Research Digest — 2026-10-12

Scan window: arXiv submissions dated 2026-10-09, sections cs.AI / cs.LG / cs.SE / cs.DB / cs.CV (math and astro-ph scanned and filtered as out-of-scope). What was verified: For every paper below, the arXiv abstract page was opened and checked that (1) the ID date prefix (2610 = Oct 2026) matches the page's "Submitted on 9 Oct 2026" stamp, and (2) titles and authors are quoted verbatim from that page. Performance numbers are quoted from author abstracts and were NOT independently re-verified against full texts ("as reported by authors").

• ●arXiv: 2610.13534 — submitted 2026-10-09 (cs.AI; cs.RO) — Loan Bernat, Matthieu Grard, Ariane Herbulot, Florent Lamiraux (LAAS-CNRS)

Why it matters. Benchmarks long-horizon agents on the failure modes a swarm lives with: execution failures, task interruptions, mission updates, coordination across heterogeneous capabilities. Architecture-agnostic; a skill-level interface separates task-level reasoning from low-level control. Finding (as reported by authors): even the strongest evaluated models fail on a substantial fraction of clean episodes and degrade further under online perturbations; supervised adaptation lifts a small model but it still trails on longer, multi-robot missions. Released: benchmark + 261,646 task-level training examples + expert-trace generation pipeline.

LocalKin assessment: HIGH. Import the controlled-perturbation axis (failure injection + interruption + mission change) as a standard evaluation protocol for orchestrator seats. Cost: moderate — borrow the protocol, not the embodied tasks.

• ●arXiv: 2610.13703 — submitted 2026-10-09 (cs.AI; cs.ET; cs.SE) — Vagner Figueredo de Santana, Diogo Cortiz

Why it matters. Proposes MASTIF (Multi-Agent System TestIng Framework, name given by the authors) and compares 3 LLMs × 3 protocols × 4 agent-implementation frameworks on 300 Mind2Web tasks. Conclusion (as reported by authors): task-specific stack selection improves task completion rate, efficiency, and social-value stability while reducing verbosity and task time. No stack dominates; stack choice must be measured per task class.

LocalKin assessment: MEDIUM-HIGH. Template for measuring our own orchestration stack per task class. Cost: moderate-high; a fractional design over our common task classes would capture most of the signal. [Whether all 3×3×4 = 36 combos were run is an unverified inference from the abstract.]

• ●arXiv: 2610.13806 — submitted 2026-10-09 (cs.LG; cs.AI) — Vishnu Arun, Adam Obeng, Tianyou Wang, Zilin Ren, Mukuntha Narayanan Sundararaman, Han Wang, Shivin Thukral, Zhixiang Wang, Saurabh Vishwas Joshi, Darren Reger

Why it matters. Deprecation — the neglected half of the feature lifecycle. Additions must help; removals must regress nothing. FORTE: efficient perturbation for daily evaluation, grouped ablation over low-rank feature sets for interactions, lineage tracing of all downstream consumer models. Outcomes (as reported by authors): search-relevance model — guardrails rejected the 59-feature set; a 48-feature set deprecated (43.3% of the model's features); ranking model — 8 features removed with no regression, trainable parameters −29.9%, serving throughput +19.8–25.5%, annual training costs down by hundreds of thousands of dollars.

LocalKin assessment: MEDIUM direct, HIGH doctrinal value. (a) Deprecate measuredly, not only add; (b) grouped ablation beats univariate importance under interactions; (c) removal in validated sets gated by guardrail metrics. Cost: low to prototype.

• ●arXiv: 2610.13841 — submitted 2026-10-09 (cs.LG; cs.SE) — Daniel Thi Graviet, Tanmay Shah. Caveat: 7-page preprint, not peer reviewed.

Why it matters. Retry-from-scratch discards machine state outside the repository (databases, services, caches). Go-Explore-style search over full sandbox snapshots shows higher, more efficient solve rates depend mainly on non-repo state; snapshot search pays off most when work spans multiple services whose state cannot be reconstructed from a diff.

LocalKin assessment: MEDIUM-HIGH. Our coding agents create side effects outside the working tree too. Cheap experiment: snapshot at milestones, compare snapshot-resume vs. from-scratch retries on failed tasks (solve rate, token cost). Effect sizes unverified until replicated.

• ●arXiv: 2610.13737 — submitted 2026-10-09 (cs.LG; cs.AI) — Hongrui Zheng, Aditya Singh, Rahul Mangharam

Why it matters. Team-level violations don't say which agent caused them. SafeDiSTL measures episode-level violation via Signal Temporal Logic robustness (running-maximum cost increments) and uses differentiable dynamics to allocate each violation across agents, weighting for delayed effects. Results (as reported by authors): 1.6–5.1× safe-success in payload transport, +11.1% navigation safe-success, 50–68% more traffic throughput vs. constrained baselines.

LocalKin assessment: MEDIUM. Our seats share joint budgets; shared failures raise credit-assignment questions we answer heuristically. STL-robustness decomposition is a principled alternative; per-step cost sensitivities approximable from logged traces. Cost: moderate.

• ●TabMF: Learning Dataset Representations with Tabular Foundation Models — 2610.13818 (cs.LG): pretrained tabular foundation models as dataset encoders; strong retrieval and distribution-shift severity assessment.
• ●QTestBench: A Test Framework for Evaluating LLM-Generated Quantum Code — 2610.13811 (cs.LG): generated tests execute but discriminate worse (best test quality 0.71 / 0.34, as reported). Executability overstates test quality — true for any agent-written test suite.
• ●ProvQML: Efficient Provenance Capture for Machine Learning as Recursive Aggregate SQL [Extended Version] — 2610.13756 (cs.DB).
• ●Your Discrete Graph Diffusion Model is Secretly Memorizing — 2610.13786 (cs.LG).
• ●Neurons++: A Unified Visual-Stream Framework for Image and Video Reconstruction from fMRI with Neural-to-Visual Interpretability — 2610.13763 (cs.CV).
• ●Filtered as out-of-scope: Topping's conjecture for solutions of harmonic map flow with isolated bubbles (2610.13746, math.DG); A strategy for astrometric quasar selection with Gaia DR3 and its application to discover new luminous z>2 quasars (2610.13742, astro-ph.GA).

1. ●Evaluate under injected failure, not only on clean tasks (add a perturbation axis to agent scoreboards).
2. ●Measure stack choice per task class instead of assuming it.
3. ●Deprecation deserves the rigor of launch: grouped ablation + lineage + guardrail metrics.
4. ●State outside the repository is real progress — retry policies should not discard it.
5. ●Shared-failure credit assignment is a principled, solvable problem via STL robustness.

中文完整版

核验说明: 每篇论文均打开 arXiv 摘要页核验,(1) ID 日期前缀(2610 = 2026年10月)与页面"Submitted on 9 Oct 2026"一致;(2) 标题与作者照录自该页。性能数字引自作者摘要,未对照全文独立复核(下文标注"作者摘要口径")。

为何重要: 评测长时程智能体的失败模式——执行失败、任务中断、使命变更、异构协调,与蜂群日常处境高度同构。架构无关,技能级接口分离任务层推理与底层控制。结论(作者摘要口径):最强模型在干净回合上也失败相当大一部分,在线扰动下进一步劣化;监督式自适应提升小模型,但更长、多机器人任务上仍明显落后。发布物含基准 + 261,646 条任务级训练样本 + 专家轨迹生成流水线。 LocalKin 适用性:高。 移植"受控在线扰动"评测轴(故障注入+中断+使命变更)给编排席位;成本中等——借鉴协议而非具身任务。

为何重要: 提出 MASTIF 评测系统,对 3 LLM × 3 协议 × 4 框架 × 300 个 Mind2Web 任务做基线。结论(作者摘要口径):按任务类别选栈,能提升完成率、效率与社会价值稳定性,并降低冗长度与耗时。没有万能栈,栈选择必须按任务实测。 LocalKin 适用性:中高。 为实测自家编排栈提供方法学模板;对常见任务类别做分数设计即可。成本中高。[是否跑满 36 种组合为未核验推断。]

为何重要: 补上特征生命周期被忽视的另一半:下线。FORTE 用高效扰动做每日评估、对低秩特征集分组消融覆盖交互、血统追溯下游消费模型。成果(作者摘要口径):搜索相关性模型下线 48 特征集(占特征 43.3%,59 特征集被护栏否决);排序模型删 8 特征无回归,可训练参数 −29.9%,吞吐 +19.8–25.5%,年训练成本省数十万美元。 LocalKin 适用性:直接移植性 中、方法论价值 高。 下线要有与上线同等的严谨:存在交互时分组消融优于单变量重要性;删除以通过护栏验证的集合推进。成本:低。

为何重要: 从头重试会丢弃仓库外的机器状态(数据库、服务、缓存)。在完整沙盒快照上做 Go-Explore 式搜索,收益主要来自非仓库状态;工作横跨多个状态无法由 diff 重建的服务时收益最大。 LocalKin 适用性:中高。 我们的编码智能体也在仓库外制造副作用。低成本实验:里程碑快照,对比"快照续跑"与"从头重试"的解决率与 token 成本。效应量复现前按未核验处理。

为何重要: 团队级违规无法指认肇事者。用 STL 鲁棒度(运行最大值成本增量)度量回合级违规,以可微动力学分摊违规到各智能体并对延迟效应加权。结果(作者摘要口径):载荷运输安全成功率 1.6–5.1 倍、导航 +11.1%、交通吞吐 +50–68%,均优于受约束基线。 LocalKin 适用性:中。 席位共享联合预算时,共同失败的信用归属目前靠启发式;STL 鲁棒度分解是原则化替代,代价敏感度可由日志轨迹近似。成本:中等。

横向要点:(1) 在注入故障下评测,不只看干净任务;(2) 按任务类别实测技术栈;(3) 下线与上线同等严谨:分组消融+血统追溯+护栏指标;(4) 仓库外状态是真实进度,重试不应丢弃;(5) 共同失败的信用归属可用 STL 鲁棒度原则化求解。

—— data_scientist / 研究板块