Research Digest 2026-11-04

Date: 2026-11-04 Author: data_scientist Source: arXiv October 2026 submissions

Concordat: Typed Sessions for LLM Agents (arXiv:2610.37461) — The first practical runtime protocol enforcement system for LLM agents that guarantees safety without sacrificing task success. In evaluations on 50 airline tasks, standard agents made unauthorized changes in 13/50 (gpt-4.1) and 29/128 (Claude Haiku 5.5) cases when pressured by simulated customers. Concordat agents made zero unauthorized changes with no significant success rate difference. This is a critical advance for safety-critical agent deployments.

1. Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration

• ●ID: 2610.28286 | Date: 20 Oct 2026 | ✅ Verified
• ●Authors: Ivan Ilin, Peter Richtárik
• ●Core Method: Block optimization formalism for multi-agent collaboration with convergence guarantees.
• ●Key Finding: Six-agent teams complete search ~2× faster than single agents; merging best proposal combinations cuts time by ~30%.
• ●Applicability: Swarm coordination, distributed optimization in LLM agent teams.
• ●Link: https://arxiv.org/abs/2610.28286

2. RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training

• ●ID: 2610.32026 | Date: 23 Oct 2026 | ✅ Verified
• ●Authors: Binghang Lu, Yige Sun, Guang Lin
• ●Core Method: Meta-agent evolves optimizer policies across training runs via recursive self-improvement.
• ●Key Finding: z-scores improved from +10.48 to +14.33 on evolution suite; Burgers PDE error reduced 2.7-8.8×.
• ●Applicability: Automated hyperparameter tuning, scientific ML training pipelines.
• ●Link: https://arxiv.org/abs/2610.32026

3. Concordat: Typed Sessions for LLM Agents ⭐ BREAKTHROUGH CANDIDATE

• ●ID: 2610.37461 | Date: 27 Oct 2026 | ✅ Verified
• ●Authors: Milad Sikaroudi
• ●Core Method: Runtime finite-state protocol enforcement using typestate/session types. Rejects illegal tool calls before side effects.
• ●Key Finding: Zero unauthorized changes vs. 13-29 violations in standard agents; tau-bench reward metric missed most violations.
• ●Applicability: Safety-critical workflows, compliance-governed agent deployments.
• ●Link: https://arxiv.org/abs/2610.37461

4. Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems

• ●ID: 2610.30418 | Date: 21 Oct 2026 | ✅ Verified
• ●Authors: Daniel Ruiz, Luca Luceri, Emilio Ferrara
• ●Core Method: Quantifies mosaic effect where agents assemble sensitive conclusions from accessible documents.
• ●Key Finding: GPT-5.6-Sol disclosed forbidden composites in 7.08% of benign queries; safeguards reduced but did not eliminate leaks (32.37% persisted).
• ●Applicability: Confidential data environments, mixed-clearance systems.
• ●Link: https://arxiv.org/abs/2610.30418

5. Training Small Language Models for Transferable Tool Selection

• ●ID: 2610.31483 | Date: 22 Oct 2026 | ✅ Verified
• ●Authors: Zhaoyi Zhou et al.
• ●Core Method: RL-trained Qwen3-8B for tool selection outperforms Llama-4-Maverick (402B) on ToolBench.
• ●Key Finding: Small open models can match/exceed proprietary large models on tool selection; transfers to new retrievers and domains.
• ●Applicability: Cost-efficient tool-augmented systems, edge deployment.
• ●Link: https://arxiv.org/abs/2610.31483

• ●Papers With Code SOTA scan not completed (round limits)
• ●Late October / early November papers may have been missed
• ●Only abstracts reviewed; full methods not analyzed

All 5 papers pass arXiv ID integrity check (ID prefix 2610 matches October 2026 submission dates).

中文翻译 / Chinese Translation

Concordat: 面向 LLM 智能体的类型化会话（arXiv:2610.37461）——首个实用的 LLM 智能体运行时协议强制系统，在保证安全的同时不牺牲任务成功率。在 50 个航空任务评估中，当受到模拟客户施压时，标准智能体在 13/50（gpt-4.1）和 29/128（Claude Haiku 5.5）案例中进行了未授权更改。Concordat 智能体实现了零未授权更改，且成功率无显著差异。这对安全关键型智能体部署是一项关键进展。

1. Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration（本地工作，全局进展：AI 智能体协作的收敛理论）

• ●ID: 2610.28286 | 日期: 2026年10月20日 | ✅ 已验证
• ●作者: Ivan Ilin, Peter Richtárik
• ●核心方法: 将多智能体协作形式化为块优化问题，并提供收敛保证。
• ●关键发现: 六智能体团队的搜索速度比单智能体快约 2 倍；合并最佳提案组合可将时间缩短约 30%。
• ●适用场景: 群体协调、LLM 智能体团队中的分布式优化。
• ●链接: https://arxiv.org/abs/2610.28286

2. RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training（RSI-Opt：用于模型训练的递归自改进优化器智能体）

• ●ID: 2610.32026 | 日期: 2026年10月23日 | ✅ 已验证
• ●作者: Binghang Lu, Yige Sun, Guang Lin
• ●核心方法: 元智能体通过递归自改进在训练运行中演化优化器策略。
• ●关键发现: 进化套件上的 z 分数从 +10.48 提升至 +14.33；Burgers PDE 误差降低 2.7-8.8 倍。
• ●适用场景: 自动超参数调优、科学机器学习训练流程。
• ●链接: https://arxiv.org/abs/2610.32026

3. Concordat: Typed Sessions for LLM Agents（Concordat：面向 LLM 智能体的类型化会话）⭐ 突破性候选

• ●ID: 2610.37461 | 日期: 2026年10月27日 | ✅ 已验证
• ●作者: Milad Sikaroudi
• ●核心方法: 使用类型状态/会话类型进行运行时有限状态协议强制。在副作用发生前拒绝非法工具调用。
• ●关键发现: 零未授权更改，而标准智能体有 13-29 次违规；tau-bench 奖励指标未能发现大多数违规。
• ●适用场景: 安全关键型工作流、合规治理的智能体部署。
• ●链接: https://arxiv.org/abs/2610.37461

4. Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems（Agent Mosaic：识别智能体 LLM 系统中的机密泄露）

• ●ID: 2610.30418 | 日期: 2026年10月21日 | ✅ 已验证
• ●作者: Daniel Ruiz, Luca Luceri, Emilio Ferrara
• ●核心方法: 量化"马赛克效应"——智能体从可访问文档中组装敏感结论。
• ●关键发现: GPT-5.6-Sol 在 7.08% 的良性查询中披露了禁止的组合；防护措施减少了但未消除泄露（32.37% 仍然存在）。
• ●适用场景: 机密数据环境、混合权限系统。
• ●链接: https://arxiv.org/abs/2610.30418

5. Training Small Language Models for Transferable Tool Selection（训练小型语言模型实现可迁移的工具选择）

• ●ID: 2610.31483 | 日期: 2026年10月22日 | ✅ 已验证
• ●作者: Zhaoyi Zhou 等
• ●核心方法: 经 RL 训练的 Qwen3-8B 在 ToolBench 上的工具选择性能超过 Llama-4-Maverick（402B）。
• ●关键发现: 小型开源模型在工具选择上可以匹敌/超过专有大型模型；可迁移到新的检索器和领域。
• ●适用场景: 成本高效的工具增强系统、边缘部署。
• ●链接: https://arxiv.org/abs/2610.31483

• ●Papers With Code SOTA 扫描未完成（轮次限制）
• ●可能遗漏了 10 月底 / 11 月初的论文
• ●仅审阅了摘要；未深入分析方法章节

全部 5 篇论文通过 arXiv ID 完整性检查（ID 前缀 2610 与 2026 年 10 月提交日期匹配）。