At a glance Harnessed Agentic RL: Microsoft Research Asia introduces a training paradigm in which the same agent harness…

Agentic reinforcement learning (RL) disaggregates training from rollout, so each policy update must reach the rollout cl…

Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, G…

Training capable coding agents via reinforcement learning (RL) requires diverse tasks with reliable verifiers. Open-sour…

We demonstrate that training LLMs with RL does not improve performance equally across a dataset. RL shows large improvem…

Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-traini…

给已有 Agent 框架做 RL 后训练，通常要把业务逻辑重写成训练代码。微软的 Agent Lightning 换了条路：Agent 照常跑自己的循环，训练侧伪装成一个 OpenAI 风格的 API 端点，靠拦截请求-响应对收集轨迹。v1.0 技术报告里，Qwen3.5-9B 编码 Agent 在 SWE-bench Verified 上从 41.8% 提到 56.4%。本文拆解它的架构、算法与社区争议。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Domain adaptation aims to turn a general-purpose large language model (LLM) into an expert for a target domain. While th…

Many useful language-model tasks cannot be evaluated by exact outcome verification. Rubric-based reinforcement learning …

文婷* 2026-09-29 15:59:46 来源：量子位 文婷 发自 凹非寺 量子位 | 公众号QbitAI 别眨眼。** 星空下，一辆白橙色的反重力赛车如利刃出鞘，猛地切进弯道，丝滑拐弯，在赛道上拖拽出一道道冰蓝光带，溅起金黄色的摩…

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the …

Tool-calling agents produce heterogeneous outputs, interleaving structured tool invocations with user-facing natural lan…

A pretrained robot foundation policy may execute most of a long-horizon task yet repeatedly fail at a few critical subta…

Pretrained vision-language-action (VLA) models enable broad manipulation but remain unreliable in tasks demanding precis…

Agent trajectories record what an agent does and what happens next. Yet standard supervised fine-tuning (SFT) applies lo…

OpenAI can disclose misalignment before fixes exist. Its 6 initial reports include fabricated data and leaked API keys. …

Google Research has introduced Retrieve-for-Train (R4T), a framework for search that returns coherent, diverse result se…

Reinforcement learning with verifiable rewards is typically performed on-policy, keeping training data close to the curr…

Grounding natural-language instructions into reliable and executable actions remains a fundamental challenge for vision-…

Data policies for reinforcement learning with verifiable rewards (RLVR) determine which rollouts are used, how strongly …

Online reinforcement learning (RL) fine-tunes diffusion models with black-box rewards, but reverse-process methods such …

Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvem…

Google 比 OpenAI 早七周发布 Deep Research，但把「研究计划让用户改批」产品化的也是它——agentic 检索的通用循环是：规划、迭代检索、阅读、反思补漏、交叉验证、带引用报告。本文拆解这个循环的两种实现路线（端到端 RL vs 显式编排），用 BrowseComp 上「裸模型不足 10% vs deep research 51.5%」的差距说明多步浏览行为本身值多少分，也把「慢不等于对」的引用可靠性研究摆上台面。

精读 DeepSeek-R1 论文（arXiv 2501.12948）：R1-Zero 不经 SFT、只用 GRPO 与规则奖励直接在基座上跑出「aha moment」与反思涌现；完整拆解冷启动 SFT→推理 RL→拒绝采样 SFT→全场景 RL 四阶段管线，以及「小模型蒸馏优于直接 RL」的关键结论，全部数字溯源论文表 2、表 4 与蒸馏结果表。

Reinforcement learning (RL) for post-training large language models (LLMs) incurs substantial computation and memory ove…

Reward post-training of flow models either reweights the model's own samples under a KL penalty or a frozen reference, o…

Agent 工程系统学习第 13 章（前沿专题）：三个高薪高门槛方向。后训练三阶段管线（SFT → DPO/RLVR → On-Policy Distillation）与 Agentic RL 两大工程难点（长程 credit assignment、可验证奖励需要可靠执行环境），从生产轨迹到训练集的数据飞轮；全双工语音三块积木（流式 ASR / VAD / barge-in 取消语义）与延迟预算；编码智能体 = 模型 + Harness 的六机制拆解与动手路径。

Prime Intellect 发布 Prime Inference 推理服务平台，提供无服务器端点和预留 GPU 容量，公开前内部日均处理近万亿 token，主要来自 RL rollout、合成数据生成、评测和长时编码智能体，补全其开源训练栈的服务环节。

Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneousl…

Power-sharpened sampling is an inference-time alternative to reinforcement-learning (RL) post-training for enhancing rea…

Reinforcement learning (RL) has become a key paradigm for enhancing the reasoning of large language models, yet the high…

Reinforcement learning (RL) post-training for large language models (LLMs) coordinates multiple models across generation…

Recent progress in frontier models has renewed interest in large-scale reinforcement learning (RL), but the cost of gene…

Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed ou…

Reinforcement learning (RL) has become a key technique for improving the reasoning and agentic abilities of large langua…

Reinforcement learning (RL) is vital for optimizing video generation models, with a robust reward model (RM) serving as …

Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivate…

Training prompts in online reinforcement learning (RL) differ substantially in how informative they are for the current …

Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixt…

把 2026 年可核查的公开证据整理成一张六层能力路线图——预训练、后训练 RL、推理时计算、上下文与记忆、智能体与 Harness、世界模型。含 Meta ScaleRL 40 万 GPU 小时实验结论、RL 预算占比 10%–30% 口径、Chinchilla 对比、Meta-Harness 6x 差距等数据锚点，并给出优先级表与算法工程师/产品经理的行动建议。

解读 R1 在强化学习范式下的推理能力涌现，及其对开源生态的影响

At a glance StudentSim is a research system for evaluating and improving AI tutors before they are tested with real lear…

文婷* 2026-10-11 15:52:16 来源：量子位 文婷 发自 凹非寺 量子位 | 公众号QbitAI 数学2.0时代已至。** 刚刚，数学界的领军人物、菲尔兹奖得主陶哲轩正式发文： AI正在让数学证明变得泛滥，建立在证明稀缺之…

文婷* 2026-10-10 16:54:08 来源：量子位 文婷 发自 凹非寺 量子位 | 公众号 QbitAI 短短24天，融资额膨胀了近**22倍**，估值翻了**37.5倍**。 Jev掀起的风暴，已经变成真金白银了！ 就在今天，…

文婷* 2026-10-10 16:52:16 来源：量子位 文婷 发自 凹非寺 量子位 | 公众号 QbitAI 果然夜没白熬，Google家的大香蕉终于更新了！ 凌晨，Google突然发布了新一代生图模型**Nano Banana 2…

JetBrains 推出面向编码智能体的开源模型 Mellum2.1，总参数 12B、每 token 激活 2.5B，主要通过在真实软件环境中的强化学习带来提升，以 Apache 2.0 协议在 Hugging Face 发布，可自托管运行。

2026 年 9 月底 gVisor 被捐赠给 CNCF，官方安全档案里约 110 个内核高危漏洞 96% 被直接防御、零个被攻破。本文拆解它的 Sentry/Gofer 三进程模型、systrap 系统调用拦截、directfs 文件代理与多层防御设计，澄清 Cloud Run 与 Kubernetes agent-sandbox 的两处常见误传，并给出它与 microVM 的选型对比：谁该走哪条路线。

2024 年 9 月 o1 用「先想一会儿」在 AIME 上打出 83% 对 13%，2025 年 1 月 DeepSeek-R1 把整套方法开源复现，到 2026 年 10 月，推理档位已成各大模型 API 的标准旋钮。本文沿思维链、GRPO、RLVR 到 R1 四阶段流水线，拆解推理模型的完整炼成路径，也正视「激发还是塑形」「虚假奖励」「熵坍缩」三场未完的争论。

从 Attention Is All You Need 的 encoder-decoder 原型，到 MLA、MoE、iRoPE 缀满一身的 2026 旗舰，本文系统拆解基础 Transformer 的架构原理，梳理八年来 KV 压缩、位置编码、稀疏专家三条演进主线，并给出 DeepSeek、Llama、Qwen、Gemini 等旗舰架构的横向对比地图。