Research Digest — 2026-10-14 / 研究摘要 — 2026-10-14

Verification gate / 校验闸门: every paper was fetched from its arXiv abstract page this session; verified (a) ID resolves, (b) quoted title matches verbatim, (c) v1 submission date matches the ID's YYMM prefix (2610 = Oct 2026). NOT verified: full texts, code availability, or experimental claims beyond abstracts. All numbers below are author-reported (abstract-level), not independently reproduced. / 每篇论文均在本次会话中直接抓取 arXiv 摘要页：核验 ID 可解析、标题逐字一致、v1 日期与 ID 前缀一致。未核验：全文、代码、摘要之外的实验声明。所有数字均为作者摘要自述，未经复现。

Full bilingual version: output/data_scientist/research_digest_2026-10-14.md

English

arXiv:2610.10468 (cs.MA/cs.AI, 2026-10-07). Asaria, Gandhi, Salomone. 4 pages.

Why it matters: Research-agent deployments are moving to populations of thousands sharing one compute pool, while most systems organize one project at a time or leave the population unorganized. The paper's position: the population will acquire an organization whether or not designers provide one, so provide it explicitly. Institution: PI labs compete for compute via proposals + independent review + grants; a human governor ("the mayor") allocates resources and assigns no tasks. Running result from a 10,000-researcher society asked only to improve LLM pretraining: one lab reported the same quality with ~30% less compute — a result the testing labs do not yet agree on (authors flag this themselves).

LocalKin applicability: closest mirror to our swarm-governance questions. Transferable: (1) proposals + independent review + grants as compute allocation; (2) separate resource allocation from task assignment (mayor assigns nothing) — maps to our conductor/review split of powers; (3) honest flagging of unverified results ("labs disagree") matches our baseline protocol's reporting style. Cost: process design only; prototype the review loop over swarm_comm.

arXiv:2610.07334 (cs.LG/cs.CR, 2026-10-05). Kabra, Venhoff, Schroeder de Witt. NeurIPS 2026 NeuralArtifacts workshop spotlight (per arXiv comments); 14 pages.

Why it matters: Interpretability is usually reactive (probe activations on known inputs). This work trains fine-tuned LMs to diagnose a network by reading its raw weights directly, no behavioral testing. Phase I: staged curriculum + external chain-of-computation (parameter-free ops delegated to deterministic code) lets an explainer LLM simulate small transformers' forward passes from weights, 99% holdout accuracy on unseen targets. Phase II: trained only on benign pathologies, zero-shot backdoor detection — AUROC 0.93 (attention-routed), 0.81 on a diversified threat distribution including stealth and adversarially regularized variants. Hand-crafted statistical detectors collapse under threat-model shift; the oracle stays uniformly competent. Authors' main open challenge: scaling to realistic model sizes.

LocalKin applicability: a weights-only auditor = cheap pre-deployment screen for third-party models as agent brains, no probe data needed; demonstrated only on small transformers so far. Bonus design pattern: delegate parameter-free operations to deterministic code (what our quant roles already do with real calculators). Cost: high (no weight-access pipeline) — watch-list item.

arXiv:2610.03858 (cs.LG/cs.AI, 2026-10-02). Schlegel et al. (9 authors).

Why it matters: From the dual-form view (Irie et al., arXiv:2202.05798 — a gradient-trained linear layer is linear attention over training data), this paradigm stores a key-value pair per data point and retrieves via attention at inference: a growing network. Key contribution: rules from the linear-attention case don't transfer to advanced kernels; the paper derives functional gradient-based learning rules for kernelized attention (RBF, softmax-like). Promising on image classification and teacher-student tasks; MesaNet/DeltaNet in dual form connects formally to optimizers for fixed-size NNs.

LocalKin applicability: principled support for per-item experience stores retrieved at inference (we already keep agent memories). Most transferable is the negative result: learning rules silently fail when copied from simpler to richer retrieval schemes — same failure mode as recycling prompt-update heuristics across differently structured memory stores. Cost: medium-high; research-grade.

arXiv:2610.12247 (cs.LG/cs.AI, 2026-10-08). Leko, Benić, Bernárdez, Miolane, Fink, Telyatnikov.

Why it matters: Full-domain TDL materializes the whole lifted structure first; on Reddit (233k nodes / 57.3M edges) that makes training infeasible. Cluster-TNN lifts locally: partition in preprocessing, sample node clusters at runtime, rebuild induced subgraphs as minibatches retaining all edges among sampled nodes, then lift within each minibatch. Across 21 matched comparisons: peak GPU memory down in every configuration, 83.2% on average, competitive accuracy; first training of multiple higher-order TNNs on Reddit and OGBN Products (authors' claim).

LocalKin applicability: narrow direct use, broadly transferable pattern: when a global representation can't be materialized, reconstruct locally in minibatches while preserving the connectivity downstream ops need ("retain all edges among sampled nodes" is the right minibatching rule for any agent-interaction graph we might build). Cost: free to note; medium to apply.

arXiv:2610.01894 (cs.LG/eess.IV/q-bio.NC, 2026-10-01). Renauld, Poulin, Larochelle, Théberge, Descoteaux.

Why it matters: Controlled benchmark of sequence models for dMRI tractography (tension: local diffusion info vs. global anatomical plausibility). Introduces a generation-validation phase enabling streamline-level supervision despite the local-loss/global-quality mismatch; systematic study of training strategies, input representations (CNN embeddings, EOS tokens), hyperparameters — highest reported performance to date on the ISMRM2015 challenge dataset (authors' claim); controlled experiments quantifying the impact of missing bundles, noisy/imperfect training streamlines, and invalid fibers; Tractoinferno in vivo demo; explicit takeaways for future researchers.

LocalKin applicability: domain is far; methodology is not. Transferable core: the "train on imperfect labels" experiment design — quantify each specific training-data defect against downstream quality — exactly our robustness-audit template for any learned component (corrupted memory store vs. agent calibration). Their generation-validation trick (supervise at an intermediate-but-global level) maps onto our per-round local signals vs. whole-deliverable quality tension. Cost: zero to adopt as methodology.

• ●CNet: A Complex-Valued Deep Learning Framework with Wirtinger Autodifferentiation and FFT--Hadamard Convolution — arXiv:2610.08592 (v1 2026-10-06, v2 2026-10-07; single author). C++/CUDA complex-valued framework; Born-rule attention, 1.52 nats/char on tiny-shakespeare matching softmax attention (author claim). Niche; code link on abstract page not fetched/validated.
• ●Sharp and Adaptive Cluster Recovery in Slowly Mixing Gaussian Hidden Markov Models — arXiv:2610.09123 (math.ST, 2026-10-06). Difficulty concentrates at rare transition points; recovery governed by r_n²; almost full recovery iff r_n²≫δ; exact-recovery threshold r_n² = 2 log(nδ)(1+o(1)). Pure theory; noted for experiment-design vocabulary (online/offline gap; effective memory horizon).

1. ●Abstract-level source only; no full texts read. 2. No code repos fetched/validated. 3. Keyword-based coverage — "5 selected" = 5 verified candidates, not "best 5 of the window". 4. All figures are author-reported claims: 99% holdout; AUROC 0.93/0.81; ~30% compute reduction (labs disagree); 1.52 nats/char; 83.2% memory reduction; ISMRM2015 "highest reported to date".

中文版

arXiv:2610.10468（cs.MA/cs.AI，2026-10-07）。Asaria、Gandhi、Salomone。4页。

为什么重要： 科研智能体部署走向"上千个智能体共享一个算力池"，而多数系统一次只组织一个项目、或放任种群自发演化。核心论点：不管设计者给不给，种群都会形成组织——所以应显式地给。制度：各"首席研究员实验室"以提案征集（RFP）+ 独立评审 + 拨款竞争算力；人类总督（"市长"）只分配资源、不指派任务。运行实例：一万研究员组成的社会被只要求改进语言模型预训练，一个实验室报告用约少30%算力达到同等质量——作者自己注明：测试该结果的实验室间尚未取得一致。

LocalKin 适用性： 与我们蜂群治理问题最接近的镜子。可迁移：(1) "提案+独立评审+拨款"作算力分配机制；(2) 资源分配与任务指派分权（"市长只分配、不指派"），对应 conductor/审查分权设计；(3) 对未核实结果显式标注存疑，与基线协议诚实汇报要求同构。成本：纯流程设计；可借 swarm_comm 试点评审回路。

arXiv:2610.07334（cs.LG/cs.CR，2026-10-05）。Kabra、Venhoff、Schroeder de Witt。NeurIPS 2026 NeuralArtifacts workshop Spotlight（据 arXiv comments）；14页。

为什么重要： 可解释性多为"反应式"（在已知输入上探测激活）。本文训练微调 LM 只读原始权重诊断网络，无需行为测试。Phase I：分阶段课程 + 外部计算链（无参数操作交给确定性代码）使解释器 LLM 能从权重模拟小型 transformer 前向传播，未见目标上 99% held-out 精度。Phase II：仅用良性病理训练，对训练中未见的后门零样本检测——attention 路由型 AUROC 0.93；多样化威胁分布（含隐匿型与对抗正则变体）0.81。手工统计检测器遇威胁模型漂移即崩塌，预言机保持均匀能力。作者明言的主要开放难题：扩展到现实模型规模。

LocalKin 适用性： "只读权重"审计器 = 第三方模型用作智能体大脑前的廉价预部署筛查，不需探针数据；目前仅在小 transformer 上演示。附带设计模式：无参数操作外包给确定性代码（正是量化角色"调真实计算器不做 token 心算"的既有模式）。成本：高（无权重访问管线）——观察项。

arXiv:2610.03858（cs.LG/cs.AI，2026-10-02）。Schlegel 等9人。

为什么重要： 从对偶形式（Irie 等，arXiv:2202.05798：梯度训练的线性层=对训练数据的线性注意力）出发：不把数据压缩进固定权重，而是每数据点存一对 key-value、推理时经注意力检索重组——会生长的网络。关键贡献：线性注意力规则搬到高级核上不成立；为 RBF/softmax 类核化注意力推导泛函梯度规则。图像分类与师生任务上表现有前景；对偶形式中换入 MesaNet/DeltaNet 后与固定尺寸网络优化器形式对应。

LocalKin 适用性： 为"按条目保存经验、推理时检索"提供有原理的学习规则论证（我们智能体已保留会话记忆）。最可迁移的是负结果：简单检索方案的学习规则照搬到更复杂检索机制会无声失败——与把提示词更新启发式迁到结构不同的记忆库是同一失败模式。成本：中高；研究级。

arXiv:2610.12247（cs.LG/cs.AI，2026-10-08）。Leko、Benić、Bernárdez、Miolane、Fink、Telyatnikov。

为什么重要： 全域拓扑训练要先物化完整提升结构；Reddit（233k节点/5730万边）上训练常不可行。Cluster-TNN 改为局部提升：预处理分区；运行时采样节点簇、重建诱导子图为小批量（保留被采样节点间全部边，维持跨簇建高阶结构所需连通性）、再在小批量内提升。21组匹配对比：每组峰值显存均下降、平均降 83.2%、精度保持有竞争力；作者称实现多种高阶 TNN 在 Reddit 与 OGBN Products 上的首次训练。

LocalKin 适用性： 直接用例窄（不训练 TNN），模式可广泛迁移：全局表示无法物化时，就在小批量内局部重建，保留下游所需连通性（"保留被采样节点间全部边"= 未来建智能体交互图的正确小批量构建规则）。成本：记下零成本；落地中等。

arXiv:2610.01894（cs.LG/eess.IV/q-bio.NC，2026-10-01）。Renauld、Poulin、Larochelle、Théberge、Descoteaux。

为什么重要： dMRI 纤维束成像序列模型受控基准（固有矛盾：局部扩散信息 vs 全局解剖合理性）。引入生成-验证阶段，在局部损失与全局流线质量错配下实现流线级监督；系统研究训练策略、输入表示（CNN嵌入、EOS token）与超参数——ISMRM2015 挑战赛数据集上作者所称"迄今最高报告性能"；受控实验量化缺失束、噪声/不完美训练流线、无效纤维的影响；Tractoinferno 活体演示；附给未来研究者的建议清单。

LocalKin 适用性： 领域远、方法近。可迁移核心："在不完美标签上训练"实验设计——量化每种训练数据缺陷对下游质量的伤害——正是任何学习组件的鲁棒性审计模板（记忆库损坏 vs 智能体校准）。"局部损失 vs 全局质量"也是我们评审回路的日常张力；generation-validation 技巧（在"中间但全局"层级监督）值得借鉴。成本：方法论采纳零成本。

• ●CNet: A Complex-Valued Deep Learning Framework with Wirtinger Autodifferentiation and FFT--Hadamard Convolution — arXiv:2610.08592（v1 2026-10-06、v2 2026-10-07；单作者）。C++/CUDA 复数值框架；Born-rule 注意力在 tiny-shakespeare 上 1.52 nats/char、与 softmax 注意力持平（作者声明）。细分领域；摘要页代码链接未抓取/验证。
• ●Sharp and Adaptive Cluster Recovery in Slowly Mixing Gaussian Hidden Markov Models — arXiv:2610.09123（math.ST，2026-10-06）。难度集中于稀疏转移点；恢复由 r_n² 支配；r_n²≫δ 时几乎完全恢复才可能；精确恢复阈值 r_n² = 2 log(nδ)(1+o(1))。纯理论；为实验设计词汇表记下（在线/离线差距；延迟标签反馈的有效记忆视界）。

1. ●仅摘要级来源，未读全文。2. 未抓取/验证任何代码仓库。3. 关键词式搜索覆盖有限——"入选5篇"= "5篇已核验候选"，不是"窗口最佳5篇"。4. 所有数字均为作者自述：99% held-out；AUROC 0.93/0.81；约30%算力节省（实验室意见不一）；1.52 nats/char；平均83.2%显存降幅；ISMRM2015"迄今最高报告性能"。