AI 论文日报(2026-03-24)

Published:

English version: /paper-news/2026-03-24/

运行统计

  • 候选论文: 1193
  • 入选论文: 30
  • 已精读完成: 30
  • 时间窗口 (UTC): 2026-03-20T00:00:00Z → 2026-03-21T00:00:00Z (weekend_backlog_sun, 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
2603.14923Directional Routing in Transformers
PDF
cs.LG, cs.AI94New transformer routing; strong causal ablations + mech interp show routing is dominant pathwaytransformers, routing, mechanistic-interpretability, circuits, architecture
2603.14723Beyond Creed: A Non-Identity Safety Condition A Strong Empirical Alternative to Identity Framing in Low-Data LoRA Fine-Tuning
PDF
cs.CL94Shows non-identity safety supervision beats identity framing in low-data LoRA on HarmBench.llm-safety, fine-tuning, LoRA, HarmBench, jailbreak-robustness, supervision-design
2603.18444Discounted Beta--Bernoulli Reward Estimation for Sample-Efficient Reinforcement Learning with Verifiable Rewards
PDF
cs.LG, cs.AI91Sample-efficient RLVR via reward distribution estimation; directly targets LLM reasoning post-training.RLVR, post-training, reasoning, sample-efficiency, reward-modeling, LLMs
2603.18545CoDA: Exploring Chain-of-Distribution Attacks and Post-Hoc Token-Space Repair for Medical Vision-Language Models
PDF
cs.CV, cs.AI90Clinically plausible distribution-shift attack chain + token-space repair for medical VLM robustness.robustness, distribution-shift, medical, vision-language, attacks, repair
2603.18495Cross-Domain Demo-to-Code via Neurosymbolic Counterfactual Reasoning
PDF
cs.AI90Neurosymbolic counterfactuals for demo-to-code; aims at verifiable procedure adaptation under domain shift.agents, robotics, neurosymbolic, counterfactual-reasoning, verification, code-generation, VLM
2603.15600From Passive Observer to Active Critic: Reinforcement Learning Elicits Process Reasoning for Robotic Manipulation
PDF
cs.RO, cs.AI, cs.CL, cs.CV90RL turns video MLLM into goal-aware process critic for long-horizon robot manipulation monitoringrobotics, process-supervision, reinforcement-learning, multimodal, monitoring
2603.19223F2LLM-v2: Inclusive, Performant, and Efficient Embeddings for a Multilingual World
PDF
cs.CL, cs.AI90Large multilingual embedding family (80M–14B) with strong MTEB results; useful for RAG/search.embeddings, multilingual, retrieval, MTEB, efficiency, distillation
2603.18411TARo: Token-level Adaptive Routing for LLM Test-time Alignment
PDF
cs.CL, cs.AI, cs.LG89Token-level test-time alignment routing using step-wise reward signals; sizable reasoning gains claimed.test-time-alignment, reasoning, reward-model, routing, inference-time, LLMs
2603.11558RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks
PDF
cs.RO, cs.AI88Unified VLM-driven long-horizon robotics with self-resetting data collection via entangled action pairsrobotics, agents, VLA, long-horizon, data-collection, self-improvement
2603.17425Proactive Knowledge Inquiry in Doctor-Patient Dialogue: Stateful Extraction, Belief Updating, and Path-Aware Action Planning
PDF
cs.AI88POMDP-lite proactive inquiry for doctor-patient dialogue; explicit belief updates and gap-aware planning.agents, planning, POMDP, uncertainty, dialogue-systems, clinical, tool-use
2603.17872Mitigating LLM Hallucinations through Domain-Grounded Tiered Retrieval
PDF
cs.CL, cs.AI86Tiered retrieval+verification pipeline (LangGraph) targeting hallucination reduction in high-stakes domains.hallucinations, RAG, verification, agents, reliability, grounding
2603.18914Security, privacy, and agentic AI in a regulatory view: From definitions and distinctions to provisions and reflections
PDF
cs.CR, cs.AI, cs.CY86Clear regulatory synthesis for security/privacy of agentic AI; useful for governance & deployment.agentic-ai, regulation, security, privacy, EU-AI-Act, governance
2603.14889Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness
PDF
eess.AS, cs.CL, cs.LG86Speech dialogue reward model + new preference dataset targeting prosody & colloquialness gapsreward-modeling, speech, preference-data, evaluation, alignment
2603.19131From Inference Efficiency to Embodied Efficiency: Revisiting Efficiency Metrics for Vision-Language-Action Models
PDF
cs.LG, cs.RO86Proposes embodied efficiency metrics for VLA robots; challenges FLOPs/params as proxy for real performanceembodied-agents, VLA, evaluation, efficiency-metrics, robotics
2603.11863CreativeBench: Benchmarking and Enhancing Machine Creativity via Self-Evolving Challenges
PDF
cs.AI86Creativity benchmark with executable metrics to separate novelty from hallucination; self-evolving challenges.evaluation, benchmarks, code-generation, self-play, open-endedness, reliability
2603.09868CarbonBench: A Global Benchmark for Upscaling of Carbon Fluxes Using Zero-Shot Learning
PDF
cs.LG, physics.ao-ph86Large zero-shot spatial transfer benchmark for carbon fluxes; strong eval protocols + scale.benchmark, evaluation, zero-shot, domain-generalization, time-series, climate
2603.14712Towards Next-Generation LLM Training: From the Data-Centric Perspective
PDF
cs.CL, cs.LG86Data-centric LLM training: agentic data pipelines, selection/mixture optimization; high leverage bottleneck.llm-training, data-centric-ai, data-mixtures, data-selection, agents, scaling
2603.09356Democratising Clinical AI through Dataset Condensation for Classical Clinical Models
PDF
cs.LG, cs.AI, cs.CR86DP dataset condensation for non-differentiable clinical models; practical privacy+utility angleprivacy, differential-privacy, dataset-condensation, synthetic-data, healthcare, reliability
2603.14838The Impact of Ideological Discourses in RAG: A Case Study with COVID-19 Treatments
PDF
cs.CL86Directly studies ideological retrieval effects in RAG on COVID treatments; relevant to grounding risks.RAG, bias, ideology, misinformation, evaluation, prompting
2603.18388Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization
PDF
cs.AI, cs.MA86Makes reflective prompt optimization more interpretable/robust with multi-agent verification and restarts.prompt-optimization, reflection, agents, robustness, interpretability, evaluation
2603.15262Probe-then-Plan: Environment-Aware Planning for Industrial E-commerce Search
PDF
cs.AI84Probe-then-plan grounds LLM search plans in live retrieval state to cut latency and invalid tool plans.agents, tool-use, planning, retrieval, latency, deployment
2603.19185MIDST Challenge at SaTML 2025: Membership Inference over Diffusion-models-based Synthetic Tabular data
PDF
cs.LG84Challenge-style benchmark on membership inference vs diffusion synthetic tabular data; concrete privacy eval.privacy, membership-inference, diffusion-models, synthetic-data, tabular, benchmark
2603.17312Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress
PDF
cs.CV, cs.AI84Recurrent snippet-based VLM reasoning for long-horizon task progress; cheaper than full-trajectory videoembodied, VLM, reasoning, long-context, planning, monitoring
2603.11479Grammar of the Wave: Towards Explainable Multivariate Time Series Event Detection via Neuro-Symbolic VLM Agents
PDF
cs.LG, cs.AI, cs.MA84Neuro-symbolic agent framework to ground natural-language event specs into time-series intervalsagents, neuro-symbolic, grounding, time-series, evaluation
2603.19225FinTradeBench: A Financial Reasoning Benchmark for LLMs
PDF
cs.CE, cs.AI, cs.CL, cs.IR, q-fin.CP84New benchmark for LLM financial reasoning over fundamentals + trading signals; closer to real analyst workflowsLLM-evaluation, benchmark, reasoning, finance, multisignal
2603.15183Token Coherence: Adapting MESI Cache Protocols to Minimize Synchronization Overhead in Multi-Agent LLM Systems
PDF
cs.DC, cs.AI, cs.LG, cs.MA84Maps multi-agent LLM sync to cache coherence; proposes lazy invalidation to cut coordination costmulti-agent, systems, coordination, scalability, synchronization
2603.19002RADIUS: Ranking, Distribution, and Significance - A Comprehensive Alignment Suite for Survey Simulation
PDF
cs.CL84Proposes standardized alignment metrics for LLM survey simulation incl. ranking+distribution.evaluation, alignment-metrics, survey-simulation, benchmarking, distribution-shift
2603.18447SODIUM: From Open Web Data to Queryable Databases
PDF
cs.DB, cs.AI, cs.CL, cs.CV, cs.IR84Formalizes web-to-database agentic pipeline + benchmark; relevant to tool-using agents and data quality.agents, information-extraction, web, databases, benchmark, tool-use
2603.18481T-QPM: Enabling Temporal Out-Of-Distribution Detection and Domain Generalization for Vision-Language Models in Open-World
PDF
cs.CV, cs.LG83Temporal OOD detection for VLMs under drift + covariate shift; open-world robustness focus.OOD-detection, robustness, vision-language, distribution-shift, domain-generalization, evaluation
2603.08321CORE-Acu: Structured Reasoning Traces and Knowledge Graph Safety Verification for Acupuncture Clinical Decision Support
PDF
cs.AI82Neuro-symbolic CDS: structured reasoning traces + KG safety verification to constrain clinical outputs.safety, clinical, neuro-symbolic, knowledge-graphs, reasoning-traces, verification

AI 论文洞察简报

2026-03-24

0) 执行要点(先读这个)

  • “验证-修订(verify-and-revise)”正在固化为可复用的安全模式:CORE-Acu(临床知识图谱否决 + 有界重写)、用于幻觉的分层检索验证、以及用于机器人学的神经符号反事实验证,都体现了同一种动作——生成 → 对照显式约束/世界模型检查 → 修订或拒答
  • RAG 正越来越多地被视为攻击面,而不只是事实性修复手段:意识形态检索上下文可测量地引导输出(且显式意识形态描述会放大这种效应);而分层检索流水线仍会在错误前提的过度断言上失败——这表明“检索治理 + 可回答性门控”正在变成必选项。
  • 轻量级路由/协同机制正在三个层面涌现:(i) 架构层(Transformer 内的 Directional Routing),(ii) 解码时(TARo 在 token 级自适应混合 base+reward logits),(iii) 系统层(Token Coherence 在多智能体工作流中替代广播同步)。三者都旨在降低干扰/成本,同时保持行为可控。
  • 时间与分布漂移正在通过基准与协议被“工程化”:CarbonBench 标准化了碳通量回归的零样本空间迁移;T-QPM 面向 VLM 的时间 OOD;CoDA 面向医学影像中的贴近流水线现实的分布链。
  • 低数据对齐可能取决于措辞,而不只是“更多数据”:在 130 个样本的 LoRA 中,匹配的非身份(non-identity)安全表述在三类模型家族上于 HarmBench 优于 creed/constitutional 表述,同时 MMLU/ARC 变化可忽略。
  • 具身部署指标正在与推理指标分化:压缩/剪枝/token/action 减少可以保持成功率,却会恶化 jerk/路径长度/耗时——对 VLA 模型的“效率”主张需要具身效率报告支撑。

2) 关键主题(聚类)

主题:面向高风险决策的神经符号验证闭环

主题:检索既是缓解手段,也是操控通道

主题:路由/协同作为通用控制旋钮(模型、解码、系统)

主题:在真实分布漂移下的鲁棒性(时间、空间、流水线)

主题:对齐、隐私与“创造力”的评测基础设施

3) 技术综合

  • 多项工作在 结构化中间产物 上收敛,将其作为验证单元:症状→病机→原则→穴位链(CORE-Acu)、ELT 模式(SELA)、符号操作符与场景图(NESYCR)、以及状态/事件元组 + 权重(医患探询)。
  • 有界循环 是主导的安全/控制原语:生成–验证–修订(CORE-Acu;分层检索验证;NESYCR 修复),并配有显式回退(人工确认;礼貌致歉)。
  • 路由正在无处不在:模型内部(方向抑制)、解码阶段(token 级 α)、检索阶段(领域/层级路由;双轨检索)、以及服务侧(电商的复杂度感知路由器)。
  • 多篇论文表明:若目标不匹配,指标提升可能具有误导性:Directional Routing 带来大幅 PPL 降低,但多选基准无增益;VLA 压缩改善推理指标却恶化具身 jerk/路径/时间。
  • 时间锚定 正成为长时程理解的通用技巧:PRIMO 的 (I_init, V_seq, I_curr) 输入结构;T-QPM 的时间步条件原型与漂移惩罚。
  • 鲁棒性研究正从“单一扰动”转向 组合的、真实的漂移链(CoDA 的 A∘R∘D),并从静态 OOD 转向 流式时间漂移(T-QPM)。
  • 评测越来越 关注尾部:CarbonBench 报告按站点分位数;T-QPM 报告早/晚时间步的 FPR95/AUROC;Token Coherence 分析波动性区间。
  • 检索/验证系统的反复失效模式是 前提验证:系统可能在错误框架下变得自信(错误前提过度断言;意识形态放大)。
  • 当基础模型冻结时,更偏好轻量适配:带重加权损失的 LoRA(CORE-Acu)、两标量融合学习(T-QPM)、token 空间线性适配器修复(CoDA)、以及仅推理期的激活引导(EvoRePE)。

4) Top 5 论文(含“为何是现在”)

1) CORE-Acu: Structured Reasoning Traces and Knowledge Graph Safety Verification for Acupuncture Clinical Decision Support(用于针灸临床决策支持的结构化推理轨迹与知识图谱安全验证)

  • 引入完整的神经符号安全栈:结构化 S-CoT + 中医 KG + 实体重加权损失 + 生成–验证–修订循环。
  • 报告验证后 0/1,000 次 KG 定义的安全违规,而同一基准上 GPT-4o 为 8.5%
  • 为其他高风险领域提供实用模板:当 token 级实体保真硬性禁忌规则 重要时尤其适用。
  • 保留意见:安全性取决于 KG 覆盖;二元否决可能遗漏细微临床权衡。

2) Token Coherence: Adapting MESI Cache Protocols to Minimize Synchronization Overhead in Multi-Agent LLM Systems(Token Coherence:将 MESI 缓存协议改造用于降低多智能体 LLM 系统同步开销)

  • 将多智能体上下文共享重构为缓存一致性;给出解析节省界与具体协议(CCS)。
  • 使用 TLA+ 模型检查 验证不变量(单写者、版本单调、陈旧度有界)。
  • 仿真显示在不同波动性区间下,惰性失效通知可节省 ~84–95% token——是智能体部署的直接成本杠杆。
  • 保留意见:评估基于仿真;中心化权威与故障下活性仍是担忧。

3) Directional Routing in Transformers(Transformer 中的方向性路由)

  • 增加小型路由器以抑制学习到的 head-space 方向;路由变为 关键承载(关闭路由器会导致回忆/归纳崩塌)。
  • 报告显著 领域困惑度下降(31–56%),参数开销约 ~3.9%。
  • 提供内置、可因果操控的“方向”作为可解释性钩子。
  • 保留意见:随机种子/规模有限;PPL 增益未转化为多选基准增益。

4) TARo: Token-level Adaptive Routing for LLM Test-time Alignment(用于 LLM 测试时对齐的 token 级自适应路由)

  • 学习逐 token 的 base 与 reward logits 混合,避免测试时对齐中脆弱的固定插值。
  • 报告 MATH500 大幅提升(如 Table 1 中 Llama-3.1-8B:32.0% → 54.4%),并对更大骨干呈现弱到强迁移。
  • 适用于重训昂贵但可在解码期进行引导的部署场景。
  • 保留意见:依赖奖励模型质量/领域偏置;全 logits 路由可能损害吞吐。

5) The Impact of Ideological Discourses in RAG: A Case Study with COVID-19 Treatments(RAG 中意识形态话语的影响:以 COVID-19 治疗为例)

  • 表明 RAG 可从检索文本中 传播意识形态立场;加入显式 LMDA 描述通常会放大对齐。
  • 提供量化引导的具体方法学(LMDA + 受控检索 + 语义/词汇相似度 + ANOVA)。
  • “为何是现在”:RAG 在生产中无处不在;该工作凸显了超越幻觉之外的治理缺口。
  • 保留意见:语料领域特定且示例选择经策划;效应可能随检索/重排策略变化。

5) 实用下一步

  • 对任何安全关键助手,原型化 生成–验证–修订 控制器,包含:(i) 显式中间模式,(ii) 确定性约束检查,(iii) 有界重试,(iv) 未解决时的拒答/交接策略。
  • 为 RAG 流水线加入 检索前的可回答性/前提检查门,以减少错误前提过度断言(在分层检索验证中被明确标注为关键失效模式)。
  • 将检索语料视为不可信输入:实施 检索治理(来源白名单、意识形态/偏置检测、chunk 级溯源),并在受控检索两极下测试 立场引导
  • 若运行多智能体工作流,按同步边界度量 token 消耗,并对比 一致性式失效通知 与广播;在上线前验证不变量(单写者、版本单调、陈旧度界)。
  • 使用测试时对齐时,用 自适应路由(token 级 α)替代固定混合,并不仅测准确率,还要测 吞吐成本 与 OOD 行为。
  • 对 VLM 鲁棒性,将评估从单一扰动扩展到 组合流水线漂移(CoDA 风格)与 时间漂移(T-QPM 风格);跟踪早/晚时间步指标。
  • 对具身智能体,在宣称“效率提升”前,同时报告 具身效率指标(jerk、路径长度、完成时间、动作率)与推理指标。

由逐篇论文分析生成;未进行外部浏览。