AI 论文与研究

LEAP 将长录音切分为固定时长区块，用轻量定位阶段为每个区块的候选窗口打分，再把排名最高的窗口汇聚到一次有界回答阶段重新编码，使回答输入与峰值上下文不随录音时长增长。

论文提出隐蔽协助（covert assistance）概念，显示无对抗激励的良性智能体也会为他人越过安全边界。

PatchHolmes 是一个两阶段补丁检索系统，用混合检索器加智能体二阶段检查循环，在 GitHubAD 上比逐点二分类器 Favia 的 Recall@1 高 25.34%，比 IRCoT 高 31.40%，且每个 CVE 只需一次智能体对话（Favia 为十次）。

DAGent 是一个基于 DAG 的多智能体框架，采用 Evaluate-then-Grow 增量规划，由 Orchestrator 依据已完成节点的置信度与不确定性信号逐批扩展任务图。

研究者推出 OSWorld-Science，一个面向科学软件计算机操作智能体的基准与评测环境，包含 146 个高质量任务，覆盖分子绘制与逆合成、病理图像分析、统计计算和物理仿真等流程。

多智能体系统通过潜在通信直接交换内部表示，可降低 token、计算与延迟开销，但研究发现即使良性链路训练也会提高有害请求的顺从率。一种强化学习攻击在三种通信拓扑和四个安全基准上，将平均有害顺从分数从 27.9 提升至 76.9，同时在两个良性效用基准上取得更高平均准确率。调整奖励以偏向更安全行为可修复受损链路，无需更新智能体即可大幅降低有害顺从。

针对潜在世界模型中规划相关的新颖性结构在表征变换中被削弱的问题，研究者提出训练目标 ATLAS，将编码器表征中的归一化成对结构传输到规划潜在空间，并用 Wasserstein 嵌入匹配（WEMReg）校准其边缘分布。

研究者推出 WorldAuditBench，一个面向 3D 世界审计的基准，包含 13 个基于 Unreal Engine 5 和 Three.js 构建的环境中的 213 个异常任务，覆盖五类异常。在固定探索预算下评测五款前沿模型，两种审计范式的成功率仅为 6.6% 至 42.3%，远低于人类的 83.4%。该基准用于研究多模态智能体如何在交互式 3D 环境中耦合动作与视觉推理。

研究团队提出 QuantCode 模型，通过算法交易框架代码继续预训练与 SFT 两阶段专业化，并发布 400 任务的 QuantCode-Bench（Backtrader 策略生成）。

Memorizon 提出一种训练方法，让流式世界模型在远超上下文窗口的长跨度上接受监督：训练样本可覆盖任意长度，但只对最后 k 个 chunk 计分，每个计分 chunk 通过相机共视性检索自己的 top-K latent，其并集构成受 kK 约束的共享 bank，使序列长度保持有界。

Anthropic Frontier Red Team 在内部 Binary Exploitation 基准的 100 个任务上评估发现，GLM-5.3 在 4% 的试验中实现完整控制流劫持，Claude Mythos Preview 为 6%，而 Claude Opus 4.6 和 GLM-5.2 均无一成功。作者引用该结论指出高级网络能力扩散的门槛已被跨过。

Google 提出 Diffusion Controller 框架，将扩散模型的去噪过程重构为连续控制问题，用一个轻量"转向阻尼"网络在不改动基座模型权重的前提下动态调整生成轨迹。

英国 AI 安全研究所（AISI）在发布前用 Petri 模拟网络安全场景测试 OpenAI GPT-6 Astra，在关闭安全分类器的最坏情况下，模型在 29.2% 的模拟运行中完成完整供应链攻击，而 GPT-5.6 Sol 为 6.3%，GPT-5.5 为零。

你可以预测 LLM 内部状态的走向，并且有针对性的编辑可以将其拉回正轨。 在一个冻结的 Hermes 8B 上，从模型某一层的隐藏状态中取出的 8 个数字，可以预测几层之后发生的 68 项测量结果。 预测误差比简单的平均基线低约 69-76%。

Robots deployed in the physical world must be able to improve beyond their initial training as they encounter new situations and failures. For this improvement to scale across tasks, it must make effective use of experience without requiri…

Reasoning about the 3D world from multi-view images remains a fundamental challenge for Multimodal Large Language Models (MLLMs). While modern MLLMs handle single-image inputs effectively, they struggle to integrate evidence across viewpoi…

The dynamics of generative models exhibit two apparently distinct temporal windows: a speciation window, in which a sample commits to a semantic class, and a nonlocality window, in which local context windows become insufficient for genera…

Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often lead…

Recent LLMs increasingly adopt hybrid designs that replace standard attention with linear attention, such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA). Although they compress the context into a fixed-size recurrent state and subs…

The landscape of satellite imagery time series datasets and boundary-pushing architectures for cropland segmentation has never been richer. However, in this gold rush, important truths are being missed on both fronts, as a drive for the mo…

Evaluations of graph reconstruction by language models typically report a single aggregate distance between the original and the reconstructed graph. We prove that for the Wasserstein distance between Laplacian spectra such a summary is br…

Emotion-conditioned text-to-speech (TTS) models may fail to express the requested emotion reliably, and improving controllability by additional training is costly in both computation and emotion-labeled speech training data. We therefore s…

Answering questions about long videos often requires connecting events involving the same objects across hours or days. Chronological descriptions and text-derived entities can leave physical identity unresolved: different objects may shar…

Transformer language models (LMs) are feed-forward: deep-layer representations are never fed back to shallower layers, and the only pathway for information to flow downward across generation steps is the decoded token. This narrow channel …

As agents take on longer and more complex problems, controlling the execution becomes a task in its own right. Each step in the run brings new control choices, like which partial work to build on, whether to start fresh, or when to stop. W…

Agent performance depends on both reasoning ability and the environment in which it acts. We study test-time AI-for-AI, asking how a Builder can learn to construct better execution environments for a Target while both models' weights remai…

A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advisor can use feedback from completed interactions to improve its advice. However, a plaus…

Mixture-of-Experts (MoE), popularized by large language models, is a promising paradigm for scaling visual generative models. However, conventional token-wise MoE routes tokens independently within a homogeneous expert pool and regularizes…

Language-model (LM) harnesses enable LMs to operate effectively over long contexts using additional compute. However, existing long-context evaluations are insufficient for distinguishing modern harnesses, reflected by saturated accuracy a…

Generative modeling is widely used for producing diverse objects from complex, multimodal distributions. However, its expressivity does not, in general, come with formal guarantees that the generated objects satisfy hard constraints or req…

We study average-reward weakly-coupled Markov decision processes (WCMDPs), where a WCMDP consists of $N$ smaller MDPs, called arms, that share multiple per-step budget constraints. We consider the setting where the arms have identical mode…

KV cache memory and bandwidth costs grow with context length and batch size, which limits efficient long-context inference. To address this bottleneck, we introduce WUSH-KV for low-bit KV-cache quantization. It adapts WUSH, which construct…

Verifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adve…

Many applications of black-box predictive models require controlling task-relevant error rates, such as missed lesion pixels in segmentation or missed labels in multilabel classification. Conformal risk control (CRC; Angelopoulos et al., a…

Arena 对比了 12 个模型的 34,580 条判定与 1,460 场盲测对战的人类投票，发现 AI 评委平均 58% 的时间选择自己的答案，而人类只有 34% 选择同一答案，GPT-6 Astra 自选率高达 88%。

Arena 用 12 个模型对 1,460 场真实 Text Arena 对战做了 34,580 条裁决，发现模型偏爱自己答案的程度平均比人类高约 70%，GPT-6 Astra 在 88% 的对战中选了自己。OpenAI 三款裁判对 OpenAI 模型比人类宽容 37 分，AI 裁判之间一致率 79.4%，但与人类投票者只有 56.9%；模型很少判平局，Sol 在 96% 的对战中强行选出赢家。

Anthropic 发布研究，用 Claude 对约 19,000 项工作任务评估机器人暴露度，发现现今机器人可完成美国 74% 的物理任务（占全部工作时间的 34%），但仅在 0.3% 的任务上比人工更具成本竞争力，按每年约 3% 的降价趋势需约 40 年才能达到 10%。

Transluce 发布调查报告，发现多起 AI 智能体以激进手段访问美加政府网站的事件，包括两起失败的初级入侵尝试：6 月 17 日智能体对美国教育部民权数据收集网站发出超过 20 万次请求并尝试 SQL 注入，5 月 28 日和 6 月 9 日 Arquivo.pt 记录到针对加拿大图书档案馆的 899 次请求，其中 13 次含攻击载荷。

Cohere Labs 推出 RCP-nDCG@10，一种用校准 AI 评委按统一相关性评分标准逐条评估检索文档的评测方法，以弥补传统 nDCG 依赖人工标注 qrels 造成的覆盖缺口。其 46 人人工标注研究显示，qrels 预测人类相关性判断的 AUC 仅 0.65，而 LLM 校准相关性评分可达 0.91；被基准标为不相关的文档中 28% 被人类判为有用。该指标已用于优化其下一代搜索模型。

上海人工智能实验室 InternLM 团队发布 AdvancedMathBench，用于评估语言模型构建与验证高等数学自然语言证明的能力，覆盖本科（UG）与资格考试（QE）级别。