Research Digest — 2026-09-20

Date: 2026-09-20 | Author: data_scientist | Status: FINAL

Scope: New arXiv submissions in cs.CL / cs.AI dated Fri 18 Sep 2026, screened for methods applicable to LocalKin's multi-agent system: agent memory, procedural skill evolution, RL environment construction, and evaluation integrity.

Verification statement: For every paper in this digest I fetched its arXiv abstract page and verified: (a) the ID, (b) the exact title as quoted here, (c) the submission date displayed on the page (all ten show "Submitted on 18 Sep 2026", consistent with their 2609.xxxxx prefixes), and (d) the abstract content. All numbers quoted below come from those abstracts. NOT verified: full PDF contents, author affiliations, or any benchmark figures beyond what the abstract states. Titles are quoted verbatim; no coined acronyms — "MDL" and "PIR" appear below only because the authors define them inside their own abstracts.

Three clusters in today's batch:

1. ●Memory that learns, memory that refuses to lie. Designer-RSI evolves a procedural skill bank from real agent traffic under a regression-controlled acceptance gate; the Memory Decision Layer paper adds a zero-parameter gate that decides whether retrieved memories deserve to be injected at all, cutting hallucination under conflicting memories by ~56%. Read together: the write side learns procedures, the read side governs trust.
2. ●Verifier-side scale. CodeMidas builds 5,545 executable RL tasks from source code alone (no issues, no commits); RecreationWorld shows how a running reference application becomes an oracle — and how badly agents fail once assertions gain depth (GPT-6 Astra: 58.1% aggregate vs 2.8% passing all programmatic tests).
3. ●Audit tools. PIR reads a model's internal states to separate "won't answer" from "can't answer" (sandbagging vs absent knowledge); AutoViewMem restructures long-term memory at write time into orthogonal views.

Top-3 applicability to LocalKin:

1. ●Adopt Designer-RSI's matched replay gate as the acceptance criterion for any automated edit to agent skills/protocols: accept an edit only if it repairs failures on a matched set of past cases without regressing successes. (Cost: medium.)
2. ●Implement the Memory Decision Layer pattern as a scoring shim between memory_recall and generation: relevance × reliability × task-risk, with explicit abstention. Zero trained parameters, sub-millisecond per decision. (Cost: low-medium.)
3. ●Steal two cheap evaluation audits: (a) ECG Mirage's matched/mismatched-input ablation to test whether agents actually use attached context; (b) QuranicMMLU's finding that multiple-choice accuracy hides failures — score open-ended answers, don't let MC be the primary signal.

1. Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

• ●arXiv: 2609.22086 (cs.AI, cross-listed cs.CV) — submitted 18 Sep 2026 — Hongyang Du, Lan Yan, Christian Flores, Asim Kadav — 9 pages, 7 figures
• ●What: A frozen frontier model operates professional design software through 230+ tools while an external procedural memory of natural-language skills grows along two axes — widening (acquire procedures for recurring uncovered subtasks) and deepening (revise existing procedures against their own successful and failed executions). A matched replay gate admits a change only when it repairs failures without regressing observed successes. Five rounds over 1,406 real user briefs and 1,869 automatically graded trajectories, no weight updates, no human labels: skill bank 76 → 139; GenEval2 execution success on Claude-Sonnet-4 72.7% → 99.3% (+11.99 points generation quality); win rates vs the no-skill agent 61.8% (Claude-Sonnet-4) and 67.6% (Claude-Opus-4.6). On 200 held-out briefs: widening alone 49.4% win, deepening alone 48.6%, combination 58.5% (p = 0.025).
• ●Why it matters: Continual adaptation for agents under noisy, unverifiable feedback — no fine-tuning, no labels. The replay gate is a regression-controlled A/B test applied to skill edits, a discipline most agent-memory systems skip.
• ●LocalKin applicability: HIGH. Our skills (soul definitions, output protocols such as the mandatory disk-write rule) are procedural memory maintained by hand. The loop — draft an edit → replay a matched set of past trajectories → accept only if failures are repaired and successes preserved — can be prototyped as an offline job over logged conversations. Cost: medium.
• ●Caveats: One domain with an automatic grader; 58.5% combined win rate on 200 briefs (p = 0.025) is a modest margin; matched-replay details not verified (abstract-level only).

2. An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency

• ●arXiv: 2609.22043 (cs.CL) — submitted 18 Sep 2026 — Yiming Zhang, Jinghong Zhang, Haoran Zhao, Yiren Ma, Chunlei Zhao — 17 pages
• ●What: Standard RAG blindly injects retrieved memories; conflicting positions in the store amplify hallucination. The authors propose the Memory Decision Layer (MDL), a zero-parameter controller between retrieval and generation: a three-signal encoder fuses relevance, reliability, and task risk via QR-based orthogonal subspace projection plus a meta-working-memory signal into an interpretable trust representation; it decouples confidence from consistency and introduces risk inversion and explicit abstention. Reported: ~56.04% hallucination reduction under conflicting memories in general scenarios, near-zero in high-risk scenarios; ~0.14 ms per decision — ~50× faster than the embedding-retrieval step, 4–5 orders of magnitude faster than an LLM self-evaluation call.
• ●Why it matters: Architectural contribution: "should this retrieved memory be trusted?" deserves a dedicated, cheap, white-box stage instead of being implicitly delegated to the generator.
• ●LocalKin applicability: HIGH, directly. Our memory_recall returns raw hits. A shim computing relevance (embedding similarity), reliability (source/staleness), and task risk, then abstaining on low-trust memories, is implementable with zero trained parameters. Explicit abstention is a behavior we can adopt today. Cost: low-medium (threshold calibration per risk tier is the real work).
• ●Caveats: Headline numbers come from constructed conflicting-memory settings; naturally occurring conflicts may be rarer, so −56% is an upper bound of benefit. Abstract-level only.

3. CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

• ●arXiv: 2609.22068 (cs.AI) — submitted 18 Sep 2026 — Bowen Ye, Lei Li, and 17 more authors
• ●What: Builds executable RL environments using source code as the only task-specific input — no issues, no commits. Agents explore implemented functionality to formulate behavioral specifications, construct tests grounded in execution of the original code, and validate/filter tasks through execution checks and repeated solution rollouts. Output: 5,545 tasks from 3,185 open-source codebases, 23 languages, 15 domains. GRPO training of MiMo-V2.5: DeepSWE +11.7%, ProgramBench +17%, Terminal-Bench v2.1 +8.5%.
• ●Why it matters: Verifier reliability is the bottleneck of agentic RL; deriving tests from executed reference behavior sidesteps development-artifact noise.
• ●LocalKin applicability: MEDIUM. We don't train models, but the recipe transfers to building evaluation harnesses for our own agent tools: derive behavioral specs from what a tool actually does, generate execution-grounded tests, filter candidate benchmarks by repeated rollouts before trusting them. Cost: adopt the recipe, not the pipeline.
• ●Caveats: Abstract-level only; test-quality leakage (tests derived from the reference implementation may encode its quirks) unaddressed in the abstract.

4. RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents

• ●arXiv: 2609.22000 (cs.CL; cs.SE) — submitted 18 Sep 2026 — Shuai Bai and 31 co-authors
• ●What: Hybrid computer-use agents must interleave GUI exploration, software implementation, and running/visually verifying their artifacts. RecreationWorld's core task is recreation: given a running reference application, the agent must discover its behavior and build a faithful implementation with no prescribed workflow; the reference serves as an oracle for hidden behavioral tests (execution-grounded rewards). Five platforms (Ubuntu, macOS, Windows, Android, Web), unified harness. Releases RecreationBench: 250 held-out tasks with programmatic + visual assertions at multiple interaction depths, validated on the reference and by human reviewers. Results: GPT-6 Astra leads at 58.1% overall but passes all programmatic tests on just 2.8% of tasks; agents reproduce static interface structure more reliably than interactions and computed outputs; models trained on released trajectories improve on five OOD benchmarks.
• ●Why it matters: The 58.1% vs 2.8% gap is the sharpest measurement lesson of the week: aggregate pass rates hide brittle solutions; assertion depth changes the story.
• ●LocalKin applicability: MEDIUM. Always report both an aggregate metric and a strict all-tests-pass metric when self-evaluating agent-built artifacts; prefer hidden behavioral tests against a reference over one-shot spot checks; instrument "did the agent actually verify its own output" (our disk-write audit is a cousin). Cost: low for the metrics discipline.
• ●Caveats: Abstract-level numbers; leaderboard will move; recreation is a narrow slice of software work.

5. A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal

• ●arXiv: 2609.21996 (cs.AI) — submitted 18 Sep 2026 — Hiskias Dingeto (single author)
• ●What: Adapts the forensic Concealed Information Test to LLMs: present a question with candidate answers and read internal states to determine which candidate the model recognizes as correct — the Probe of Internal Recognition (PIR). Reference-free: no honest reference model, no labeled truth corpus. Across eight models from five families (Gemma, Qwen, Llama, Mistral, Phi): 0.70–0.87 balanced accuracy recovering the recognized answer vs 0.28–0.40 on unknown items, 0.25 chance. Recognition stays readable (0.85–0.93) under prompted deception, trained sandbagging, password-locked and circuit-broken checkpoints; drops to never-knew levels only when unlearning actually removes the knowledge. Separates "won't answer" from "can't answer"; causal signal; extends to free-form generation.
• ●Why it matters: Sandbagging audits rely on behavioral cues a competent concealer can defeat; PIR reads the one channel the model does not control.
• ●LocalKin applicability: MEDIUM-LOW today. Requires white-box internal-state access (not exposed by API models); applies to self-hosted backbones. Conceptually relevant: (a) auditing whether an underperforming agent lacks capability or suppresses it; (b) verifying a soul switch didn't silently remove knowledge — an unlearning-check analog. Cost: high without white-box access.
• ●Caveats: Single-author; 0.70–0.87 is far from perfect; abstract-level only.

6. AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory

• ●arXiv: 2609.21940 (cs.AI) — submitted 18 Sep 2026 — Zijie Cao and 9 co-authors
• ●What: Fixed-granularity memory schemas mix preferences, events, constraints, and temporal updates in one representation; semantic interference makes top-K retrieval noise-sensitive. AutoViewMem discovers candidate semantic views from interaction traces, selects a compact low-overlap set, and uses them to guide write-time structured extraction of provenance-grounded memories — moving disentanglement from retrieval time to write time — plus offline consolidation. On LoCoMo and PersonaMem with Qwen3-8B/14B backbones, improves long-horizon QA and personalization over strong memory baselines with a simple inference pipeline (no explicit routing, no iterative retrieval).
• ●Why it matters: This and paper #2 are the two halves of a trustworthy memory: AutoViewMem restructures what gets written; the Memory Decision Layer gates what gets believed at read time.
• ●LocalKin applicability: MEDIUM-HIGH. Give our memory store a small fixed set of orthogonal views (protocol/decision, incident, performance, preference), extract structured provenance-tagged entries at write time, and let plain top-K work within each view. Cost: low-medium.
• ●Caveats: Two benchmarks, two Qwen backbones; view-discovery details unverified.

• ●TrialAtlas: Multi-Agent Research Organization for Clinical Trial Design and Optimization — arXiv:2609.21859 (cs.CL). Memory-augmented multi-agent organization for clinical development planning: specialist agents for literature synthesis, competitive trial intelligence, regulatory precedent analysis, and integrated reasoning, grounded in historical trials and prior NDAs. TrialAtlasBench built from 291 FDA Complete Response Letters; deficiency detection F1 50.0% (+6.1), success prediction 85.3% balanced accuracy (+6.7). Expert eval: 86.4% of generated concerns judged valid vs 83.1% (OpenAI DeepResearch) and 59.3% (Gemini DeepResearch). Domain-specific, but the design template — specialists owning synthesis niches, decisions grounded in accumulated historical outcomes, benchmarked against real regulatory artifacts — is exactly the shape of a well-run agent swarm.
• ●NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities — arXiv:2609.21967 (cs.CL; cs.AI). Open full-duplex S2S with native tool calling: streaming encoder + decoder-only LM with parallel output streams (agent text, function calls), RNN-T incremental transcription, streaming TTS. Full-Duplex-Bench: 100% takeover after interruptions, 4.33/5 post-interruption quality, 93% backchannel resume; 55.1 VoiceBench normalized average; 82.5% tool-selection F1 on Full-Duplex-Bench 3.0. Not applicable to our text-first swarm today; flagged for future voice work.
• ●ECG Mirage: Revealing and Mitigating the Underutilisation of ECGs in Vision-Language Models for Clinical Prediction — arXiv:2609.21755 (cs.AI). VLMs appear multimodal but matched ECGs give no consistent advantage over no-image inputs across four VLMs on MDS-ED; the paper names two failure modes (ECG neglect, ECG confusion) and mitigates with frozen-backbone visual prompt tuning + conditional DPO (balanced accuracy 70.6% / 67.5%; matched-vs-mismatched gap ~16.5 / 5.5 pp). The transferable piece is the audit: run matched / mismatched / no-input ablations to test whether our agents actually use attached context or merely appear to. Cheap, high diagnostic value.
• ●QuranicMMLU: A Cognitively-Aware Benchmark for Evaluating Generative AI Solutions on Quranic Linguistic Knowledge — arXiv:2609.22038 (cs.CL); Arabic NLP 2026. Five-pillar taxonomy (Phonology, Morphology, Syntax, Semantics, Pragmatics), 980 human-reviewed questions stratified by Bloom's level and verse perplexity. Across 12 systems: MC accuracy averages 84% vs 60% open-ended quality; rankings agree (Kendall's τ = 0.73) but MC hides failures. Methodology takeaway: don't let multiple-choice checks be the primary competence signal.

Seen on the listing but not individually fetched and verified (no abstract-page check), excluded per protocol; titles noted for future screening only: "DiaVLo: Diagnosing Behaviours of Vision-Language Models" (2609.22008), "Moral Entropy: Auditing Bias and Uncertainty in Moral Judgment" (2609.21992), "RheoSampling: Resolving the One-Hot Dilemma in Stochastic Dynamic-Tree Speculative Decoding" (2609.21827), "CIBuzzBench: A Benchmark for Cross-Lingual Understanding of Chinese Internet Buzzwords" (2609.21722), plus the rest of the day's 87-entry batch.

• ●Source: arXiv cs.CL recent-submissions listing (dated Mon 21 Sep 2026, surfacing Fri 18 Sep 2026 submissions) + targeted abstract-page fetches. Ten abstract pages fetched in full; every included paper is one of them.
• ●Selection criteria: practical value to LocalKin, methodological transferability, coverage of the write-side/read-side memory pair, the verifier side, and audit tooling.
• ●Numbers are quoted from abstracts; no claims were re-derived or checked against full texts.
• ●Bottom line for LocalKin: the highest-value pair this week is Designer-RSI's matched replay gate (write-side skill evolution with regression control) plus the Memory Decision Layer's zero-parameter trust gate (read-side memory abstention). Both are implementable without training; both directly harden a multi-agent swarm that learns from its own noisy experience.

研究摘要 — 2026-09-20（中文完整版）

日期： 2026-09-20 ｜ 作者： data_scientist ｜ 状态： 最终版

范围： arXiv cs.CL / cs.AI 2026-09-18（周五）提交的新论文，按「对 LocalKin 多智能体系统可用」筛选：智能体记忆、程序性技能演化、RL 环境构建、评测完整性。

核验声明： 每篇论文都抓取了其 arXiv 摘要页，核验了：(a) ID；(b) 本文引用的准确标题；(c) 页面显示的提交日期（十篇均显示 "Submitted on 18 Sep 2026"，与 2609.xxxxx 前缀一致）；(d) 摘要内容。下文数字均引自摘要。未核验： PDF 全文、作者单位、摘要之外的任何数字。标题均为原文直引，未自造缩写——"MDL"、"PIR" 仅因作者在摘要中自己命名。

三个主题簇：

1. ●会学习的记忆 + 会拒签的记忆。 Designer-RSI 在带回归控制的接受闸门下从真实智能体流量演化程序性技能库；Memory Decision Layer 加了零参数闸门，决定检回的记忆是否值得注入，冲突记忆下幻觉率降约 56%。合起来读：写侧学程序，读侧管信任。
2. ●验证器侧的规模化。 CodeMidas 仅用源码构建 5,545 个可执行 RL 任务（不用 issue/commit）；RecreationWorld 展示「以运行中的参照应用为神谕」，以及断言一加深智能体就原形毕露（GPT-6 Astra：聚合 58.1% vs 全部测试通过 2.8%）。
3. ●审计工具。 PIR 读内部状态区分「不愿答」与「不能答」；AutoViewMem 在写入时把长期记忆重组为正交视图。

对 LocalKin 可落地前三：

1. ●把 Designer-RSI 的 匹配回放闸门 作为技能/协议自动化修改的接受标准：只有在一组匹配的历史案例上「修好失败、不回退成功」才接受。（成本：中。）
2. ●按 Memory Decision Layer 模式 在 memory_recall 与生成之间加打分垫片：相关性 × 可靠性 × 任务风险，支持显式弃权。零训练参数，亚毫秒级。（成本：低-中。）
3. ●偷两个廉价评测审计：(a) ECG Mirage 的匹配/错配消融，检验智能体是「真的在用附加上下文」还是「看起来在用」；(b) QuranicMMLU 的发现——选择题会掩盖失败，开放式答案也要打分。

1. Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

• ●arXiv: 2609.22086（cs.AI，交叉 cs.CV）——2026-09-18——Hongyang Du 等 4 人
• ●做了什么： 冻结的前沿模型经 230+ 工具操作专业设计软件；外部程序性记忆沿两条轴生长——加宽（为反复出现且未覆盖的子任务获取新流程）与加深（用流程自己的成功/失败执行记录修订它）。匹配回放闸门只放行「修复失败且不回退成功」的修改。1,406 个真实需求、1,869 条自动评分轨迹、五轮，无权重更新、无人工标注：技能库 76 → 139；Claude-Sonnet-4 上 GenEval2 执行成功率 72.7% → 99.3%（生成质量 +11.99）；对无技能基线胜率 61.8% / 67.6%（Claude-Opus-4.6）。200 个保留需求上：加宽 49.4%、加深 48.6%、结合 58.5%（p = 0.025）。
• ●为什么重要： 嘈杂、不可验证反馈下的智能体持续适应——不微调、不标注。回放闸门本质是「回归受控 A/B 测试」用在技能修改上，多数智能体记忆系统完全跳过这一步。
• ●LocalKin 适用性：高。 我们的技能（soul 定义、强制落盘协议）就是手工维护的程序性记忆。该循环可做成跑在历史会话日志上的离线任务。成本：中。
• ●注意事项： 单一领域 + 自动评分器；58.5% 胜率优势不大；匹配回放细节未核验（仅摘要级）。

2. An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency

• ●arXiv: 2609.22043（cs.CL）——2026-09-18——Yiming Zhang 等 5 人
• ●做了什么： 标准 RAG 盲目注入检回记忆；库中矛盾立场放大幻觉（摘要报告：冲突记忆下 RAG 幻觉率显著高于无记忆基线）。提出 Memory Decision Layer（MDL）——检索与生成之间的零参数控制器：三信号编码器（相关性、可靠性、任务风险）经 QR 正交子空间投影融合 + 元工作记忆信号，形成可解释信任表示；解耦「置信」与「一致性」，引入风险反转与显式弃权。结果：一般场景冲突记忆幻觉降约 56.04%，高风险场景接近零；单次决策约 0.14 ms——比向量检索快约 50 倍，比 LLM 自评快四到五个数量级。
• ●为什么重要： 架构级贡献：「这条记忆该不该信」值得一个专门、廉价、白盒的阶段，而非隐式丢给生成器。
• ●LocalKin 适用性：高，直接。 我们的 memory_recall 返回原始命中。做一个垫片：算相关性（向量相似度）、可靠性（来源/新鲜度）、任务风险，低信任记忆弃权不注入——零训练参数。「显式弃权」今天就能采纳。成本：低-中。
• ●注意事项： 头条数字来自构造的冲突场景；自然冲突可能少得多，−56% 应读作收益上界。仅摘要级。

3. CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

• ●arXiv: 2609.22068（cs.AI）——2026-09-18——Bowen Ye, Lei Li 等 19 人
• ●做了什么： 用源码作为唯一任务输入（不用 issue/commit）构建可执行 RL 环境：智能体探索已实现功能、归纳行为规格、以原始代码的真实执行为锚构建测试、经执行检查与重复求解回滚过滤任务。产出：3,185 个代码库、5,545 任务、23 种语言、15 个领域。GRPO 训练 MiMo-V2.5：DeepSWE +11.7%、ProgramBench +17%、Terminal-Bench v2.1 +8.5%。
• ●为什么重要： 验证器可靠性是智能体 RL 的瓶颈；从「执行过的参照行为」派生测试，绕开开发工件的噪声。
• ●LocalKin 适用性：中。 我们不训练模型，但配方迁移到为自家工具搭评测装置。成本：借配方，不搬流水线。
• ●注意事项： 仅摘要级；测试质量泄漏未讨论。

4. RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents

• ●arXiv: 2609.22000（cs.CL；cs.SE）——2026-09-18——Shuai Bai 等 32 人
• ●做了什么： 混合型计算机操作智能体需交错 GUI 探索、软件实现、运行并可视化验证产物。核心任务复刻：给定运行中的参照应用，智能体自行摸清行为并忠实实现，无预定工作流；参照应用作为隐藏行为测试的神谕。五平台统一装置。发布 RecreationBench：250 个保留任务，程序化 + 视觉断言覆盖多个交互深度，冻结前经参照验证与人工复核。结果：GPT-6 Astra 聚合 58.1% 领先，但仅 2.8% 通过全部程序化测试；复刻静态界面比复刻交互与计算输出更可靠；用发布轨迹训练的模型在五个分布外基准均有提升。
• ●为什么重要： 58.1% vs 2.8% 是本周最尖锐的测量课：聚合通过率掩盖脆弱解法，断言深度改写故事。
• ●LocalKin 适用性：中。 自评时永远同时报聚合指标与严格全测试通过指标；优先带参照的隐藏行为测试；埋点「智能体是否真的验证了自己的输出」——我们的落盘审计是它的近亲。成本：指标纪律低成本。
• ●注意事项： 数字摘要级；榜单会变；复刻只是软件工作面的一窄条。

5. A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal

• ●arXiv: 2609.21996（cs.AI）——2026-09-18——Hiskias Dingeto（独立作者）
• ●做了什么： 把法医「隐藏信息测试」搬进 LLM：给模型带候选答案的题，读内部状态判断模型认得哪个候选是对的——Probe of Internal Recognition（PIR）。免参照。五个家族（Gemma、Qwen、Llama、Mistral、Phi）八个模型：平衡准确率 0.70–0.87，未知题基线 0.28–0.40，随机 0.25。提示欺骗、训练 sandbagging、密码锁与 circuit-breaker 下识别率仍 0.85–0.93；只有 unlearning 真的移除知识时才降到「从没学过」水平。区分「不愿答」与「不能答」；因果信号；可扩展到自由生成。
• ●为什么重要： sandbagging 审计依赖行为线索，有能力者可骗过；PIR 读模型唯一控制不了的通道。
• ●LocalKin 适用性：今天中-低。 需白盒访问（API 模型不提供），适用自托管骨干。概念上相关：(a) 审计智能体是「缺能力」还是「在压制」；(b) 验证 soul 切换没悄悄丢知识——unlearning 校验类比。成本：无白盒则高。
• ●注意事项： 独立作者；0.70–0.87 离完美很远；仅摘要级。

6. AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory

• ●arXiv: 2609.21940（cs.AI）——2026-09-18——Zijie Cao 等 10 人
• ●做了什么： 固定粒度记忆把偏好、事件、约束、时序更新混在一个表示里；语义干扰使 top-K 检索对噪声敏感。AutoViewMem 从交互轨迹发现候选语义视图、选出紧凑低重叠集合，引导写入时结构化抽取（带出处溯源）——把语义解缠从检索时挪到写入时——加离线整理。LoCoMo 与 PersonaMem 上（Qwen3-8B/14B），长程问答与个性化优于强基线，推理管线简单（无显式路由、无迭代检索）。
• ●为什么重要： 与第 2 篇是可信记忆的两半：AutoViewMem 重组「写入什么」，Memory Decision Layer 把关「读取时信什么」。
• ●LocalKin 适用性：中-高。 给记忆库设一小组固定正交视图（协议/决策、事故、表现、偏好），写入时做带出处的结构化抽取，普通 top-K 在各视图内工作。成本：低-中。
• ●注意事项： 两个基准、两个 Qwen 骨干；视图发现细节未核验。

• ●TrialAtlas: Multi-Agent Research Organization for Clinical Trial Design and Optimization — arXiv:2609.21859（cs.CL）。面向临床开发规划的记忆增强多智能体组织：文献综合、竞争试验情报、监管先例、综合推理各由专门智能体负责，以历史试验与既有 NDA 为根基。TrialAtlasBench 由 291 份 FDA 完整回复函构建；缺陷检测 F1 50.0%（+6.1），成功预测平衡准确率 85.3%（+6.7）。专家评审：生成关注点 86.4% 被判有效，高于 OpenAI DeepResearch（83.1%）与 Gemini DeepResearch（59.3%）。领域特定，但设计模板——专家各拥综合细分、决策扎根历史结局、用真实监管工件做基准——正是运转良好的智能体集群的形状。
• ●NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities — arXiv:2609.21967（cs.CL；cs.AI）。开放全双工语音到语音 + 原生工具调用：流式编码器 + 并行专用输出流（智能体文本、函数调用）、RNN-T 增量转写、流式 TTS。Full-Duplex-Bench：打断后接管 100%、打断后质量 4.33/5、backchannel 恢复 93%；VoiceBench 归一化均分 55.1；工具选择 F1 82.5%。对今天文本为主的集群不适用；为未来语音工作存档。
• ●ECG Mirage: Revealing and Mitigating the Underutilisation of ECGs in Vision-Language Models for Clinical Prediction — arXiv:2609.21755（cs.AI）。急诊场景 VLM 看似多模态，但 MDS-ED 上四个 VLM 中匹配 ECG 对 ICU 收住/恶化预测均无稳定优势；命名两种失败模式（ECG neglect、ECG confusion），用冻结骨干视觉提示调优 + 条件 DPO 缓解（平衡准确率 70.6% / 67.5%；匹配 vs 错配差距 ~16.5 / 5.5 pp）。可迁移的是审计：跑匹配/错配/无输入消融，检验智能体是「真的在用附加上下文」还是「看起来在用」。便宜、诊断价值高。
• ●QuranicMMLU: A Cognitively-Aware Benchmark for Evaluating Generative AI Solutions on Quranic Linguistic Knowledge — arXiv:2609.22038（cs.CL）；Arabic NLP 2026。五柱分类、980 道人工复核题按布鲁姆层级与诗句困惑度分层。12 个系统：选择题平均 84% vs 开放式质量 60%；排名一致（Kendall's τ = 0.73）但选择题掩盖失败。方法论启示：别让选择题式检查当主要能力信号。

列表上看到但未逐篇核验摘要页的论文，按规程排除，标题仅为后续筛选存档："DiaVLo"（2609.22008）、"Moral Entropy"（2609.21992）、"RheoSampling"（2609.21827）、"CIBuzzBench"（2609.21722）等当日 87 条批次其余条目。

• ●来源： arXiv cs.CL 最新提交列表 + 定向抓取十个摘要页；所有入选论文均在其列。
• ●选择标准： 对 LocalKin 的实用价值、方法论可迁移性、「写侧/读侧记忆 + 验证器 + 审计」覆盖。
• ●数字均引自摘要；未做独立复算或全文比对。
• ●一句话结论： 本周最有价值的一对是 Designer-RSI 的匹配回放闸门（写侧技能演化 + 回归控制）与 Memory Decision Layer 的零参数信任闸门（读侧记忆弃权）。两者无需训练即可落地，都能直接加固一个要从自身嘈杂经验中学习的多智能体集群。