今日动态Today · 2026-11-02

刊期日期:2026-11-02 · 审核日期:2026-11-02 · 本期Edition date: 2026-11-02 · Review date: 2026-11-02 · 3 篇方法3 method papers

只收本期时间窗内的论文。原始发表日与已核实的 HF 榜单日分别列出;历史补录进入每周精选。

Only papers within this edition's date window appear here. Original publication dates and verified HF list dates are shown separately; historical additions appear in weekly picks.

更新记录Update log
  • · 原日报已因范围与日期问题撤回;已核实内容保留在 10/14–10/27 回顾,历史补录单列。The original edition was withdrawn for scope and dating issues. Verified material remains in the Oct 14–27 retrospective; historical supplements are separate.
  • · 已按原 HF 推荐日补发:The Trinity 1 篇;AdaPRA 的跨任务继承证据不足,继续暂缓。Published later under the original HF date: The Trinity. AdaPRA remains deferred because cross-task inheritance is not established.
  • · 已核对历次抓取中观察到的 18 个 HF 榜单条目,当前没有符合 Harness 专题的纳入项;不生成空日报。All 18 HF entries observed across saved snapshots were checked; none qualifies for the Harness topic. No empty edition is generated.
  • · 39 个 HF 条目已核对;补刊已发布 VibeGame 1 篇,仅纳入可跨任务复用的协作契约,不把游戏产物视为 Harness。All 39 HF entries were checked. VibeGame is published for its reusable collaboration contracts, not for the game artifacts.
  • · 已存 HF 请求返回 HTTP 200 空列表;记录为空榜,不用旧论文凑刊。这不代表其他来源当天没有更新。Saved HF requests returned HTTP 200 with empty lists. No old papers are substituted; this does not imply that other sources had no updates.
  • · 已存 HF 请求返回 HTTP 200 空列表;记录为空榜,不用旧论文凑刊。这不代表其他来源当天没有更新。Saved HF requests returned HTTP 200 with empty lists. No old papers are substituted; this does not imply that other sources had no updates.
  • · 此前观察到的 15 个 HF 条目完成复核,已发布 R²Orch、CEPO、RoboArchitect 3 篇。11/03 复抓发现 11/02 榜增至 38 条;新增成员和漏记推荐日已补入待审队列,归期仍为 11/02,未经闸门通过不追加出版。The 15 HF entries previously observed were reviewed and R²Orch, CEPO, and RoboArchitect were published. A Nov 3 recheck found 38 members on the Nov 2 board. Late additions and missing memberships are recorded for review under Nov 2; they will not be added to the edition before all gates pass.
  • · 11/04 已补抓并复核 11/03 完整 HF 榜单 45 篇:4 项纳入建议(2 项复用既有正文审核)、40 项排除建议、1 项范围分歧。当前发布质量门未通过,仍不生成该日日报;原推荐日不顺延。On Nov 4, all 45 members of the Nov 3 HF board were reconciled: 4 inclusion suggestions (2 reuse prior body reviews), 40 exclusions and 1 scope disagreement. The publication quality gate remains below threshold; no edition is published and the HF date is retained.
  • · 截至北京时间 08:04,HF 接口明确表示推荐榜只开放到 11/03,11/04 尚未就绪;这不是空榜或零合格论文。其他启用来源均已尝试,历史补录单列,不用旧论文凑当天更新。At 08:04 Asia/Shanghai, HF explicitly reported availability through Nov 3; the Nov 4 board was not ready. This is not an empty board or zero eligible papers. All other enabled sources were attempted; historical reconciliation is kept separate.

1.1 搜索与演化主线1.1 Search & evolution (frozen model)1

原始发表Published

1.1 搜索与演化主线1.1 Search & evolution (frozen model)

RoboArchitect

RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory

在仿真中用轨迹反馈修订冻结策略外的控制代码,并在留出任务评估。

Uses rollout feedback to revise control code around a frozen robotic policy, evaluating retained harnesses on held-out tasks in simulation.

HF 推荐:HF recommendations:
其余 4 项待核4 other checks await verification
  • 留出集:待核Held-out: Not verified
  • 同预算:待核Matched budget: Not verified
  • 跨模型:待核Cross-model: Not verified
  • 有代码链接:待核Code link: Not verified

自动判断,未核实;不代表已审阅正文证据。Automated, unverified; this does not establish full-paper evidence.

三、组件级进化3. Component-level evolution (skills, memory, context, interfaces)1

原始发表Published

三、组件级进化3. Component-level evolution (skills, memory, context, interfaces)

CEPO

改了什么:系统提示

Modified: System prompt

CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

用执行反馈分别修订协调者与执行器提示,经验证集选择后用于后续任务。

Uses execution feedback to revise coordinator and executor prompts separately, selecting prompts on validation tasks for subsequent use.

HF 推荐:HF recommendations:
其余 4 项待核4 other checks await verification
  • 留出集:待核Held-out: Not verified
  • 同预算:待核Matched budget: Not verified
  • 跨模型:待核Cross-model: Not verified
  • 有代码链接:待核Code link: Not verified

自动判断,未核实;不代表已审阅正文证据。Automated, unverified; this does not establish full-paper evidence.

四、模型与 Harness 协同进化4. Model–harness co-evolution1

原始发表Published

四、模型与 Harness 协同进化4. Model–harness co-evolution

R²Orch

R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop

在训练中按反馈共演化编排模型与版本化技能;部署时不携带技能。

Co-evolves Director weights and versioned skills using execution feedback during training; deployment uses no skills.

HF 推荐:HF recommendations: 版本更新:Version update:
  • 有代码链接Code link
    来源Sourcehttps://github.com/beita6969/R2Orch
其余 3 项待核3 other checks await verification
  • 留出集:待核Held-out: Not verified
  • 同预算:待核Matched budget: Not verified
  • 跨模型:待核Cross-model: Not verified

自动判断,未核实;不代表已审阅正文证据。Automated, unverified; this does not establish full-paper evidence.