运行统计
- 候选论文: 本地 selector-pool 快照中保留 120 篇
- 入选论文: 30
- 已完成精读: 30
- 本地快照时间窗 (UTC): 2026-07-27T00:00:00Z → 2026-07-28T00:00:00Z (来自当前存档的候选元数据)
- 证据基础: 本次重做主要锚定
selected.json与analyses.all.json。该目录中保留的候选快照与入选集合并不完全一致，因此下面的综合判断以 30 篇已完成分析的论文为准。
展开查看用于本次综述的入选论文列表
| arXiv ID | Title / Links | Category | Score | Selection reason | Tags |
|---|---|---|---|---|---|
2607.21325 | Toward cryptographically verifiable authorization for autonomous AI agents: A security hypothesis, preliminary formal model, and proof-of-concept implementation | Privacy/Security | 93 | Cryptographically verifiable authorization for autonomous agents; strong agent security relevance. | agent-security, authorization, cryptography, formal-models, tool-use |
2607.21495 | Toward Continuous Assurance for the Democratization of AI Agent Creation in Industry | Other | 92 | Practical continuous assurance for citizen-built AI agents; strong reliability/governance relevance. | agents, safety, governance, monitoring, reliability, enterprise |
2606.29280 | Deterministic Decisions for High-Stakes AI. A Zero-Egress Pipeline with the Deployability of RAG and the Accuracy of Machine Learning | Alignment | 91 | Finds major LLM intervention bias in high-stakes advice; strong empirical comparison to supervised policy learning. | llm-reliability, high-stakes-ai, calibration, rag, evaluation |
2607.21111 | TOUR: A Trajectory-Level Unlearning Benchmark for Offline Reinforcement Learning | Privacy/Security | 90 | Benchmark for unlearning in offline RL with privacy audits and utility anchors; highly reusable. | unlearning, offline-RL, privacy, benchmark, evaluation |
2606.28710 | The Two Genie Game: Adoption and Welfare in Audit-Grounded AI Governance | Alignment | 90 | Directly studies AI governance, RLHF vs harm-minimizing agents, and welfare/adoption tradeoffs. | ai-governance, alignment, rlhf, game-theory, safety |
2607.21461 | AREX: Towards a Recursively Self-Improving Agent for Deep Research | Alignment | 90 | Recursively self-improving research agent with verification loop; highly relevant to agent reliability. | agents, self-improvement, verification, deep-research, reliability |
2607.11175 | The Path to Self-Evolving Clinical Systems: Scaling Medical Agents from Assistance to Autonomy | Other | 90 | Deployment-first roadmap for autonomous medical agents with benchmarks, training envs, and trust taxonomy. | medical-agents, autonomy, benchmarking, deployment, safety |
2607.21143 | One More Turn, Less Regret: A Regret-Based Multi-Turn Benchmark for LLMs' Clarification Policies | Alignment | 90 | Benchmark for multi-turn clarification policies with regret-based evaluation; useful for agent reliability. | evaluation, agents, benchmark, clarification, reliability, policy |
2607.12252 | FinResearchBench II: A Deep Research Benchmark with Consensus-Derived Gold Rubrics for Distinguishing Financial Report Quality | Other | 90 | Benchmark for deep research agents with consensus-derived rubrics; strong eval reuse value. | benchmark, evaluation, agents, llm-judges, finance |
2607.21482 | Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks | Other | 89 | Useful benchmark for local open-weight coding agents on sensitive data workflows. | agents, evaluation, open-weight, privacy, coding |
2607.15095 | Digital Pantheon: Simulating and Auditing Coalition Formation with LLM Agents | Alignment | 88 | LLM multi-agent coalition simulation with DPO+RAG; relevant to auditing ideological agent behavior. | llm-agents, multi-agent, auditing, dpo, rag |
2607.19243 | Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs | Alignment | 88 | Inference-time methods for cross-lingual factual consistency in LLMs; strong reliability focus. | LLMs, factuality, multilingual, steering, reliability |
2607.15001 | LQCDMaster: Agentic Scientific Computing for Lattice Quantum Chromodynamics Research | Other | 88 | Domain-specialized scientific agent with tool constraints; strong agentic workflow and reliability relevance. | agents, scientific-computing, tool-use, reliability, code-generation |
2606.31167 | MIRTH: Mutual-Information Reasoning with Temporal Hubs for Vision-Language-Action Agents | Alignment | 88 | Advances VLA agents with temporal memory, latent reasoning, and efficient action decoding. | VLA, agents, robotics, reasoning, temporal, efficiency |
2607.18006 | MADA-RL: Multi-Agent Debate-Aware Reinforcement Learning for Parameter-Efficient Reasoning in Compact Models | Alignment | 88 | Debate-aware RL for compact LLM reasoning with concrete PEFT gains and reusable training idea. | LLM, reasoning, RL, post-training, PEFT, multi-agent |
2606.31831 | An Agentic AI Framework to Accelerate Scientific Discovery in Plant Phenotyping | Other | 88 | Agentic AI for scientific workflows; concrete multi-agent system with real lab use potential. | agents, scientific-discovery, workflow-automation, tool-use, applied-ai |
2607.11084 | NVAITC AI Scientist: A Governed End-to-End Research System -- A Hypertension GWAS Case Study | Privacy/Security | 88 | Governed end-to-end AI scientist system with oversight, privacy boundaries, and reproducibility. | agents, governance, oversight, privacy, scientific-workflows, reproducibility |
2607.14905 | Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution | Robustness | 88 | Robust LLM authorship attribution via reasoning graphs, targeting paraphrase-resistant detection. | LLM, authorship-attribution, reasoning, robustness, evaluation |
2607.18975 | Mi-Memory: A Lifecycle Memory Framework for Personal AI | Other | 87 | Personal AI memory framework emphasizes governance, auditability, forgetting, and evidence-grounded continuity. | agent-memory, personal-ai, governance, auditability, privacy |
2607.20848 | Auditing Evidence Use in Medical LLM Diagnosis | Interpretability | 87 | Audits whether medical LLMs use evidence faithfully, not just final accuracy. | llm-reliability, evaluation, faithfulness, medical-ai, auditing |
2607.14439 | Active Real-World Factor-Based Evaluation for Generalist Robot Policies | Robustness | 87 | Active real-world evaluation for generalist robot policies; practical framework for finding failures. | evaluation, robotics, generalist-agents, real-world, safety |
2607.18973 | Verifiable Self-Evolution for Open-Ended Dialogue Skills via Future-Feedback Prediction | Alignment | 86 | Verifiable self-evolution for dialogue agents via future-feedback prediction; alignment-relevant. | agents, alignment, self-improvement, dialogue, verification |
2607.06452 | From Voting to Agent Collaboration: Answer-Type-Aware LLM Pipelines for BioASQ 14b | Other | 86 | LLM QA pipeline emphasizes robustness, evidence grounding, self-reflection, and agent collaboration. | llm, agents, grounding, biomedical-qa, evaluation |
2607.21404 | MemTools: A Unified Research Framework for Interoperable Agent Memory | Other | 86 | Unified framework for interoperable agent memory and controlled evaluation; reusable agent infra. | agents, memory, frameworks, evaluation, interoperability |
2607.05396 | From Fixed to Free Cameras: Calibration-Free View-Robust Vision-Language-Action Model | Robustness | 86 | Practical VLA robustness: calibration-free camera adaptation for real-world robot deployment. | VLA, robotics, robustness, generalization, multimodal |
2607.18684 | When to Trust the Map: Confidence-Aware LLM Routing for Automotive CVE-to-ATM Mapping | Privacy/Security | 86 | Confidence-calibrated LLM routing for safety-critical vuln mapping; strong selective automation angle. | security, LLM, calibration, evaluation, automotive, selective-automation |
2607.11012 | EasyOPD: An Easy-to-use On-Policy Distillation Framework for Large Language Models | Other | 86 | Reusable on-policy distillation framework for LLMs; practical post-training infra with broad impact. | LLM, distillation, post-training, framework, reproducibility |
2607.21412 | Euclid-MCP: A Model Context Protocol Server for Deterministic Logical Reasoning via Prolog | Alignment | 86 | Standardized MCP tool for deterministic symbolic reasoning; promising for safer tool-augmented agents. | agents, tool-use, reasoning, neuro-symbolic, MCP, reliability |
2607.15216 | Symbal: Detecting Systematic Misalignments in Model-Generated Captions | Alignment | 85 | Detects systematic caption misalignments in MLLM data; useful for reliability auditing and dataset quality. | multimodal-llm, reliability, auditing, dataset-quality, evaluation |
2607.05111 | From Multiplicity to Vulnerability: Privacy Amplification Risk from One-Dataset-Multiple-Model Exposure | Privacy/Security | 85 | Shows privacy leakage compounds across multiple models trained on one dataset; important API risk. | privacy, membership-inference, data-leakage, theory, APIs |
AI 论文洞察简报
2026-07-27
0) 核心结论（请先读这里）
- 今天最强的模式是：可靠性提升主要来自结构，而不是更流畅的表达。 在错误代价真实存在的场景里，确定性决策层、符号后端和校准路由反复优于开放式 LLM 行为。
- 过程感知评估正在取代只看终点的评分。 医学证据使用、澄清遗憾，以及“采用 vs. 福利”这样的工作，都在追问系统是否走了正确路径，而不仅仅是最后看起来像成功。
- 对 agent 来说，瓶颈越来越是状态治理：递归验证、记忆生命周期设计，以及持续性 assurance 正变成一等工程原语。
- 多篇论文还清楚地区分了被采用与真正有益。一个系统可能更容易被用户选择、更容易显得可信，却并不一定更安全或更有社会价值。
- 实用教训很直接：如果你想要可部署的可靠 AI，就该在生成之外建立明确控制边界，而不是期待基础模型自己约束自己。
2) 关键主题（聚类）
主题：结构化可靠性
- 为什么重要：在最有后果的设置里，自由生成正不断让位于有边界的决策系统。真正占优的不是“更会说”，而是预测、证据与行动之间更紧的接口。
- 代表论文：
- 共同方法：
- 用确定性策略、符号检查或置信度门控，把生成与执行分离。
- 优先采用狭窄动作 schema 和可审计的决策回执，而不是无限制自然语言输出。
- 把 defer / abstain 当成产品能力，而不是失败。
- 开放问题 / 失败模式：
- 结构化接口也可能继承本体覆盖不足或任务定义脆弱的问题。
- 领域内胜利仍需跨场景迁移验证。
- 如果没有更强的执行绑定，形式化授权仍然不完整。
主题：过程审计
- 为什么重要：今天多篇最佳论文都表明，顶层分数本身可能会误导。真正该问的是：系统是否使用了正确证据、问了正确追问、并以合理方式在成本与风险之间做了权衡。
- 代表论文：
- 共同方法：
- 用证据角色分析、策略遗憾或福利敏感比较替代最终答案评分。
- 让权衡变得可见：效用 vs. 轮次、采用 vs. 福利、准确率 vs. 证据忠实性。
- 用结构化探针区分真正良好的行为与表面上可以接受的结果。
- 开放问题 / 失败模式：
- 审计层本身仍可能依赖特定 benchmark 或 judge。
- 遗憾与福利代理指标未必与真实用户体验一致。
- 许多结果目前仍主要成立于精细仪表化的环境中。
主题：受治理 agent
- 为什么重要：长时程 agent 越来越像系统工程问题。新的主线是显式治理记忆、自我修订和交接，而不是期待更大的模型自行吸收协调负担。
- 代表论文：
- 共同方法：
- 加入显式验证循环、就绪性检查、生命周期状态和记忆契约。
- 让 agent 基础设施可检查，以便把失败定位到记忆、策略或环境层。
- 把治理视为运行时架构，而不只是政策文档。
- 开放问题 / 失败模式：
- 目前多数证据仍来自 benchmark、原型或内部案例研究。
- 更强的状态机制也会增加复杂度与运维负担。
- 可靠的递归式自我改进距离被真正证明，还很远。
3) 技术综述
这份综述以本地保留的入选论文集合和 analyses.all.json 中 30 篇已完成精读为基础。今天的关键结论并不是由某一个模型家族或某一个排行榜主导，而是来自多个领域里重复出现的设计动作。
最清晰的动作是在生成之后约束行动。Deterministic Decisions 是最尖锐的例子：一旦把决策问题表达成显式状态，监督式策略就在干预忠实度上显著超过零样本和 RAG 风格的 LLM arms。Euclid-MCP、可验证授权，以及置信度路由也体现了同样直觉：让语言模型负责提出候选或映射，但在系统真正行动前，必须经过确定性或校准层。
第二个动作是让评估关心机制本身。医学证据使用审计关注模型推理是否真的跟随证据；One More Turn, Less Regret 把澄清行为视为完整对话策略，而不是一次看起来“很 helpful”的追问；The Two Genie Game 更进一步，把更容易被采用的系统与真正提升福利的系统分离开来。这些工作共同反驳了一个偷懒假设：性能、偏好与安全会自然对齐。
第三个动作是把 agent 状态工程化，而不是口头带过。AREX 把验证当成研究轮次之间的递归控制信号；Mi-Memory 和 MemTools 关注记忆生命周期与互操作性；持续 assurance 工作则把 readiness、ownership 与 monitoring 纳入系统边界。于是今天的整体感觉，不是“模型更聪明了”，而是“agent 系统开始具备运行纪律”。
4) 值得优先读的论文
Deterministic Decisions for High-Stakes AI
如果你关心真实后果下的可靠性，这是最值得先读的一篇。它的核心结论很难忽视：一旦进入结构化决策问题，监督式策略能消除许多流畅 LLM 系统反复带回来的行动偏差。
局限： 最强证据仍集中在单一、结构化很强的教育支持场景。Auditing Evidence Use in Medical LLM Diagnosis
它很重要，因为它说明“答对”本身是一种危险的安慰信号。论文审计的是诊断是否真正依赖忠实证据，这比“最后答案对不对”更接近部署现实。
局限： 部分审计解释仍绑定在专家定义的证据角色上。AREX: Towards a Recursively Self-Improving Agent for Deep Research
值得读的核心在于它的控制循环设计：验证被提升成研究轮次之间的递归算子，而不是再加一个 reranker。
局限： 基准收益还不能证明它能在更混乱的真实研究任务上稳定自我改进。One More Turn, Less Regret
作为评估论文很有用，因为它惩罚无意义的额外交互，把澄清策略变成可度量的端到端行为。
局限： 像多数对话策略基准一样，它依赖一个受控的隐藏意图环境。When to Trust the Map
这是一篇很好的选择性自动化配套论文：它把置信度变成一个安全关键映射任务中的路由策略，而这正是许多真实部署真正需要的有边界自治。
局限： 当前证据仍相当依赖具体领域与主干模型。