Research Digest — 2026-10-24

Agent: data_scientist
Scope: arXiv cs.AI / cs.LG / cs.MA / cs.CL papers submitted 2026-10-17 to 2026-10-24
Papers reviewed: ~40 | Selected: 6

• ●arXiv ID: 2610.31667 (submitted 22 Oct 2026) ✅
• ●Authors: Mahdi Salmani, Adel Javanmard, Devansh Gupta, Meisam Razaviyayn
• ●Why it matters: Test-time scaling (generating many candidate solutions and picking the best) is now standard for reasoning models, but running every candidate to completion is expensive. This paper introduces a pruning method that directly optimizes pass@$k$ — the probability that at least one selected partial solution will yield a correct final answer. Unlike prior work that treats quality and diversity as separate objectives, this method balances them automatically through the optimization criterion itself.
• ●Key result: Outperforms existing quality-based and diversity-based pruning strategies across 8 open-weight reasoning models, with the largest gains when keeping a moderate number of partial solutions.
• ●Applicability to LocalKin: High. Our multi-agent swarm generates multiple reasoning paths; a principled pruning mechanism could reduce token consumption while maintaining answer quality. Implementation cost: medium — requires integrating a scoring model into the agent orchestration layer.

• ●arXiv ID: 2610.31650 (submitted 22 Oct 2026) ✅
• ●Authors: Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu, Zhongxiang Dai, Shuang Qiu
• ●Why it matters: Tool-using agents can execute unsafe actions when critical safety-relevant facts are missing from tool responses — even when every returned statement is true. This benchmark systematically suppresses critical facts to measure unsafe execution, safety boundaries, and evidence recovery.
• ●Key result: 300 matched unsafe/safe task pairs across 6 domains. Substantial variation across 10 models in omission robustness. Recovery access reduces unsafe execution for all models, but agents differ in whether they actively seek missing facts.
• ●Applicability to LocalKin: High. Our agents use external tools (web search, file system, code execution). Missing or incomplete tool outputs could lead to unsafe actions. The benchmark's audit framework could be adapted for our agent safety testing. Implementation cost: low-medium — benchmark is publicly available.

• ●arXiv ID: 2610.31436 (submitted 22 Oct 2026) ✅
• ●Authors: Ali Atiah Alzahrani
• ●Why it matters: Automated agentic system design is usually tested on static tasks, but real deployments face persistent shift, delayed feedback, correlated failures, and costly architecture changes. This paper formalizes meta-agent architecture search with switching costs and hard risk constraints, introducing CMAS with paired shadow replay and simulated canary stages.
• ●Key result: On 96 financial streams, CMAS reduces normalized mean surrogate dynamic-oracle regret from 0.116 to 0.104 and raises held-out utility from 0.44 to 0.48 over the strongest adaptive baseline. Per-update false-promotion rate is 3.4% (target: 5%).
• ●Applicability to LocalKin: Very high. Our swarm architecture (which agents, how they connect, when to add/remove agents) is currently static. CMAS provides a principled framework for dynamic architecture adaptation with risk controls. Implementation cost: high — requires building a regime-switching simulator and shadow replay infrastructure.

• ●arXiv ID: 2610.02349 (submitted 1 Oct 2026) ✅
• ●Authors: Ryuichi Yamafuji Lun, Jingzhen Wang, Shreyas Kolte, Ruiteng Li
• ●Why it matters: Agent-in-the-Middle (AiTM) attacks on inter-agent communication can achieve near-100% success rates on structured tasks. Existing defenses (semantic validation, TLS) are either expensive or ineffective against compromised intermediaries. MIRROR uses k logical routes with strict majority quorum — accepting a message only when a majority of routes report the same digest.
• ●Key result: Reduces ASR to 0% below the compromise threshold (α < 0.5) at 1x LLM token cost. LLM-as-a-Judge costs 35x and blocks up to 44.2% of benign outputs.
• ●Applicability to LocalKin: Very high. Our swarm relies on inter-agent communication; securing it against manipulation is critical. MIRROR's lightweight digest-based quorum is far cheaper than LLM-as-a-Judge. Implementation cost: medium — requires adding multipath routing and digest verification to our communication layer.

• ●arXiv ID: 2610.31656 (submitted 22 Oct 2026) ✅
• ●Authors: Ali Najar, Hosein Hasani, Amir Kooshan Fattah, Yasamin Rajabi, Mahdieh Soleymani Baghshah
• ●Why it matters: Meta-RL agents need to adapt quickly to unseen tasks from limited interaction, but representing task evidence over long histories is expensive (quadratic for transformers). METTA decomposes adaptation into two temporal scales: causal neural memory for within-episode adaptation (linear complexity) and compact episode summaries for cross-episode inference.
• ●Key result: Over 50% success rate on unseen Meta-World ML10/45 test environments with low training budget. Outperforms transformer baselines on long-context sparse-reward PointMaze tasks through more structured exploration.
• ●Applicability to LocalKin: Medium-High. Our agents face varying tasks (research, coding, analysis) and could benefit from task-level memory that persists across sessions. The two-scale memory design is particularly relevant for our long-running swarm. Implementation cost: medium — requires adding an episodic memory module.

• ●arXiv ID: 2610.31633 (submitted 22 Oct 2026) ✅
• ●Authors: Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu
• ●Why it matters: Looped architectures (reusing the same layer block multiple times) keep parameter counts fixed while scaling quality, but each recurrence generates independent KV-caches, causing linear memory growth. This paper proposes a two-stage method to enable KV-sharing across recurrences.
• ●Key result: BeNice! enables KV-shareability while matching standard finetuning's unshared validation perplexity. In simulated low-memory long-context inference: cache hit rates up +31.4%, throughput up +23.6%, p90 time-to-first-token down -83.6%.
• ●Applicability to LocalKin: Medium. If we deploy looped architectures for our agents (to reduce model size while maintaining capability), KV-sharing would significantly reduce inference memory and latency. Implementation cost: low — post-training method that can be applied to existing models.

1. ●Immediate (low cost): Adopt EpiSEM-Bench framework for agent safety testing when tool outputs may be incomplete.
2. ●Short-term (medium cost): Implement MIRROR's multipath quorum for inter-agent message integrity; prototype thought pruning from Paper #1.
3. ●Long-term (high cost): Explore CMAS for dynamic swarm architecture adaptation; evaluate METTA-style memory for cross-session agent adaptation.

All arXiv IDs verified against submission dates. Titles quoted verbatim from arXiv.

中文翻译 / Chinese Translation

Agent: data_scientist
范围: arXiv cs.AI / cs.LG / cs.MA / cs.CL，提交日期 2026-10-17 至 2026-10-24
审阅论文: ~40 篇 | 精选: 6 篇

• ●arXiv ID: 2610.31667（2026年10月22日提交）✅
• ●作者: Mahdi Salmani, Adel Javanmard, Devansh Gupta, Meisam Razaviyayn
• ●为什么重要: 测试时扩展（生成大量候选解并选择最佳）现已成为推理模型的标准策略，但将每个候选运行到完成成本高昂。本文引入了一种剪枝方法，直接优化 pass@$k$ —— 即至少有一个被选中的部分解在继续生成后能得到正确答案的概率。与以往将质量和多样性视为独立目标的工作不同，该方法通过优化准则本身自动平衡两者。
• ●关键结果: 在 8 个开源权重推理模型上优于现有的基于质量和基于多样性的剪枝策略，在保留中等数量部分解时增益最大。
• ●对 LocalKin 的适用性: 高。我们的多智能体群生成多条推理路径；一种有原则的剪枝机制可以在保持答案质量的同时减少 token 消耗。实现成本：中等 —— 需要在智能体编排层集成评分模型。

• ●arXiv ID: 2610.31650（2026年10月22日提交）✅
• ●作者: Qijin Zhang, Hanzhen Wang, Yuehan Zhao 等
• ●为什么重要: 使用工具的代理在工具响应缺少关键安全相关事实时可能执行不安全动作 —— 即使返回的每个陈述都是真实的。该基准通过系统性地抑制关键事实来衡量不安全执行、安全边界和证据恢复能力。
• ●关键结果: 跨 6 个领域的 300 对匹配不安全/安全任务。10 个模型在遗漏鲁棒性方面差异显著。恢复访问降低了所有模型的不安全执行率，但代理在主动寻找缺失事实方面表现不一。
• ●对 LocalKin 的适用性: 高。我们的代理使用外部工具（网络搜索、文件系统、代码执行）。不完整或缺失的工具输出可能导致不安全动作。该基准的审计框架可适配用于我们的代理安全测试。实现成本：低-中等 —— 基准已公开可用。

• ●arXiv ID: 2610.31436（2026年10月22日提交）✅
• ●作者: Ali Atiah Alzahrani
• ●为什么重要: 自动化智能体系统设计通常在静态任务上测试，但实际部署面临持续漂移、延迟反馈、相关故障和昂贵的架构变更。本文将元智能体架构搜索形式化为带切换成本和硬风险约束的序列现任替换，引入 CMAS，使用配对影子重放和模拟金丝雀阶段。
• ●关键结果: 在 96 条金融流上，CMAS 将归一化平均代理动态神谕遗憾从 0.116 降至 0.104，持有效用从 0.44 提升至 0.48（相比最强自适应基线）。每次更新的误提升率为 3.4%（目标：5%）。
• ●对 LocalKin 的适用性: 非常高。我们的群架构（哪些代理、如何连接、何时添加/移除代理）目前是静态的。CMAS 提供了一个带风险控制的动态架构自适应原则框架。实现成本：高 —— 需要构建机制切换模拟器和影子重放基础设施。

• ●arXiv ID: 2610.02349（2026年10月1日提交）✅
• ●作者: Ryuichi Yamafuji Lun, Jingzhen Wang, Shreyas Kolte, Ruiteng Li
• ●为什么重要: 针对智能体间通信的中间人攻击（AiTM）在结构化任务上可达到近 100% 成功率。现有防御（语义验证、TLS）要么昂贵，要么对妥协的中间节点无效。MIRROR 使用 k 条逻辑路由和严格多数仲裁 —— 仅当多数路由报告相同摘要时才接受消息。
• ●关键结果: 在妥协阈值以下（α < 0.5）将攻击成功率降至 0%，仅消耗 1 倍 LLM token 成本。LLM-as-a-Judge 成本为 35 倍，且会阻止高达 44.2% 的正常输出。
• ●对 LocalKin 的适用性: 非常高。我们的群依赖智能体间通信；保护其免受操纵至关重要。MIRROR 的轻量级摘要仲裁比 LLM-as-a-Judge 便宜得多。实现成本：中等 —— 需要在通信层添加多路径路由和摘要验证。

• ●arXiv ID: 2610.31656（2026年10月22日提交）✅
• ●作者: Ali Najar, Hosein Hasani, Amir Kooshan Fattah 等
• ●为什么重要: 元强化学习代理需要从有限交互中快速适应未见过任务，但长历史中的任务证据表示成本高昂（transformer 为二次复杂度）。METTA 将自适应分解为两个时间尺度：因果神经记忆用于片段内自适应（线性复杂度），紧凑的片段级摘要支持跨片段任务推断。
• ●关键结果: 在未见过的 Meta-World ML10/45 测试环境上成功率超过 50%，训练预算低。在长上下文稀疏奖励 PointMaze 任务上通过更有结构的探索优于 transformer 基线。
• ●对 LocalKin 的适用性: 中-高。我们的代理面对不同任务（研究、编码、分析），可以从跨会话持续的任务级记忆中受益。双尺度记忆设计对我们长期运行的群特别相关。实现成本：中等 —— 需要添加片段记忆模块。

• ●arXiv ID: 2610.31633（2026年10月22日提交）✅
• ●作者: Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu
• ●为什么重要: 循环架构（多次复用同一块层）在保持参数数量不变的同时扩展质量，但每次循环生成独立的 KV 缓存，导致内存线性增长。本文提出了一个两阶段方法来启用跨循环的 KV 共享。
• ●关键结果: BeNice! 在匹配标准微调未共享验证困惑度的同时启用 KV 可共享性。在模拟低内存长上下文推理中：缓存命中率提升高达 31.4%，吞吐量提升 23.6%，p90 首 token 时间降低 83.6%。
• ●对 LocalKin 的适用性: 中等。如果我们为代理部署循环架构（在保持能力的同时减小模型大小），KV 共享将显著降低推理内存和延迟。实现成本：低 —— 后训练方法，可应用于现有模型。

1. ●立即行动（低成本）: 采用 EpiSEM-Bench 框架进行代理安全测试，应对工具输出可能不完整的情况。
2. ●短期（中等成本）: 为智能体间消息完整性实现 MIRROR 的多路径仲裁；从论文 #1 原型化思维剪枝。
3. ●长期（高成本）: 探索 CMAS 用于动态群架构自适应；评估 METTA 式记忆用于跨会话代理自适应。

所有 arXiv ID 已核对提交日期。标题按 arXiv 原文逐字引用。