AI 论文日报(2026-04-05)

Published:

English version: /paper-news/2026-04-05/

运行统计

  • 候选论文: 2272
  • 入选论文: 30
  • 已精读完成: 30
  • 时间窗口 (UTC): 2026-04-03T00:00:00Z → 2026-04-04T00:00:00Z (weekend_backlog_unknown, expanded=0)
展开查看用于总结的论文列表
arXiv ID标题 / 链接分类评分入选理由标签
2604.00419G-Drift MIA: Membership Inference via Gradient-Induced Feature Drift in LLMs
PDF
cs.LG, cs.AI93White-box LLM membership inference via gradient-induced representation drift; stronger privacy auditing signal.privacy, membership-inference, LLM-security, white-box, auditing
2604.00430Secure Forgetting: A Framework for Privacy-Driven Unlearning in Large Language Model (LLM)-Based Agents
PDF
cs.MA, cs.CR93Framework for privacy-driven unlearning in LLM agents; timely for deployed agent safety/privacy.LLM-agents, unlearning, privacy, data-deletion, security, governance
2604.01161Reasoning Shift: How Context Silently Shortens LLM Reasoning
PDF
cs.LG92Finds context can silently shorten reasoning traces; robustness issue for test-time scaling modelsllm-reasoning, robustness, test-time-scaling, long-context, evaluation
2604.01147SERSEM: Selective Entropy-Weighted Scoring for Membership Inference in Code Language Models
PDF
cs.SE, cs.CR92Stronger membership inference for code LMs; practical contamination/privacy auditing via AST-weighted signalsprivacy, membership-inference, data-contamination, code-llms, security, memorization
2604.00860Policy Improvement Reinforcement Learning
PDF
cs.LG92Adds inter-iteration “did policy improve?” feedback to RLVR; targets drift/collapse in LLM post-trainingRLVR, post-training, policy-improvement, reasoning, stability, verification
2604.01014AutoMIA: Improved Baselines for Membership Inference Attack via Agentic Self-Exploration
PDF
cs.CR, cs.CV90Agentic self-exploration to auto-discover stronger MIA strategies; reusable auditing framework.privacy, membership-inference, agents, red-teaming, automation
2604.00478The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents
PDF
cs.AI90Agent anti-sycophancy gating + auditor veto loop; concrete eval on TruthfulQA adversarial dialogsllm-agents, sycophancy, guardrails, behavioral-gating, oversight, evaluation
2604.00938WARP: Guaranteed Inner-Layer Repair of NLP Transformers
PDF
cs.LG, cs.AI90Provable inner-layer transformer repair against adversarial perturbations; bridges robustness+verification.robustness, adversarial, transformers, formal-methods, model-repair, verification
2603.28101Heddle: A Distributed Orchestration System for Agentic RL Rollout
PDF
cs.LG90Trajectory-centric orchestration for agentic RL rollouts; tackles long-tail tool-call bottlenecksagentic-RL, LLM-agents, systems, tool-use, scaling, scheduling
2604.01128Paper Reconstruction Evaluation: Evaluating Presentation and Hallucination in AI-written Papers
PDF
cs.CL, cs.AI, cs.LG90New framework to measure hallucination/presentation risks in agent-written papers; timely eval methodologyevaluation, hallucinations, agent-reliability, scientific-writing, benchmarks, auditing
2604.00555Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents
PDF
cs.AI, cs.CL, cs.SE89Ontology-constrained neurosymbolic agent architecture to reduce hallucination and enforce complianceagents, governance, neurosymbolic, hallucinations, enterprise, compliance
2604.00442Execution-Verified Reinforcement Learning for Optimization Modeling
PDF
cs.AI, cs.CL89Execution-verified RL with sandboxed solver as verifier for NL→optimization code; reusable agentic training recipeagents, tool-use, RLVR, verifiable-rewards, code-generation, sandboxing, optimization
2603.28309VulnScout-C: A Lightweight Transformer for C Code Vulnerability Detection
PDF
cs.CR88Lightweight transformer + curated dataset for C vuln detection; practical secure-dev impact and benchmark value.code-security, vulnerability-detection, dataset, efficient-LLM, software-security
2603.28386COvolve: Adversarial Co-Evolution of Large-Language-Model-Generated Policies and Environments via Two-Player Zero-Sum Game
PDF
cs.AI88Adversarial co-evolution of LLM-generated envs/policies for automated curricula; strong agent generalization.agents, curriculum, adversarial-training, LLM-codegen, evaluation, continual-learning
2604.01221HippoCamp: Benchmarking Contextual Agents on Personal Computers
PDF
cs.AI, cs.CV88Realistic benchmark for contextual PC/file agents with large multimodal data + dense trajectories for analysisagent-benchmark, tool-use, context, multimodal, evaluation, personal-data
2603.29328Beyond Corner Patches: Semantics-Aware Backdoor Attack in Federated Learning
PDF
cs.CR, cs.AI, cs.CV, cs.DC, cs.LG86Realistic semantic in-distribution backdoors in federated learning; stronger threat model than patch triggers.backdoors, federated-learning, adversarial-ML, security-eval, robustness
2603.28281Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback
PDF
cs.LG86Provable robustness to ε-fraction corrupted preference data in offline multi-agent RLHF settingrlhf, offline-rl, multi-agent, robustness, data-corruption, theory, alignment
2603.28673FL-PBM: Pre-Training Backdoor Mitigation for Federated Learning
PDF
cs.LG, cs.CR, cs.DC86Client-side pre-training data filtering to mitigate FL backdoors; practical security anglebackdoors, federated-learning, data-poisoning, ml-security, defenses
2603.08269SAIL: Test-Time Scaling for In-Context Imitation Learning with VLM
PDF
cs.RO, cs.AI86Test-time compute scaling (MCTS) for VLM imitation; reusable recipe for robust robot agentsagents, robotics, VLM, test-time-scaling, MCTS, imitation-learning, planning
2603.21656TrustFed: Enabling Trustworthy Medical AI under Data Privacy Constraints
PDF
cs.LG, cs.CY86Federated uncertainty quantification with distribution-free finite-sample coverage under heterogeneityuncertainty-quantification, federated-learning, reliability, privacy, healthcare, conformal
2603.29399ELT-Bench-Verified: Benchmark Quality Issues Underestimate AI Agent Capabilities
PDF
cs.AI, cs.DB86Auditor-Corrector finds benchmark flaws; shows agent capability underestimation and improves evaluation rigoragent-evaluation, benchmarks, data-quality, auditing, data-engineering, ELT
2603.28589Towards a Medical AI Scientist
PDF
cs.AI, cs.LG86Autonomous “AI Scientist” tailored to clinical research with evidence grounding and traceability.agents, autonomous-research, medical, evidence-grounding, traceability, LLM
2603.29199AEC-Bench: A Multimodal Benchmark for Agentic Systems in Architecture, Engineering, and Construction
PDF
cs.AI85Open benchmark for multimodal agentic AEC tasks; includes harness techniques and baselinesagent-evaluation, benchmarks, multimodal, tool-use, real-world-tasks
2603.29182Dummy-Aware Weighted Attack (DAWA): Breaking the Safe Sink in Dummy Class Defenses
PDF
cs.LG, cs.CR84Shows dummy-class defenses fool AutoAttack; proposes DAWA for proper robustness evaluation.adversarial-robustness, evaluation, attack-methods, security, benchmarks
2603.29410AGFT: Alignment-Guided Fine-Tuning for Zero-Shot Adversarial Robustness of Vision-Language Models
PDF
cs.CV, cs.AI, cs.LG84Adversarially robust VLM fine-tuning while preserving cross-modal alignment and zero-shot performance.vision-language, adversarial-robustness, alignment-preservation, fine-tuning, zero-shot
2603.22904Separating Diagnosis from Control: Auditable Policy Adaptation in Agent-Based Simulations with LLM-Based Diagnostics
PDF
cs.AI84Separates LLM diagnosis from deterministic control for auditability in adaptive policy interventionsauditability, LLM, governance, agent-based-simulation, control, safety
2604.00706AfrIFact: Cultural Information Retrieval, Evidence Extraction and Fact Checking for African Languages
PDF
cs.CL84Fact-checking dataset for 10 African languages; exposes retrieval gaps and supports grounded verificationfact-checking, retrieval, low-resource, multilingual, grounding, misinformation
2603.09331Reward-Zero: Language Embedding Driven Implicit Reward Mechanisms for Reinforcement Learning
PDF
cs.LG84Language-embedding implicit rewards for RL; could impact instruction-following agents & reward hacking analysis.reinforcement-learning, language-reward, implicit-reward, agents, specification-gaming
2604.00835Agentic Tool Use in Large Language Models
PDF
cs.CL84Structured survey of LLM tool-use paradigms, failure modes, and eval gaps; useful map for agent safety worktool-use, agents, survey, evaluation, failure-modes, LLM-systems
2603.11709Scaling Laws for Educational AI Agents
PDF
cs.AI84Proposes “agent scaling law” + structured AgentProfile; relevant to evaluating/engineering agent capability growth.agents, scaling-laws, agent-evaluation, multi-agent, education, specifications

AI 论文洞察简报

2026-04-05

0) 执行要点(先读这个)

  • “测试时扩展(test-time scaling)”正从文本走向具身控制:SAIL 表明,通过在连续轨迹上用 MCTS 花更多推理算力可以显著提升成功率(45 个节点时平均成功率 25%→73%),随后再蒸馏以恢复低延迟。
  • 验证(verification)正在成为跨领域的主导训练/评估原语:求解器作为验证器(EVOM)、仿真器/数字孪生作为验证器(SAIL)、用于智能体的基准验证器(AEC-Bench、HippoCamp),以及能纠正基准本身的审计流水线(ELT-Bench-Verified)。
  • 联邦学习安全进入军备竞赛:一种主动的客户端侧缓解(FL-PBM)可在测试的交通标志设置上将 ASR 压到接近 0–5%;而更贴近现实的语义感知攻击(SABLE)即使在鲁棒聚合器下也能实现高 ASR——说明角落贴片触发器的评估已不再具有代表性。
  • 隐私审计正转向更强威胁模型与自动化:白盒梯度“特征漂移”MIA(G-Drift)在问答基准上报告接近上限的 AUC;AutoMIA 用智能体去发现跨 VLM 的 logits 级 MIA;SERSEM 表明代码特定的 MIA 需要结构感知加权才能超过通用基线。
  • 鲁棒性评估本身也在被攻击:DAWA 展示标准对抗目标会高估 dummy-class 防御的鲁棒性(例如 CIFAR-10 上某领先防御的鲁棒准确率 58.61%→29.52%)。
  • 长上下文可能悄然降低深思熟虑:Reasoning Shift 发现在无关长前缀/多轮/打包子任务下,推理轨迹最多缩短约 50%,自我验证减少——这对把子任务嵌入长历史的智能体流水线很关键。

2) 关键主题(聚类)

主题:测试时算力与搜索以获得鲁棒性(具身 + 智能体)

主题:验证器无处不在(执行验证、仿真验证与评测框架验证的学习/评估)

主题:现实异质性下的联邦不确定性与安全

主题:隐私审计与成员推断在多样化(白盒、智能体化、领域特定)

主题:鲁棒性评估陷阱与带保证的修复

主题:在真实工作流中基准化与审计智能体能力

3) 技术综合

  • 搜索 + 评分正在跨模态收敛:SAIL 用 VLM 导出的逐帧进度奖励做 MCTS;COvolve 用收益矩阵 + MSNE 在档案上稳定——两者都是“在学习到的评估器引导下,对候选进行种群/搜索”。
  • 表征空间成为新的路由层:TrustFed 通过嵌入距离把测试样本分配给客户端;SABLE 显式分离触发与干净特征;两者都把嵌入当作隐私/鲁棒约束下的操作接口。
  • 可结果验证的 RL 正在扩展到数学之外:EVOM 用求解器执行作为奖励;类似“验证器回路”也出现在 SAIL(数字孪生执行)与基准 harness(AEC-Bench/HippoCamp)中。
  • 基准正确性成为一等变量:ELT-Bench-Verified 显示 33% 的列不匹配可归因于基准本身,修复后 SRDT 22.66%→32.51%,意味着许多“智能体失败”其实是评估失败。
  • 对抗鲁棒性需要防御感知目标:DAWA 表明若攻击目标不匹配防御机制(dummy sink),鲁棒性会被高估;这呼应了评估中更广泛的目标不匹配担忧。
  • 对齐/鲁棒微调正转向“结构保持”目标:AGFT 使用预训练的软图像→文本分布(加校准)以在提升鲁棒性的同时保留 CLIP 对齐。
  • 智能体安全日益成为“行为控制平面”:Silicon Mirror 用基于风险的门控 + 生成器-评论家重写;Secure Forgetting 用转换模型生成遗忘提示与记忆编辑——两者都是不改基础权重的编排级控制。
  • 长上下文流水线可能降低安全裕度:Reasoning Shift 表明加入无关上下文会减少自我验证行为,这可能与累积长历史的智能体系统产生不良交互。
  • 系统瓶颈正在变成基础设施瓶颈:Heddle 显示通过以轨迹为中心的调度/放置/资源分配,rollout 吞吐可提升至 2.5×——若验证器重循环成为标准,这将至关重要。

4) Top 5 论文(含“为何是现在”)

1) SAIL: Test-Time Scaling for In-Context Imitation Learning with VLM(VLM 的上下文模仿学习测试时扩展)

  • 将脆弱的一次性 VLM 轨迹预测变为 对完整轨迹做 MCTS,并结合 VLM 评分 + 步级反馈。
  • 随算力强扩展:从 1 次 rollout 到 45 个节点,平均成功率 25%→73%;真实世界 BlockIntoBowl 5/6 成功。
  • 蒸馏将执行时间 644.72s→72.306s,使“多想再压缩”可落地。
  • 质疑点:依赖与试验匹配的仿真器/数字孪生;sim-to-real 差距(位姿/接触)仍会导致失败。

2) TrustFed: Enabling Trustworthy Medical AI under Data Privacy Constraints(隐私约束下的可信医疗 AI)

  • 在异质性下实现实用的联邦保形预测:表征感知的客户端分配最大聚合阈值
  • 大规模评估(>43 万张图像,六种模态),在异质/不平衡下经验覆盖率接近标称值。
  • 仅交换 标量距离与阈值,契合隐私约束。
  • 质疑点:邻域大小选择是经验性的;仅限分类与单模态任务。

3) Execution-Verified Reinforcement Learning for Optimization Modeling(用于优化建模的执行验证强化学习)

  • 用求解器作为确定性验证器进行仅结果 RL(无过程轨迹),采用 GRPO/DAPO 更新。
  • 在部分基准上匹配/略超过程 SFT(如 OptiBench 62.95% vs 60.96%),并展示 零样本求解器迁移优势。
  • 给出具体冷启动配方(通过跨求解器翻译做小规模 SFT,再执行-RL)。
  • 质疑点:约束/语义错误仍是主要残余失效模式;需要执行 harness 基础设施。

4) DAWA: Breaking the Safe Sink in Dummy Class Defenses(打破 dummy 类防御中的安全汇)

  • 揭示系统性评估缺陷:标准攻击会落入 dummy “安全汇”,从而高估鲁棒性。
  • DAWA 的目标同时针对真实与 dummy logits;鲁棒准确率显著下降(如 CIFAR-10 上 PGD-AT+DUCAT 58.61%→29.52%)。
  • 计算高效,易于集成到评估套件。
  • 质疑点:仅在 CIFAR-10/100 的 ℓ∞ 下展示;更广数据集/威胁模型未展示。

5) Reasoning Shift: How Context Silently Shortens LLM Reasoning(推理迁移:上下文如何悄然缩短 LLM 推理)

  • 发现当同一问题被嵌入到无关长上下文/多轮/打包子任务中时,推理轨迹最多缩短约 50%
  • 句子级分析显示自我验证减少,且更可能在首次答案后停止。
  • 与长上下文智能体与工具使用系统高度相关(它们常把子任务嵌入大历史中)。
  • 质疑点:上下文是合成的(如长莎士比亚前缀),深度轨迹分析主要聚焦单一模型。

5) 实用下一步

  • 面向具身智能体:原型化“轨迹搜索 + 学习评分器”回路(MCTS/beam),测量成功率随节点预算变化;再测试蒸馏以恢复延迟(SAIL 风格)。
  • 面向验证器式 RL:若你有确定性检查器(求解器、编译器、仿真器),实现仅结果 RL 与分组更新,并跟踪哪些错误类型仍残留(EVOM 强调约束错误)。
  • 面向 FL 部署:用语义、分布内触发器(SABLE 风格)评估后门防御,而不仅是角落贴片;并在自适应攻击者下单独测试客户端清洗(FL-PBM)。
  • 面向联邦医疗 ML 的不确定性:加入表征感知路由 + 保守阈值聚合(TrustFed),并扫描邻域大小以绘制覆盖率–集合大小前沿。
  • 面向隐私审计:在可行时至少运行一种白盒 MIA(G-Drift)与一种灰盒场景下的自动化策略搜索(AutoMIA);若审计代码模型,再与代码领域特定 MIA(SERSEM)对比。
  • 面向鲁棒性评估:若使用 dummy-class 防御,纳入dummy-aware 成功准则与 DAWA 类损失;否则你可能在基准化“安全汇”,而非鲁棒性。
  • 面向长上下文智能体:增加监控以记录“推理 token 预算使用量”和自检行为随上下文长度变化;测试上下文压缩或子任务隔离是否能恢复验证(由 Reasoning Shift 启发)。
  • 面向基准/harness:为基准审计预留时间——ELT-Bench-Verified 显示修正可显著改变结论;把验证脚本视为模型的一部分。

由逐篇论文分析生成;未进行外部浏览。