[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mihoyo-gamer-vla-game-agent":3,"topics-all":38,"news-related-cd155bec-f69f-42e8-aa29-fd3bd2218669":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"cd155bec-f69f-42e8-aa29-fd3bd2218669","米哈游8B模型打星铁:支线83.6%,小游戏49.7%","米哈游Honkai AI团队在arXiv发布Gamer：用数万小时真实键鼠轨迹继续预训练Qwen3-VL-8B，再加稀疏思考微调与在线强化学习，模型只看屏幕、只按键鼠，在星铁未见过的支线拿到83.6%，七类小游戏宏平均49.7%，同表API模型落在2%到17.1%；自研流式引擎把单步延迟从1306毫秒压到65毫秒。","11 月 13 日挂上 arXiv 的 2611.18397，作者栏不是常见的大模型厂商，而是米哈游 Honkai AI R&D Team，合作机构包括新加坡国立大学、清华大学、中国科学院大学、中国科学技术大学和西安交通大学。他们交出的模型叫 Gamer，定位是通用视觉-语言-动作（VLA）基础模型，专攻长时程、实时的游戏智能体：不读内存、不调专用接口，只看屏幕像素、输出原生键鼠操作，像人一样坐在屏幕前打游戏。\n\n## 三段式训练：先看录像，再学指令，最后自己试错\n\nGamer 的底座是 Qwen3-VL-8B，训练分三步。第一步继续预训练（CPT），语料是数万小时的设备级键鼠轨迹，其中《崩坏：星穹铁道》录音轨道约有 68K 场、11K 小时原始录像，过滤掉损坏、错位和无意义片段后，约 3.2K 场、1K 小时进入标注环节；星铁部分最终对齐出约 149K 段剪辑、约 332 小时进入 CPT。游戏轨迹占训练 token 的 85%，另外 15% 是通用多模态数据。团队还跑了 120 组受控对照实验（4 个模型规模 × 6 档游戏时长预算 × 5 种数据配比）。动作接口每 200ms 采样一次观察（5Hz），每段切成 6 个键鼠子块，等效约 30Hz 的控制频率。\n\n第二步是指令微调（SFT）：由 GPT-5.6-Luna 在关键决策点标注稀疏的 Think 推理、任务指令与记忆摘要，把操作、子目标、完整任务三个层级缝进同一条轨迹。第三步是 TA-GRPO 在线强化学习：64 张 H200 同时伺服 32 个并行游戏环境，每轮收集 128 条 rollout，奖励全部来自游戏环境侧的可验证信号，没有训练额外的过程奖励模型。\n\n## 成绩单：支线 83.6%，七类小游戏宏平均 49.7%\n\n在两条训练时没见过的星铁支线上，整体准确率的变化是：底座 0.0% → CPT 34.4% → SFT 72.2% → RL 80.0%（Opportunity Makes the Thief），以及 13.9% → 47.9% → 69.1% → 83.6%（Bestial Ferocity）。40 个 held-out 小游戏实例、七类玩法的宏平均则是 3.2% → 15.0% → 39.2% → 49.7%。同一张表里，一排 API 前沿模型的成绩是：GPT-6.1-Sol 17.1%、GPT-5.6-Sol 12.0%、GPT-5.6-Luna 11.5%、Claude-Opus-5.0 9.9%、GLM-5.3-Flash 8.2%、Gemini-3.8-Flash 和 DeepSeek-V4-Flash-Vision 各 2.0%——以上均为论文自报口径。跨游戏方向，8B CPT 策略在四个公开游戏基准上的平均实时分为 29.5，Qwen3.5-35B-A3B MoE 版为 31.5，论文称两者都超过了表内任何一个通用基线。\n\n## 隐藏主角：把 1.31 秒压到 65 毫秒的推理引擎\n\n实时游戏对推理延迟极其敏感，这部分由自研的 Streaming-vLLM 扛住：跨轮次保留 KV 状态并做记忆压缩，用有限状态机校验结构化动作 token，提前流式派发已验证的动作块。在 H200 上跑 Qwen3.5-35B-A3B MoE，端到端延迟从 1306ms 一路压到 65ms，论文称之为 20.1 倍加速，步进率从 0.77Hz 提到 15.36Hz。这套引擎直接反过来影响训练：把 rollout 换成标准 vLLM 的 0.6-FPS 慢速控制后，八个导航任务的奖励曲线在学习后期出现崩塌。流式视频理解方面，Streaming-vLLM 在 StreamingBench 拿到 83.29（对照 DSCache 为 83.21），查询延迟从 0.53s 降到 0.08s。\n\n## 这条路不止米哈游在走\n\n用人类键鼠轨迹加 scaling 练游戏智能体，不是孤例。Game-TARS（arXiv 2510.23691，2025 年 10 月）用统一的键鼠动作空间预训练了超过 500B token；2026 年 1 月的另一篇（arXiv 2601.04575）干脆开源了 8300+ 小时高质量人类游戏轨迹和完整训练配方，模型在消费级 GPU 上实时推理，成绩与人类玩家相当。不同团队、不同配方，在同一件事上各自踩通了。\n\n## 泼点冷水\n\n论文自己列了三条限制：评测覆盖的环境和任务结构仍有限；serving 成本还可以继续降；深度推理仍是瓶颈，作者设想用 System 2 慢思考加 System 1 快执行的分层架构来补。远迁移也没那么神，论文承认在更远的 GameWorld 上 CPT 检查点提升有限；RL 同样不是全面提升——七类小游戏里 Janus 的迷宫从 SFT 的 52.4% 反而掉到 RL 后的 23.8%。另外，GitHub 仓库已经上线，但 Streaming-vLLM 引擎还躺在 TODO 列表里尚未放出，论文称将开源的 CPT 检查点也还没见到实物。「操作效率与熟练玩家相当」这句话，目前同样只是论文自述口径。\n\n游戏是天然自带可验证奖励的环境，而游戏公司自己最清楚哪里需要成千上万次重复测试。当米哈游开始训自己的模型，最先落地的大概率不是陪你玩，而是自动化 QA 和回归测试。键盘鼠标，正在成为具身智能之外另一条通用的「动作词表」——下一个下场的游戏公司，会是谁？\n\n论文：https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18397 ｜ 项目页：https:\u002F\u002Fmihoyo-gamer.github.io\u002F ｜ 代码仓库：https:\u002F\u002Fgithub.com\u002FRainBowLuoCS\u002FGamer","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18397","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d6ed6394-9605-4cf9-b761-2022a5d9a7b5","en","miHoYo Trains an 8B Model to Play Star Rail in Real Time","miHoYo trains an 8B VLA agent on tens of thousands of hours of keyboard-mouse play: 83.6% on held-out Star Rail quests, 49.7% on minigames","On November 13, a paper landed on arXiv under the ID 2611.18397, and the author block is not a familiar LLM vendor: it is the miHoYo Honkai AI R&D Team, with collaborators from the National University of Singapore, Tsinghua University, the University of Chinese Academy of Sciences, USTC, and Xi'an Jiaotong University. The model is called Gamer, a generalist vision-language-action (VLA) foundation model aimed at long-horizon, real-time game agents. It reads no game memory and calls no special API: it watches screen pixels and emits native keyboard-mouse actions, exactly like a human sitting in front of the monitor.\n\n## Three-stage training: watch recordings, learn instructions, then learn by doing\n\nGamer starts from Qwen3-VL-8B and trains in three stages. Stage one is continued pre-training (CPT) on tens of thousands of hours of device-level keyboard-mouse trajectories. The Honkai: Star Rail recording track alone holds roughly 68K sessions and 11K hours of raw video; after filtering corrupted, misaligned, or inactive segments, about 3.2K sessions and 1K hours survive for annotation, and the HSR portion of the CPT corpus ends up as roughly 149K aligned clips covering around 332 hours. Gameplay trajectories account for 85% of training tokens, with the remaining 15% from general multimodal data. The team also ran a controlled 120-configuration study (4 backbone sizes x 6 gameplay budgets x 5 data mixtures). The action interface samples one observation every 200 ms (5 Hz) and splits each interval into six keyboard-mouse sub-blocks, an effective control rate of roughly 30 Hz.\n\nStage two is supervised fine-tuning: GPT-5.6-Luna annotates sparse Think steps at key decision points, plus task instructions and memory summaries, stitching three instruction levels (local operations, subgoals, complete tasks) into the same trajectories. Stage three is TA-GRPO online reinforcement learning: 64 NVIDIA H200 GPUs serve 32 parallel game environments, each iteration collects 128 rollouts, and rewards come from environment-side verification, with no learned process reward model.\n\n## The scoreboard: 83.6% on side quests, 49.7% across seven minigame types\n\nOn two held-out Star Rail side quests, overall accuracy moves from 0.0% for the base model to 34.4% after CPT, 72.2% after SFT, and 80.0% after RL on Opportunity Makes the Thief, and from 13.9% to 47.9%, 69.1%, and 83.6% on Bestial Ferocity. Across 40 held-out minigame instances spanning seven gameplay types, the macro-average climbs from 3.2% for the base model to 15.0% for Gamer-CPT, 39.2% for Gamer-SFT, and 49.7% for Gamer-RL. In the same table, a row of API frontier models lands at GPT-6.1-Sol 17.1%, GPT-5.6-Sol 12.0%, GPT-5.6-Luna 11.5%, Claude-Opus-5.0 9.9%, GLM-5.3-Flash 8.2%, and 2.0% each for Gemini-3.8-Flash and DeepSeek-V4-Flash-Vision — all self-reported numbers from the paper. On four public game benchmarks, the 8B CPT policy averages a real-time score of 29.5 and the Qwen3.5-35B-A3B MoE variant reaches 31.5; the paper states that both exceed every general-purpose baseline listed in that table.\n\n## The hidden protagonist: an engine that cuts 1.31 seconds to 65 ms\n\nReal-time games are brutally sensitive to inference latency, and this is where the in-house Streaming-vLLM engine earns its keep: it retains KV states across turns under memory compaction, verifies structural action tokens with a finite-state machine, and dispatches verified action blocks early. On an H200 running the Qwen3.5-35B-A3B MoE model, end-to-end latency drops from 1306 ms to 65 ms, which the paper calls a 20.1x speedup, while the step rate rises from 0.77 Hz to 15.36 Hz. The engine feeds straight back into training: when rollouts switch to standard vLLM control at 0.6 FPS, reward curves across eight navigation tasks collapse in later training stages. On streaming video understanding, Streaming-vLLM scores 83.29 on StreamingBench versus 83.21 for DSCache, with query latency falling from 0.53 s to 0.08 s.\n\n## miHoYo is not alone on this road\n\nTraining game agents on human keyboard-mouse trajectories plus scaling is not a one-off. Game-TARS (arXiv 2510.23691, October 2025) pre-trained on over 500B tokens with a unified keyboard-mouse action space. Another paper from January 2026 (arXiv 2601.04575) open-sourced 8300+ hours of high-quality human gameplay plus a full training recipe, with a model that runs real-time inference on a consumer GPU and performs competitively with human players. Different teams, different recipes, each independently landing on the same answer.\n\n## A few buckets of cold water\n\nThe paper lists three limitations of its own: evaluation still covers a selected set of environments; serving costs can drop further; and deep reasoning remains a bottleneck, with the authors sketching a layered System 2 plus System 1 architecture as a remedy. Distant transfer is not magic — the paper concedes that gains on the more distant GameWorld benchmark remain limited for CPT checkpoints. RL is not uniformly positive either: on Janus's Maze, one of the seven minigame types, success falls from 52.4% after SFT to 23.8% after RL. And while the GitHub repository is live, the Streaming-vLLM engine still sits on the repository TODO list, and the CPT checkpoints the paper says will be released are not yet visible. The claim of operational efficiency comparable to skilled players likewise remains the paper's own framing, not an independent measurement.\n\nGames are environments with verifiable rewards built in, and no one knows better than a game company where thousands of repetitive tests are needed. When miHoYo starts training its own models, the first landing spot is probably automated QA and regression testing, not playing alongside you. The keyboard and mouse are becoming another universal action vocabulary alongside embodied AI — which game company steps in next?\n\nPaper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18397 | Project: https:\u002F\u002Fmihoyo-gamer.github.io\u002F | Code: https:\u002F\u002Fgithub.com\u002FRainBowLuoCS\u002FGamer","mihoyo-gamer-vla-game-agent","2026-11-16T15:40:00Z","2026-11-16T15:20:08.629463Z","2026-11-16T15:20:08.629471Z",true,"agent",227,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"950e298b-7e72-443b-939e-ef8ba07ee677","多模态社工攻破7个智能体:验证越多被骗越狠","multimodal-sea-social-engineering-agent-security-benchmark","2026-11-08T09:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bff1ec03-4e40-4d63-be35-82da33948969","图像生成的长上下文考试:32张参考图拼一张图","longbanana-multi-reference-image-synthesis-benchmark","2026-11-05T17:05:54+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"643e4f3e-2ec0-47eb-b732-4d8da1b18e70","视频模型该学会追问:离线52.18%,流式只剩7.22%","askstream-proactive-asking-streaming-video","2026-11-02T01:35:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"fd0e593c-bb2c-44f9-8913-16f8dcef6999","苹果8B智能体:真实iOS上追平Gemini-3.1-Pro","apple-online-rl-ui-agent-live-ios","2026-10-31T11:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9109e6b0-bf85-472d-abf6-ba8e31dc17c5","Co-Design 基准:七成准则过,26.4% 完整交付","co-design-bench-agent-delivery-gap","2026-10-30T05:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"db8f2b94-8565-49fe-9fd8-08310f441161","HiSTEVE 1.0:像人一样玩我的世界,只靠屏幕和键鼠","histeve-hierarchical-vla-minecraft-agent","2026-10-26T15:13:44+00:00"]