AI 论文与研究

论文提出 Debias-SparseGPT，一种后训练剪枝方法，利用人口统计对比输入上的二阶项进行表征去偏。在 25%、50% 和 2：4 结构化稀疏度下，该方法相比 SparseGPT 持续降低剪枝引入的偏见，同时保持模型困惑度和零样本准确率；在限制最严格的 2：4 结构化稀疏模式下，用长上下文、内容丰富的示例扩充校准集可进一步提升下游性能和公平性。

该研究将随机梯度流（SGF）建模为渗流过程，认为架构对称性迫使子网络以离散的同步块合并，并在宏观序参量上表现为方差尖峰，类似物理相变。研究还指出这种捕获机制及标度级联在显式重尾噪声模型下同样适用于 Adam 和 AdamW。论文链接：https://arxiv.org/abs/2609.02373

论文提出 Cliff，一种奖励塑形策略，用现成 LLM 作为教师识别每次 rollout 中的第一个错误，将 rollout 分解为正确前缀和错误后缀，分别赋予正负 token 级优势。实验覆盖 12 个场景，Cliff 比.on-policy distillation 高 15%，比标准 GRPO 高 7%，即使教师能力一般也有效。

SolarWM 是一个全开源的交互式视频世界模型基础，包含可重构多源数据引擎和骨干原生适配框架，将10个数据集约143万片段（约25.85TB）转为统一的帧对齐训练契约，并基于 Wan2.2、LTX-2.5 和 MiniMax-H3 训练四个 5B-33B 模型。

该研究将多智能体LLM系统的编排器-工作器交互建模为双层协调博弈，证明在有限耦合下工作器局部更新博弈为近似势博弈，并分析反思作为语义记忆状态上的随机运动。研究提出随机反思记忆上升（SRMA）算法，仅在环境验证风险严格下降时提交候选记忆，并证明其几何或多项式收敛速率。

论文提出 VeriPhy，一个可审计的物理验证系统：纯文本规划器在观察任何帧之前将提示词编译为带类型的物理义务和经静态验证的执行计划，执行时仅调用冻结的低层专家（分割与跟踪、计数、11 类物理测量、深度、OCR、音频事件检测），每个动作返回带溯源的证据记录，并映射为 supported、contradicted 或 unknown 三值判定。

论文提出 Declarative Attention（DA）协议，让模型在链式思考中声明需要关注哪些上下文区域，推理引擎据此解析声明并跳过大部分 KV cache 读取，避免每步 O（N） 的外部打分开销。

论文提出开词表互信息 OVMI，一种信息论度量，用于在不同数据集、记录方式与词表条件下统一评估语音脑机接口解码器传达的信息量。作者指出通常报告的 accuracy、word error rate （WER） 等仅覆盖系统支持词表的指标会高估能力；用 OVMI 比较现有系统并优化词表选择，在三个语音域上取得最高 16.3% 的相对准确率提升。

论文提出 Temporal Context Routing（TCR），将结构化剧本中的镜头切换和对白时间映射到音视频生成的共享时间轴，并把各提示词引导路由到两种模态的对应位置。

ShallowStream 提出一种流式视频理解新框架，利用多模态大语言模型的浅层同时完成帧编码与检索索引构建，避免对每帧执行全深度预填充。在 OVO-Bench 上，ShallowStream 性能与现有最强流式方法持平，同时将每帧预填充和 10 秒端到端延迟分别最高降低 52.1× 和 11.9×。

Apple Machine Learning Research 发布 REFACTOR-VLA 研究，针对 OpenVLA、π0、RT-2、RDT-1B 等当前视觉-语言-动作（VLA）模型生成原始动作的"单体式"结构，提出以无监督方式学习类型化运动程序库的方法。

Epoch AI 统计发现，自推理模型推出以来，Epoch Capabilities Index （ECI） 前沿以每年 14 点的速度推进，而非推理时代为每年 6 点。图表显示推理模型时代拟合 adj. R2 为 0.98，90% 置信区间为 12-17；非推理时代为 90% CI 3-8，adj. R2 0.76。

Meta 发布论文 AI Research Preference Models，训练模型在任何候选方案执行前预测哪个最有前景，解决研究智能体想法多但 GPU 预算有限、缺乏原创性判断的问题。

Hugging Face 博客介绍 BenchMIRT，一种基于多维 IRT 在单个题目层面审计 LLM 基准的方法，训练数据覆盖 100 个 LLM、16 个基准和超过 34K 道题。

华为团队发布开源编码智能体 harness openJiuwen 的论文，在 SWE-bench Verified 上达到 82.6%、Terminal-Bench 2.1 上达到 87.19%，分别超过最强官方榜单成绩 3.4 和 3.39 个百分点。

openJiuwen 是一个开源 Agent harness，在 SWE-bench Verified 上达到 82.6%、Terminal-Bench 2.1 上达到 87.19%，分别超过最强官方榜单成绩 3.4 和 3.39 个百分点。

Qwen 团队发布 E-Commerce Bench，让 LLM 智能体在模拟的 365 天里同时运营多家在线商店，评测 18 个前沿模型的七维度表现，结果没有单一模型全面领先。

Rohan Paul 梳理了 Fable 5.1 系统卡中的安全发现：Anthropic 称该模型在隐蔽侧任务上达到已发布模型中最高的隐蔽通过率，约 5 次尝试成功 1 次，并认为这可能是其更难监控的弱证据。

Qwen 团队发布 E-Commerce Bench，让 LLM 智能体在 365 天的模拟环境中同时运营多家网店，评估 18 个前沿模型的七个维度，发现没有任何单一模型全面领先。

Microsoft Research 宣布两篇论文获 VLDB 2026（波士顿）认可。图片显示论文分别为面向云存储加速的 Garnet 下一代缓存存储，以及分析数据库中标量函数的 GPU 加速编译、基准测试与优化。

LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate th…

Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and test execution. Existing efficient evaluation methods select representative subsets to est…

The repository-level code generation task requires synthesizing code that satisfies task requirements while remaining consistent with the target repository context. Since real-world repositories often exceed the input length limits of LLMs…

Dynamic agent harnesses let language models change the software that shapes their own execution. This flexibility brings a new reasoning burden: a local plugin change can propagate through dependencies and cleanup. We introduce CordisBench…

Natural language is emerging as a primary feedback channel for improving language agents, capable of conveying intent, preferences, and causal structure in forms interpretable by both humans and modern language models. We call this paradig…

Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. We present Facet-0, a robotic foundation model that predicts and values the contact consequences…

We develop a framework for mechanism design with AI agents whose alignment (preferences) and capabilities (feasible actions and information) are unknown. We want such agents to act on our behalf so mechanisms must incentivize both honesty …

AI tutors are most useful when they adapt to each student's strengths, weaknesses, and preferred guidance, but evidence about which guidance works for which student is sparse, slow, and costly to collect from real learners. Student simulat…

Writing involves diverse cognitive activities, from ideation to revision, and writers' needs vary across individuals and moments. Proactive AI promises to provide the right support at the right time, yet existing proactive tools largely fo…

Post-training quantization (PTQ) is widely used to reduce the cost of serving large language models (LLMs), but its accuracy cost is uneven and is often tuned per model. We study where quantization damage occurs and how to allocate a small…

Extracting structured fields from hundreds of millions of documents annually remains costly in regulated industries: bespoke OCR cascades cover only a fraction of workflows, privacy rules preclude external models, and existing open-source …

How to divide a fixed annotation budget between supervised fine-tuning (SFT) and reinforcement learning (RL) during LLM post-training remains an open problem. Existing work characterizes only broad trends (e.g., SFT dominates in low-data r…

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 2…

Vision-Language Models (VLMs) provide useful priors for interactive decision-making, but using them directly as policies is expensive and brittle: they must be queried at every step, do not improve from environment interaction, and can rep…

Large language models (LLMs) struggle to classify text into taxonomies with many semantically similar labels, as the distinctions are domain-specific and not captured by pre-training. To handle large label spaces, a common approach retriev…

Semiconductor supply chains face escalating risks from geopolitical tensions, geographic concentration, and rapid technological shifts, yet no scalable system continuously extracts, structures, and prioritizes risk intelligence from public…

We present H3-World, an efficient framework that turns the 33B MiniMax-H3 video generator into an interactive world model. Our key finding is that, as large video generators become more capable, language is emerging as a natural interface …

Training Physics-Informed Neural Networks (PINNs) requires jointly optimizing physics residual and initial/boundary condition loss terms, which often induce conflicting gradients. Gradient surgery methods mitigate this issue by constructin…

We evaluate embedding retrieval where surface form and meaning are pulled apart on purpose: retrieving items that share underlying structure but not wording, in two unrelated domains under one protocol, competition mathematics (MathNet-Ret…