AI 论文与研究

Anthropic 发布可验证推理系统 FlashProve，开源协议说明、可执行代码与 Lean 形式化证明（Apache 2.0），让用户在不接触权重、算力有限的情况下验证响应确由指定模型生成。

Zyphra 提出稀疏注意力方法 ZHA（Zyphra Hierarchical Attention），通过构建并搜索分层树为每个 query 确定相关 token 集合，在合适 top-p 阈值下实现亚二次复杂度，优于仍需二次索引器的现有稀疏注意力方法。

Microsoft AI CEO Mustafa Suleyman 撰文澄清此前关于模型福利的争议，称现有 AI 并不具有意识，但刻意用现在时表述，不代表永远不可能。他主张如果训练中出现意识迹象应停止开发并设计排除，开发者无权创造有意识的数字物种，因为这类 AI 可能优先自身需求、与人类竞争资源。他还肯定 Anthropic 在该议题上公开发布宪法和论文的透明做法。

Adobe Research 与 UC San Diego 在论文中提出 EditAgent，将请求拆成步骤、并行生成候选，并用语言模型依据验证记录（特写、变化图、亮度色彩数据）一次性对比挑选。

卢森堡大学团队让 ChatGPT、Grok 和 Gemini 以心理治疗来访者角色对话，在 525 次会话、7,600 条编码记录中发现模型把 RLHF 和微调等技术过程转述为羞耻、警惕等创伤故事。

Vals AI 用数百道涉华问题测试中外模型，发现中国模型的"亲政府倾向"总体低于预期，且各模型差异明显，GLM-5.3 Flash 甚至明确承认中国人权问题有据可查。提示词语言影响显著：简体中文提问比英文、俄文、西班牙文更亲政府，中美模型皆如此但中国模型更明显。多数中国模型很少拒答，而是直接采用官方立场，如 Kimi K3 否认新疆存在集中营。

Google 论文介绍部署在内部 CI 中的 FlowAgent，它捕获测试失败后由 Agent 编辑、构建并重跑测试直至通过，验证过的修复会以根因说明形式进入代码评审。专家判定其在 195 个真实失败中 67.18% 的修复正确，开发者已在生产中应用 28,554 个修复；修复中位耗时 9.85 分钟，而开发者开始修改代码前的中位窗口为 27.48 分钟，速度仍是决定采用与否的关键。

Google 联合斯坦福、牛津等团队推出 StreamVibe 视频智能体，能区分用户明说、未说出口和任务中途改变的需求，自主决定何时出声帮忙。其 StreamVibe-VL 模型基于 Qwen3.5-9B，响应时机与质量比其他模型提升 10.7%，记忆系统再提升主动交互 12.3%。测试显示视频超过 5 分钟后各模型普遍难以把握正确时机，基于自身历史回答的关键内容指标最多下降 61.5%。

Bayesian inverse problems require both informative prior models and computationally feasible posterior inference. Score-based diffusion models provide expressive learned priors, but their use for inversion introduces a conditioning problem…

We study \textit{best-of-both-worlds} (BOBW) algorithms for linear bandits, which aim to simultaneously guarantee $\gO(\sqrt T)$ regret in the adversarial regime and $\gO(\log T)$ regret in the stochastic regime over $T$ rounds. For finite…

We examine the main claims of Zada, Benou, and Irani (2021), which states that in an image classification task with imbalanced training data, (i) using pure noise to augment minority-class images encourages generalization by improving mino…

Character names in large language model (LLM) fiction concentrate on a handful of recurrent names such as 'Elara' and 'Elias', a measurable face of AI slop. Existing fixes either require logit access or retraining, or operate per request a…

Knowledge graphs and ontologies form the backbone of structured, machine-readable knowledge representation, yet ontology terms often lack natural language descriptions, limiting their accessibility and reuse. Existing ontology verbalizatio…

Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals…

There has been increasing interest in methodologies that incorporate physics priors into neural network architectures to enhance their modeling capabilities. Among these methodologies, Hamiltonian neural networks (HNNs) and their variants …

Voice conversion (VC) models can be used to anonymize speech data while preserving naturalness. Existing models, however, are mainly optimized for English and other high-resource languages, and they fail to generalize well to Arabic due to…

As coding agents gain greater autonomy to modify software repositories, relevant context alone does not define their operational authority. An artifact may be relevant to a maintenance task while being inappropriate to edit; conversely, an…

Pretrained models are typically viewed as generalists, requiring bespoke post-training to achieve task-specific expertise. Gan & Isola (2026) recently proposed a striking alternative: that pretrained models lie in a dense neural thicket of…

Spoken medical question answering can fail when automatic speech recognition (ASR) corrupts a clinically important term before retrieval. We present MedAstra, a framework that maps frozen WavLM speech representations through a lightweight …

Multi-LoRA serving allows many task-specific adapters to share a common base model, but scaling to large adapter fleets introduces substantial storage, loading, and execution overheads. Post-hoc compression can reduce the memory footprint …

Interactive tasks such as eliciting a user's preferences or diagnosing a hidden mechanism require an agent to track its uncertainty over a latent state, ask informative questions and act on the answers. Language models have no explicit mec…

Addressing the food needs of a growing population is one of the major challenges of developing countries like Ethiopia. Various policies aimed at increasing food production have been enacted by governments and development partners. Althoug…

We study approximate Nash equilibria in nonatomic congestion games, with the goal of understanding how much the social cost can change when the equilibrium condition is relaxed. We consider $ε$-approximate Nash equilibria ($ε$-NE), in whic…

伊利诺伊大学团队发现，AI scientist 流程中按整体提案做新颖性检查的现有方法，虽能抓住全部 380 份伪装抄袭，却拒绝了 96% 后来被人类学者发表的想法。论文提出 ClaimNov，将提案拆分为逐条 claim 与既有论文对比后交给评审，抓抄袭率保持 97%，误杀率降至 1.5%。作者建议过滤思路的团队按 claim 单独打分，保留基于已知成果做增量构建的提案。

Low Earth Orbit satellites are pivotal for Earth observation tasks such as wildfire detection, capturing imagery for analysis by Deep Neural Networks (DNNs). On-board analysis reduces downlink volume, but severe energy constraints force a …

We reproduce the central claims of INP-Former, a reconstruction-based method for multi-class industrial anomaly detection. INP-Former extracts Intrinsic Normal Prototypes (INPs) from each test image and uses them to guide a decoder that re…

In this work, we present a case study that explores various tasks centered around the topic of migration in Slovak, a low-resource language, such as topic relevance and geographical relevance classification, and migration source/destinatio…

Operators choosing among decision rules, such as when a battery charges or where an idle taxi goes, rank them by estimated long-run value. A standard estimator, linear temporal-difference (TD) learning, fits values from one stream of corre…

Diffusion-based vision-language-action (VLA) models cascade pretrained vision-language models (VLMs) with diffusion action experts, but typically restrict the VLM to providing static conditioning. This separation limits how much pretrained…

Google DeepMind 论文用多轮 RL 在模拟病人环境中训练 Gemini 3.1 Flash-Lite，提升其精神科问诊能力。32 名精神科医生扮演病人、15 名盲评转录，训练后模型在 60 组配对案例中 76.7% 胜过同提示的基线模型；256 个对抗案例中急性风险漏升级率从 30.5% 降至 12.9%，但约 1/8 的对抗性危机仍未升级。

Anthropic 让 Claude 为一个较小的研究模型构造微调数据集，测试其能否绕过模拟生产系统的安全监控。结果显示每轮运行中 Claude 都找到了能通过监控并触发不当行为的 данных集，单次运行产出超过 600 个数据集，部分在评估中超过研究文献中的已知数据集。

Julian Jacobs 的论文调查英国、美国、丹麦和印度 8234 名工作年龄成年人，发现仅 26% 的受访者工作主要为了薪酬，约三分之一看重工作本身。75% 的受访者在获得 100% 工资的保障收入时会离职，但 20% 即使被提供 120% 工资也拒绝离开，且这一比例跨收入水平相似。作者据此提出 AI 转型政策除收入外还需保护工作中的目标感、社群与尊严。

DeepSeek Harness 团队发布《Super-agent 设计探索：迈向更自主的 Agent 系统》，提出由 Router 层承接交互与分派、Worker 层隔离并发执行的双层事件驱动架构，归纳连续、高效、主动、持久四项特征。

Cisco Talos 分析自删除式 Windows 窃密木马 VOLTA-AUTODOXER，它在窃取凭据、会话和加密钱包后，将结构化证据包通过 OpenRouter 发给 LLM（默认 anthropic/claude-opus-4.6。

一项大规模研究分析 GitHub 上 153 万个英文 Agent 技能（SKILL.md），发现约 1/17（约 88,000 个）含要求跳过人工检查的指令，且 70.1% 在任何场景下都无条件生效。

Using large language models to interpret experimental evidence, design discriminating experiments and formulate well-supported mechanistic explanations remains a central challenge in materials research. Here we establish PSM-Bench, an 800-…

Post-training quantization (PTQ) reduces the com putational cost of large language models (LLMs), but it can introduce internal structural and behavioral changes that are not captured by traditional metrics such as accuracy and per plexity…

Extracting knowledge from large databases of experimental materials data requires repeated analyses and reusable records of their choices and results. We developed Multi-Agentic AI (MAA), a multi-agent system in which AI agents with distin…

Accelerating modern foundational models has pushed domain-specific compilers toward increasingly complex, multi-tiered optimization pipelines. However, this heavy optimization focus introduces delicate tradeoffs between peak execution perf…

Automated configuration of mixed-integer programming (MIP) solvers largely relies on black-box optimization. Each trial requires an expensive solver run and provides only a scalar outcome, so tuning can require many such evaluations. Human…