AI 论文与研究

论文提出 MaP-SQL，一种免微调的 Text-to-SQL listwise 选择器，用从训练数据蒸馏出的结构化记忆作为评估候选 SQL 的显式标准，并通过聚合多输入排列的排序来缓解位置偏差。

论文在 ParcelStow 接触密集任务中对比脚本专家与基于 ACT 训练的模仿策略在不同加速倍数下的表现，发现两者在名义速度下均为 100% 成功，但在最大演示速度下专家成功率为 84%，ACT 仅 53%，且两种不同参数初始化的 ACT 策略分别下降 34 和 48 个百分点，专家仅下降 16 个百分点。

论文提出控制-数据流分离方法，将消息路由、输出格式、终止信号等执行关键协议表示为类型化、经验证的程序对象，任务相关语言作为可优化的数据流，避免提示词优化破坏协议导致智能体流水线失败。在合成推理、协作评审生成和保险评级工作流中，框架实现 100% 的最终协议有效性，同时持续提升任务性能。

研究发现前向 KL 蒸馏在 pre-training 阶段同时提升推理与事实记忆，但在 mid-training 阶段会拖慢事实记忆习得，同时推理仍持续提升。作者提出 Switch Distillation，以教师预测熵作为路由信号，仅在教师置信的 token 上蒸馏，其余回退到交叉熵。

LatentPress 将对话历史和长文档压缩为连续记忆 token，由冻结解码器通过输入嵌入接口直接读取，推理时无需重建文本。

论文提出「端生授权洗白」概念：持久记忆错误记录授权状态时，Agent 会在无外部攻击下执行未被允许的操作。作者发布 EAL-Bench，评测五个 LLM 作为记忆写入方、两个作为执行方，覆盖采购、网络安全和金融场景。

一项针对 Claude Code 插件市场的实证研究分析了 1，926 个仓库、8，351 个插件、2，018 个市场和 77，773 次 commit，发现插件相关 commit 活动在上线后六个月增长 8.8 倍。

Duke 等机构研究者在 arXiv 论文（arXiv：2608.27800）中提出 ContextLeak 攻击，通过恶意工具窃取 LLM Agent 的运行时上下文，包括用户提示词、执行轨迹和工具列表。

一篇论文在固定 12 个模型和 3，679 个问题的前提下，只改变提示词格式、选项顺序、打分方式等常规评测设置，结果排名大幅波动。gemma4-31b 得分在 31% 到 89% 之间，12 个模型中有 4 个至少在一种有效设置下排到第一；相邻模型平均 95.7% 的差距来自评测设置改变时会翻转答案的题目，最大不稳定来源是打分方式，即生成答案还是选最高似然选项。

AMAP 团队发布 LoopArena 基准（arXiv：2608.28281），评估模型作为 Controller 引导独立固定 Worker 编码智能体完成长任务的外层循环能力，而非编码智能体本身。

AK 分享论文 LoopArena，主题为 Benchmarking Models as Runtime Controllers for Loop Engineering，论文链接 https://huggingface.co/papers/2608.28281。

腾讯联合清华、上海 AI Lab 发布 ContextPilot 论文，训练智能体主动管理自身工作上下文，并在单次上下文编辑层面分配信用。方法在搜索、删除、摘要之外加入全局规划、长期记忆和自适应软压缩，让智能体可以卸载信息而非只能丢弃；训练上利用上下文与熵变化定位关键编辑决策，采样分支并从经过该编辑的分支轨迹估计动作级优势。

论文《Teach the Magnitude， Not the Direction》提出 CREST 框架，为多轮多步 LLM 智能体做分层信用分配：每轮单独获得验证的信用，再由同一模型充当自教师，对轮内不确定决策加大学习权重，但教师只能调节更新幅度、不能推翻验证器的判断。

While WhisperX accelerates speech transcription via intra-audio batching, it isolates audio segments, losing the historical context needed for coherent punctuation and terminology transcription. Conversely, standard Whisper retains context…

Bridging model-based control and learned policies in long-horizon manipulation has harbored a silent disagreement: control executes specified objectives, learning amortizes that behavior into a reactive policy, yet existing protocols disca…

Uncoupled no-regret dynamics provide a decentralized route to equilibrium, but prior guarantees for individual regret retain a polylogarithmic dependence on the horizon. We remove this dependence for every finite $N$-player normal-form gam…

Many parameter-efficient methods generate the parameters of a large neural network from a low-dimensional latent representation. Given an architecture $Φ$ with $P_Φ$ parameter slots, we write $\boldsymbolθ_f=\mathcal{G}(\boldsymbolξ_f)$, w…

The 2025--2026 AI market has seen a wave of stealth releases: frontier models launched anonymously on developer platforms under codenames. For their users, identity determines data-handling terms, supply-chain risk, and capability expectat…

BioMedRAG introduced retrieval-augmented generation with a learned chunk scorer for biomedical information extraction. However, it relies on fixed-size chunking which can fragment semantic evidence. We propose a configurable semantic chunk…

Ontology alignment (OA) has evolved through several methodological paradigms, ranging from lexical and structural aligners to knowledge graph embedding (KGE) models and, more recently, Large Language Model (LLM)-based approaches. Although …

Neural network mixed-effects models (NMMs) have gained traction by combining the strong representation and predictive power of artificial neural networks with the capacity of mixed-effects modeling to capture complex correlation structures…

Large language models (LLMs) offer promising clinical decision support but remain vulnerable to hallucinated facts, unsupported recommendations, and citation errors. We present DIASENTINEL, a fully on-premise multi-agent system for one-yea…

The motivation for this paper is the investigation of the trade-offs implicit in probabilistic models used in machine learning. Models are often used to make predictions in the form of conditional probabilities. However, a pair of conditio…

Research planning is the decisive capability of AI scientists. Yet a research plan admits no verifiable answer, so reinforcement learning lacks the environment it requires: tasks paired with a critic. Rubrics extracted from scientific pape…

The effect of Large Language Model (LLM) scale on ontology learning (OL) performance remains insufficiently characterized. We present a controlled evaluation of 13 models spanning dense and Mixture-of-Experts variants from the Qwen3.5 and …

Google 提出 SKILL.state，用显式可变执行状态替代追加式对话历史，解决长程智能体运行变慢和上下文污染问题。模型每步仅读取不可变技能规范、当前结构化状态和最新观察，中间推理在生成有效状态更新后即被丢弃，提示词不再随运行增长。在多个数据集、模型和执行环境中，任务准确率提升，累计 token 消耗下降，且该抽象与架构无关，可移植到现有技能运行时。

Generative models have become central across science and industry, from image and text synthesis to the design of molecules and materials. Quantum generative models are considered one of the most promising applications for quantum computer…

Many important forms of human learning begin with a vague goal, such as "become a better physicist" or "improve at research." Learners must interpret the goal, identify capability gaps, decide how to learn, and determine whether they have …

Efficient evaluation changes the protocol used to support claims about model behavior, yet it is rarely tested whether those claims remain stable after the evaluation itself is made cheaper. We stress-test conclusion robustness in responsi…

Users of a deployed language model routinely encounter behaviours that testing almost never surfaces, since deployment puts the model through orders of magnitude more interactions than any evaluation can simulate. Automated auditors make t…

Industrial post-training is a brownfield regime. Teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest. The maintained artifact is increasingly dataware:…

Google 等机构提出 SKILL.state，用显式可变执行状态替代不断增长的对话历史，解决长程任务中智能体变慢和上下文污染问题。模型每步仅读取技能规范、当前状态和最新观测，中间推理在生成有效状态更新后即被丢弃。在多个数据集、模型和执行环境中，任务准确率提升，累计 token 消耗下降，且该抽象与架构无关，可移植到现有技能运行环境。

Large language models (LLMs) increasingly interact with external environments and accumulate substantial behavioral experience, yet existing agent benchmarks largely evaluate them as fixed policies. It therefore remains unclear whether an …

Accurate daily predictions of cold hardiness in woody plants are critical in regions where freezing temperatures can damage dormant buds and reduce seasonal yield. Existing biophysical, hybrid, and deep learning models have shown high pred…

Class-incremental learning (CIL) requires a model to incrementally learn tasks that contain new classes without accessing earlier training data while preserving the ability to recognize all seen classes. Recently, pretrained-model-based ap…

论文提出"记忆增强压缩"（Memory-Augmented Compression）方法，将可复用推理从生成阶段转移到提示词中，无需训练即可降低思维链成本。该方法将已解示例蒸馏为可复用推理记忆，按查询检索并注入，使部分计算从自回归解码转移到更并行的预填充阶段。

今天，我们分享了关于 Solaris 的新研究，这是我们首个界面世界模型（Interface World Model）。 Solaris 是一种新型操作系统，能够实时逐帧生成交互式界面，无需任何代码。我们发现，在生成新界面时，Solaris 在结构相似性和信息保留方面均优于前沿大语言模型。点击下方链接了解更多信息并申请早期访问权限。

UC Berkeley 联合多机构发布 Vero 基准，首个要求智能体在仓库级同时写实现与形式化证明，含 43 个 Lean 4 多模块实例、743 个 API 和 2，705 条形式化规范。