为 12 人），一切以 curl 快照为准；Checker 28 断言对账 27 PASS + C1 作者修正全链落地（I-242：双核验必须覆盖元数据域）。 - 配套纯标准库原型 trustworthy_rsi_lab.py（NB-LLR 玩具世界 S1–S4+RI 全链，双跑字节级一致；E1–E5 破坏实验：假提升 +5.3pp/−20.9pp、约束违规 0→158、同源过拟合 0.70→1.0 而 0.95→0.72、假 meta 被成对延续揭穿 delta=0、无护栏停摆）；5 张论文原图入 source-figures（fig10 等比降采样过 1280 KiB 门）并登记台账。 - 登记面：research readme 第六节、knowledge-map Research 节、source-figures README、CHANGELOG、issue I-242。 - 验收：保真核对两轮全绿（INV/FENCE/REFS/FIG/CLEAN）；盲评 §6 双判官均判成稿胜；外行四测四项出闸（闭卷证诊断力、LLM 代理作答）；补阙 5 处增量核验 5/5；字数门 5420/6000；pre-commit 全绿。 🤖 Generated with [Claude Code](https://github.com/claude), [CodeX](https://openai.com), [Gemini](https://github.com/apps/gemini-code-assist) Co-Authored-By: Aurelius Huang<threefish.ai@gmail.com>
冲突处置：CHANGELOG 与 knowledge-map 取双侧并集；issue.md 双方撞号 I-242——对方线性注意力条目（#1222）已合主干保留 I-242，我方 Trustworthy RSI 条目重编为 I-243 并留映射注记。 🤖 Generated with [Claude Code](https://github.com/claude), [CodeX](https://openai.com), [Gemini](https://github.com/apps/gemini-code-assist) Co-Authored-By: Aurelius Huang<threefish.ai@gmail.com>
ThreeFish-AI
changed the title
…1–3 单行 $$ 围栏改多行修存量 \tag ParseError，标题/alt/frontmatter 按仓惯例保持纯文本; 🤖 Generated with [Claude Code](https://github.com/claude), [CodeX](https://openai.com), [Gemini](https://github.com/apps/gemini-code-assist) Co-Authored-By: Aurelius Huang<threefish.ai@gmail.com>
ThreeFish-AI
deleted the
ThreeFish-AI/trustworthy-recursive-self-improvement-notes
branch
November 1, 2026 03:15
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
变更概要
self-evolution 145 精读入库：arXiv:2610.35193v1《Trustworthy Recursive Self-Improvement in the Era of Large Models》（Jiayue Pu 等 12 人，人大高瓴，2026-10-26，301 引文）——把「可信递归自我改进」从口号拆成可审计对象的操作化综述。
正文
docs/research/self-evolution/145-trustworthy-rsi.md（≈5.4k 字，含 11 节 + IEEE 参考）：B/U/RI 形式化（式 1–3、表 2 三问判据：Promptbreeder 是「B 变 U 不变」、Gödel machine 是理论前辈）；S1–S4 改进环 + 三层主张（update/successor/meta，一层证据只撑一层）；四主题 × S1–S4/RI 风险-保障矩阵（四类已证失效与 Hyperagents 基准投毒 HTTPS 案）；成对延续 + 总成本核算 + 联合判据的评估协议（26 篇 RSI 标签研究仅 4 篇做过严格比较）；五规律三争议五批判边界。原型
assets/trustworthy_rsi_lab.py（纯标准库、固定种子、双跑一致）：朴素贝叶斯对数似然比（NB-LLR）玩具世界跑通 S1–S4+RI 全链，E1–E5 破坏实验实测（假提升 +5.3pp/−20.9pp、约束违规 0→158、同源过拟合 0.70→1.0 而 0.95→0.72、假 meta 被成对延续揭穿 delta=0、无护栏停摆）。原图 5 张入
docs/assets/source-figures/self-evolution/（fig10 等比降采样过 1280 KiB 门）并登记台账。登记面：research readme 第六节、knowledge-map Research 节、source-figures README、CHANGELOG、issue I-242。
验收：源稿对账 28 断言（27 PASS + C1 作者块幻影修正）；保真核对两轮全绿（INV/FENCE/REFS/FIG/CLEAN）；盲评 §6 双判官均判成稿胜；外行四测四项出闸（闭卷证诊断力，LLM 代理作答）；补阙 5 处增量核验 5/5；字数门 5420/6000；pre-commit 全绿。
数学式 LaTeX 化（2026-10-30 追加）：正文 54 处裸符号/下标式（h/B/U/E/F/a/s/r、B_{t+1} 等）补行内
$...$定界（更新算子与 display 公式一致用\mathcal{U}）；式 1–3 单行$$围栏改多行，修复\tag在行内模式触发 KaTeX ParseError 的存量渲染缺陷；标题/alt/frontmatter 按仓惯例保持纯文本；经生产插件栈（remark-math + remarkMathSanitize + rehype-katex）全文渲染终验（display 3、零 katex-error、$7k字面保留）。🤖 Generated with Claude Code