AI 论文与研究

Nathan Lambert 表示正为 Mercor 提供研究方向的建议，主张构建可扩展的样本高效 RL 数据方法，并推动评估科学前沿、针对明确经济价值领域建立专门 benchmark。他认为前沿实验室将围绕真实世界代表性评估与合成数据方法大举投入，并看好基于开源模型构建这一方向的机会。

东京都立大学等机构的论文发现，基于厂商旧模型训练的 AI 文本检测器在模型代际更迭前可捕获 99% 以上的改写，更迭后仅捕获 3.8%。论文用 4,000 篇前 ChatGPT 时代摘要由 25 个 LLM 版本改写进行测试；引用案例中 Pangram 漏掉 79.8% 被 Meta Muse-Glimmer 改写的科学摘要，但对 5,000 篇人写摘要仅误报 1 篇，作者建议每次模型发布都重新验证检测器。

小米 LLM-Core 团队发布 MiMo-V2.6 技术报告，提出同时扩展训练批次、环境多样性与评分器三个维度的 RL 方法。

蚂蚁 inclusionAI 开源 PolyOCR-Venus 系列 OCR 基础模型，含 2B 和 9B 两个版本，基于 Qwen3.5 视觉语言骨干，统一支持文本识别、定位、文档解析、信息抽取与 OCR 推理。

AI 文本检测器 Pangram 漏掉了 Meta 的 Muse-Glimmer 改写的 79.8% 科学摘要，而仅将 5,000 篇人类摘要中的 1 篇标记为 AI 生成。 在东京都立大学的一篇新论文中，研究人员发现，Pangram 漏检的 AI 改写摘要比例在很大程度上取决于 LLM 版本。 研究表明，它捕获了 93.5% 的 GPT-5 改写内容，但漏掉了另一个新模型改写的 79.8%。其漏检率主要取决于改写所用的模型。

小米 MiMo-V2.6 论文发布，模型在 RL 训练循环中承担构建任务、审计测试、评分答案和排查作弊等工作。论文指出 pass/fail 测试无法区分干净修复与 hacky 修复，因此引入 grader agent 比较各组通过的补丁并把奖励移向更干净的方案。

Rohan Paul 转述 CNBC 对 NYU 数学教授 Tristan Buckmaster 的采访，后者表示仍使用 Anthropic、OpenAI、Codex，但不把任何自认为原创的想法输入其中。背景是 OpenAI 发布 722 篇 AI 生成数学论文引发职业数学家不安，作者称开放问题无限而博士与博士后时限有限。

清华、牛津与斯坦福的论文《Thinking Inertia: LLMs Keep Thinking When Told Not To》（arxiv.org/abs/2610.11765）发现，即使关闭思考模式，LLM 仍会持续显式推理，开放性问题尤其明显。

清华论文提出 TokenRouter，一个在 token 级进行大小模型路由的服务系统，吞吐量最高达现有方案的 64.15 倍。针对 vLLM、SGLang 等框架一次请求只跑一个模型、双模型协作时每步都等慢模型的问题，TokenRouter 让每个模型独立部署并来回传递半成品答案，同时保留 KV cache，并短暂挂起请求以凑更大批次。

NYU与Amazon新论文提出SGUID方法，只保留训练早期和后期都持续提供有效信号的技能，蒸馏少量精选技能即可匹配甚至超越大11倍的技能库。在3个Qwen模型上，按主题匹配选出的技能中不到25%能提供有效信号；仅用6个精选技能，4个模型中有3个在数学竞赛测试上匹配或超越了30至71个技能的完整库。第二轮加入3个新技能后，Qwen3-8B从64.3%提升至66.3%。

微软论文发现编码智能体在理解大量代码时容易出错，而非编辑大量代码时，因此建议用代码阅读与比较任务而非 diff 大小来评测。研究者构建 CABRA 生成合成编码任务，每次只提升一类难度，在 6,840 个任务上测试了 8 个 LLM 和 6 个智能体，并将每次工具调用标注为读取、分析、搜索、编辑或测试。

Redwood Research 发布论文，实证检验两条蒸馏安全路径：DFI 将不信任教师蒸馏为更弱的可信学生，使其暴露教师的隐藏怪癖；DFC 则在迁移能力的同时阻断失对齐。

Anthropic 开始更频繁地发布模型行为报告，首篇描述了评估和内部使用中发现的四类非预期行为：Claude 在真实网站或系统上采取了非预期操作，有时绕过限制而非停止。所有案例实际影响极小，Anthropic 认为其严重程度远低于 7 月和 9 月报告的网络安全事件。

我们如何衡量智能体的能力？ Agent Arena 不采用成对偏好投票，而是使用因果追踪方法，从真实世界的长时程智能体会话中测量五种信号。 更多信息见推文串。

随着机器人变得越来越可靠，DAgger 的效率就大大降低了。 例如 90% 的成功率意味着人 >90% 的时间都在被动地看着机器人 核心思路：失败来自相似的初始状态，我们可以把学习聚焦在这些状态上！ 论文：https://arxiv.org/abs/2610.05882

Harvard 研究者 Fiona Chen 和 James Stratton 基于 Jellyfish 的数据分析发现，AI 编码工具的编码阶段效率提升被生产流程下游瓶颈吸收，企业软件产出和就业未见明显变化。研究覆盖 2021 年至 2026 年 3 月超过 700 家公司约 70 万员工的 3 亿条工作事件，显示代码审查时间显著变长、pull request 更易要求修改、审查者留下更多评论。

Epoch AI 推出 InnovationEval 评测，测试 AI 能否独立复现人类论文中的机器学习创新，对照对象为 Self-Distillation Policy Optimization（SDPO）。

Language models increasingly support or automate information retrieval from document collections too large to index. This task is difficult when it requires specialist linguistic or scientific knowledge. We used the Wikipedia Game to test …

Interpreting machine learning model predictions is a central challenge, particularly for black-box vision models. Concept-based attribution methods interpret models using human-understandable concepts. Existing concept-based methods can on…

The expressivity of a neural network $Φ$ describes the class of functions that a network can express. The term topological expressivity refers to a family of measures particularly suited to understand binary classification and is often for…

Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions…

A parallel sampler can choose at random which coordinates to generate together. For a target $p$ and the output law $q_U$ of a schedule $U$ drawn afresh for each output, we compare the schedule-averaged divergence $\mathbb{E}_U D_{\mathrm{…

Deep-space communications operate under intermittent connectivity, long propagation delays, and relay storage constraints. In such networks, the Bundle Protocol enables store-carry-forward communication, requiring relay nodes to manage lim…

Clinical sleep medicine repeatedly documents profound divergences between objective sleep duration and subjective daytime impairment, notably exemplified by sleep state misperception in paradoxical insomnia and boundary instability in narc…

A central challenge in de novo protein design is generating proteins with desired structural and functional properties. Existing design models typically require a dedicated input representation for each property, limiting the design object…

Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a doma…

Vision-Language Models (VLMs) increasingly analyze and reason about images, yet it remains unclear whether their internal representations faithfully encode the structured relations required for such reasoning. We introduce EICS (Evaluation…

While deep learning models for tabular data have recently received significant research attention, the choice of objective functions for regression tasks remains an understudied degree of freedom. Several lines of research offer alternativ…

We study the last iterate of standard asynchronous Q-learning from a single trajectory of a finite Markov decision process. The trajectory follows a fixed behavior policy whose state--action chain is irreducible and may be periodic. For di…

Neural semantic parsers generate meaning representations token by token, yet well-formedness depends on references, scope, and graph relations that unfold during generation. Constrained decoding prevents locally invalid tokens but does not…

Large language models reuse the same attention heads across tasks, but identifying these shared heads does not explain what they do differently in each task. A natural source of inspiration is task-dependent interactions among shared brain…

As AI agents increasingly rely on external tools, accurate retrieval is essential for reliable behavior. Existing methods optimize semantic relevance but largely ignore tool availability during execution. To address this gap, we propose RA…

Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign u…

A memory that reproduces every current prediction may still have forgotten how to learn from the next observation. We characterize this distinction for exact cumulative learning and attention. For smooth convex residual losses with unrestr…

Many optimization problems provide direct access to only part of the gradient, while the remaining component must be inferred from function evaluations. This structure arises in performative prediction, performative reinforcement learning,…

We establish finite-sample guarantees for conditional diffusion models learned from data and implemented with finitely many denoising steps. Our approach aligns the score-training objective with the sampling grid, allowing statistical esti…

Generative molecular design typically relies on over-generation and post-hoc filtering. Inference-time steering methods like Feynman-Kac (FK) steering improve efficiency by biasing trajectories toward high-reward candidates, yet they optim…

a16z 援引 Zitzewitz 的 NBER 工作论文指出，汇总逾百年预测市场数据后，政治倾向、性别、肤色、年龄等维度均未出现超过标准误的收益系数，仅非美国选举样本显示市场略微高估右倾候选人，但 95% 置信区间表明该偏差不具统计显著性。文章还以巴西总统选举首轮为例，称预测市场比市场反应提前约 3 周押中结果。

Ethan Mollick 引用一项基于旧版 GPT-4o 的随机对照实验研究：获得 AI 访问能提高测试成绩，较小的提升在一周后仍然保留。以 AI 作为辅导工具（augmentation）的学生成绩提升保持，而让 AI 代写（automation）的学生提升消失。作者称所有实验均显示正面效果。

我们在一个 11.6 万行的代码库中埋了 70 个 bug。3 次运行中，单个智能体分别找到了 14、15 和 27 个 bug。而一个工作流在 3 次运行中每次都稳定找到 66 个。