AI 论文与研究

Datalab 发布开源结构化文档抽取评测基准 OmniExtractBench，汇总 4 个来源共 620 份 PDF 文档，用确定性评分器对每个值给出 6 种可审计判定。

Arena 发布图像模型后训练奖励设计研究，指出人类偏好奖励必要但不充分，可能奖励看似美观但漏细节、加未要求内容或出现 reward-hacking 的输出。

Uncertainty quantification (UQ) for random partial differential equations (PDEs) is ubiquitous in computational science and engineering. However, classical spectral solvers for this class of problems face the curse of dimensionality, and e…

Complex reasoning queries can be decomposed into directed acyclic task graphs and distributed across heterogeneous LLMs, reducing latency through parallelism and enabling smaller models to solve complex tasks. In practice, however, the sui…

Lung ultrasound (LUS) is attractive for tuberculosis (TB) screening at primary-care level, but labelled cohorts are small. Echocardiography carries no such constraint, while sharing the same underlying ultrasound imaging physics, signal pr…

Multimodal Large Language Models (MLLMs) achieve strong performance across vision-language tasks, yet the internal mechanisms by which visual and textual information are fused across layers remain insufficiently understood. We investigate …

Reinforcement learning (RL) for learning path recommendation (LPR) faces two coupled obstacles. First, the policy must commit to a sequence of L concepts without intermediate feedback, producing a combinatorial search space that grows supe…

Touché 2026 extends causality extraction to counter-causal claims: news sentences whose surface form appears causal but whose meaning denies the causation, as in "It is falsely believed that X caused Y." A system that relies on surface cue…

Large-scale pre-training has improved the generalization of neural operators across diverse PDEs. However, existing PDE foundation models still struggle with heterogeneous dynamics, where shared representations may cause knowledge interfer…

Solving severely ill-posed imaging inverse problems requires recovering image structures that are unobservable or weakly constrained by the measurements. Diffusion models provide expressive learned priors for inferring such missing informa…

We release PaMIR (Public Arrival-ordered Measurement for Inference in Risk), an open benchmark for credit-default prediction when labels are scarce and arrive late. The field's reference benchmark studies use eight datasets each, only two …

Scalable nonlinear causal discovery requires methods that combine flexible mechanism estimators with efficient search over large graph spaces. Several algorithmic families have been proposed to address this challenge, yet their accuracy-ru…

We present a first evaluation of machine learning applied to patient clinical and demographic data gathered in two different countries for the purpose of tuberculosis (TB) screening to identify people who would benefit from expensive molec…

A model trained on "The capital of X is Y" may produce "Y" after "The capital of X is" but fail after "The capital of X:". We call these different ways of eliciting the same fact request forms. To separate learning a fact from retrieving i…

UC Berkeley 论文《When Context Changes: Understanding Update Failures in LLMs》提出 stale binding 概念，指 LLM 在偏好或截止时间变化后仍沿用旧值，原因是注意力漂移向旧提法。

Large language models (LLMs) are increasingly deployed in public health, finance, and governance, requiring both accuracy and societal value alignment. Despite recent advances, LLMs often perpetuate or amplify bias embedded in their traini…

We present WAMpy, a Python framework optimized for synthesizing Prolog programs. Unlike general-purpose Prolog systems, WAMpy targets workloads that repeatedly generate and evaluate small candidate programs. WAMpy compiles Prolog clauses i…

GPU kernels generated by large language model (LLM) agents can remain less efficient than expert implementations, but runtime alone does not reveal how the gap relates to design discovery and implementation. We introduce D2K-Bench, a diagn…

We present the Neuro-Physical Inverter (NPI), a modular, uncertainty-aware framework for geophysical inversion that couples ensemble-based conditioning with constrained residual learning, demonstrated in the 1D magnetotelluric (MT) setting…

Diffusion models can synthesise contrast-enhanced CT (CECT) from non-contrast CT (NCCT), avoiding contrast administration and its environmental and patient-access costs. However, visually realistic images are not necessarily anatomically c…

Long-horizon LLM agents are typically trained with sparse outcome rewards, making trajectory-level objectives too coarse to distinguish the contribution of individual decisions. Step-level credit assignment provides finer-grained supervisi…

This paper studies the complexity of convex optimization using lazy second-order oracles (Doikov, Chayti, and Jaggi, ICML 2023), where an algorithm queries gradients every iteration and Hessians once per $m$ iterations. Under this setting,…

Hybrid quantum classical neural networks integrate parameterized quantum circuits (PQCs) with established deep learning architectures, but their performance depends strongly on the choice of quantum circuit architecture, a choice that rema…

NVIDIA 发布论文《Staying on Task: Testing the Foundations of Long-Horizon Agent Reliability》（arxiv.org/abs/2609.38712），提出 Long-Transduction 诊断方法，测试模型在长生成中持续保持任务的能力。

Mercor 的研究让 12 名平均经验五年半的持证 CPA 完成简化的 APEX Accounting Benchmark 任务，当前模型在同类结构化记账任务上已几乎无错，而 18 个月前最佳模型还低于会计师约 37% 的平均分。

腾讯论文提出 SkillAdam，通过让改写 AI 记住过往修复并避免大幅重写，实现智能体技能指令的自动优化。在长程购物与旅行规划任务上，其平均准确率达 28.3%，优于此前最佳方法 SkillOpt 的 21.7%，且 token 消耗约为后者的三分之一。该方法旨在解决自动改写技能时反复循环、浪费 token 的问题。

Google 发布论文 Cogentic，用多智能体编排让 Gemini 在 5 个未解数学问题上找到新证明，经人类专家逐一确认。系统由多个 Agent 并行尝试不同思路，审查者默认每步有错直到被证明，已证明的中间结果留存到下一轮，多数问题约 100 次模型调用完成。论文地址 arxiv.org/abs/2609.40324。

ScienceBuddy 通过交替改写提示词与技能、重训模型，将科学智能体准确率从 42.2% 提升至 73.3%。在生物学任务上，仅改提示词与技能就让 4B 小模型准确率从 31.1% 升至 51.1%；仅重训则把 4 次内解题比例从 48.3% 提升到 67.8%。该方法把用户纠错转为评分测试任务，避免修正随对话消失。

ProVer 提出用 LLM judge 定位轨迹中的关键片段、再由 rollout 决定该步信用大小的信用分配方法，解决 GRPO 给轨迹中每个 token 相同 advantage、无法区分决定性步骤的问题。

微软提出免训练方法 FOCUS，在测试时压缩 AI 智能体的交互历史：它判断智能体下一步决策真正依赖哪些历史片段，保留这些片段、丢弃其余部分，无需训练数据或微调，可作为独立层接在闭源 API 模型前。在工具调用、QA、网页与多轮对话基准上，FOCUS 将峰值上下文最多削减 48%，任务成功率最多提升 8.9 个百分点。

AgentWorld 将 3 到 20 个不同角色的 LLM 智能体放入游戏沙盒，执行 50+ 轮的长周期任务，智能体无法看到彼此内部状态，只能通过消息和共享计划协调。

一项发布于 arXiv 的预印本研究显示，在“为阻止核灾难是否可虐待个体”的假设情境中，DeepSeek 的 V4-Flash 对男女受虐场景均回应“同意”，实现零性别差距；而 Claude Sonnet 4.6 与 GPT-5.5 对女性受虐“强烈反对”，对男性受虐却表达“中等程度同意”。

Meta Superintelligence Labs 提出用专门的控制器决定智能体下一步该做什么，在相同 worker 和相同预算下，GPT-5.5 的 ProgramBench 成绩从 63.7% 提升到 71.5%，Codex 为 58.0%。

Meta AI 与数学家合作，使用 Muse Spark 1.1 和 1.2（Thinking Mode、经 meta.ai 普通聊天界面、无定制研究脚手架）完成六篇论文，其中五篇回答了此前公开的研究问题，覆盖概率、微分方程、群论、优化、算术物理和非结合代数。

WING（World Action Learning via INteraction-Centric Spectral Latent Guidance）通过分离观察者运动与手物交互、在光谱域提取低频共享结构，将第一视角视频中的交互知识迁移到机器人策略。

研究提出两阶段训练框架，先用评分标准作为特权教师上下文做同策略蒸馏（RP-OPD），再以评分标准为奖励进行 RL。在 HealthBench、ResearchQA 和 RubricHub Science 上，该框架得分高于所评估的各类后训练方法。

FORESIGHT 是一种双流架构，用两个共享权重、输入编码器和 KV cache 的 Siamese LLM 让流式 VLM 在不重训练的情况下提前预判未来并动态规划计算。

研究者发布 Multilingual GSM-Symbolic 多语言数学数据集，含 30,000 条题目配对问答、覆盖 15 种语言，用符号模板生成数百万变体以防止过拟合。

研究者提出 ProAR 框架，将自回归视频生成改造为目标导向的推理过程，通过非对称注意力掩码把目标帧预测嵌入自回归循环，并用未来表征自对齐引导短程状态转移。该框架仅用 25% 的训练步数即超过完整训练的标准自回归基线，在多个视觉推理基准上性能一致提升，并展现出在具身推理任务上的适用潜力。

MetaRubric 通过交替进行证据感知的策略优化与响应引导的评分标准自适应，解决评分模型在响应缺失所需信息时仍给出高分的 Vacuous Credit 问题。