[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-xpeng-x-aut-audio-encoder-pruning":3,"topics-all":38,"news-related-108af093-b226-4372-9cf0-77323ffc5456":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},"108af093-b226-4372-9cf0-77323ffc5456","小鹏 X-AuT 给语音大模型剪枝:音频塔砍 4 层,车载推理提速 21.4%","小鹏汽车开源 X-AuT 压缩框架:把 Qwen3-ASR-0.6B 的音频编码器从 18 层渐进剪到 14 层,参数省 20.7%,16 层版错误率反而降 6.1%,车载 PPU 编码器延迟降 21.4%,权重与推理代码已放出。","语音大模型上最贵的部分,未必是那个会说话的大脑,而是那双\"耳朵\"。小鹏汽车 XPeng Omni 团队最近开源的 X-AuT 框架,给出的答案是:把耳朵剪小一点,大脑保持冻结,识别精度还能不降反升。\n\n## 剪的是音频塔,不是语言模型\n\nX-AuT 的操作对象是 Qwen3-ASR-0.6B 这类语音大模型里的音频编码器(audio tower)——负责把原始波形变成语言模型能消化的嵌入表示。团队把它从 18 层 Transformer 块渐进式剪到 14 层:先删第 1、18 层得到 16 层版,再删第 5、6 层得到 14 层版,音频塔参数从 1.86 亿降到 1.48 亿,省了 20.7%。整个过程语言模型主干完全冻结,只在注意力投影上加 LoRA 适配器,输出嵌入与 lm_head 绑定共享。\n\n## 关键数字:两层精度不降反升的窗口\n\n在十个公开中英文基准上,结果是两档实用的工程甜点:16 层版把宏平均错误率从 5.61% 压到 5.27%,相对降 6.1%;14 层版错误率 5.75%,相对只涨 2.5%,换来 20.7% 的参数削减。推理效率方面,14 层编码器在车载 PPU 上延迟降 21.4%,H800 上降 11.4%;端到端延迟分别降 4.7% 和 2.6%——幅度有限,因为未被剪枝的 28 层文本解码器占了大头。\n\n## 为什么渐进剪枝比一刀切好\n\n直接把 18 层砍到 14 层,错误率会飙到 6.73%;X-AuT 分两步走、配合恢复训练,同样 14 层只要 5.75%,差出近 1 个百分点。恢复手段是四件套:行为探针选层、表征对齐、跨尺度蒸馏、LoRA 微调。其中跨尺度蒸馏值得单说——用一个冻结的 Qwen3-ASR-1.7B 当教师,通过学到的 2048→1024 投影监督 0.6B 学生,蒸馏完教师就丢弃。同样配方下,跨尺度教师把学生错误率带到 5.55%,同尺度自蒸馏只有 8.45%,差距悬殊。\n\n## 端侧落地的算力账\n\n这套数字背后是车载场景的现实约束:车机芯片上每一毫秒延迟都直接体验相关。训练数据侧,团队用一个超 28 万小时、按转录一致性分九档置信级的源池,报告的运行全部用第一档数据。但也要泼冷水:基准值都是单次运行的最佳 checkpoint,论文自己写明没有重复种子和置信区间;14 层版在 Fleurs-en 上最大退化 0.93 个百分点,十个基准里有八个仍高于基线。开源范围也有限——放出的是 14 层 checkpoint 推理脚本和最小 LoRA 微调示例,行为探针管线、跨尺度蒸馏代码和 28 万小时数据管线都不在发布范围,许可证是 CC BY-NC 4.0,禁商用。\n\n对做端侧语音的团队,这份工作的价值不在\"又一个压缩方法\",而在它把\"剪音频塔、冻语言模型\"这条路径的精度-延迟边界画清楚了:两层以内的渐进剪枝,精度损失可以靠恢复训练抹平甚至反超;再往下剪,收益和风险就开始赛跑。当语音交互从手机助手走向车载、耳机等算力受限场景,这种给\"耳朵\"减负的精细活,会比无脑堆参数更接近落地。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.11412","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"5c8c6f85-c5ad-4174-af01-ce069be1f34a","en","XPeng's X-AuT Prunes Speech LLM Audio Tower for In-Vehicle Speed","XPeng's X-AuT prunes the Qwen3-ASR-0.6B audio encoder from 18 to 14 layers, cutting 20.7% of parameters and 21.4% in-vehicle latency; weights open.","The most expensive part of a speech LLM may not be the talking brain but the listening ears. XPeng's Omni team just open-sourced X-AuT, a compression framework whose answer is: shrink the ears, keep the brain frozen, and accuracy can even improve.\n\n## What Gets Pruned\n\nX-AuT operates on the audio tower of speech LLMs like Qwen3-ASR-0.6B — the encoder that turns raw waveforms into embeddings the language model can consume. The team progressively pruned it from 18 Transformer blocks to 14: first removing layers 1 and 18 to get a 16-layer version, then removing layers 5 and 6 to reach 14. Audio-tower parameters dropped from 186M to 148M, a 20.7% cut. Throughout, the language-model backbone stays completely frozen; only LoRA adapters are added to attention projections, with the output embedding tied to lm_head.\n\n## The Numbers: A Two-Step Sweet Spot\n\nAcross ten public Chinese-English benchmarks, the results land at two practical operating points. The 16-layer variant pushes macro-average error from 5.61% down to 5.27%, a 6.1% relative reduction. The 14-layer variant sits at 5.75% error — just 2.5% relatively higher — in exchange for the 20.7% parameter cut. On inference efficiency, the 14-layer encoder cuts latency by 21.4% on the in-vehicle PPU and 11.4% on H800; end-to-end latency improves 4.7% and 2.6% respectively — modest, because the unpruned 28-layer text decoder dominates total inference time.\n\n## Why Progressive Beats One-Shot\n\nPruning 18 layers straight down to 14 pushes error to 6.73%; X-AuT's two-step approach with recovery training reaches 5.75% at the same depth — nearly a full point better. The recovery toolkit has four parts: behavior-driven layer probes, representation alignment, cross-scale distillation, and LoRA finetuning. The cross-scale distillation deserves a closer look: a frozen Qwen3-ASR-1.7B teacher supervises the 0.6B student through learned 2048-to-1024 projections, then is discarded after training. Under the matched recipe, the cross-scale teacher brings the student to 5.55% error versus 8.45% for same-scale self-distillation — a wide gap.\n\n## The On-Device Math\n\nBehind these numbers is the reality of automotive compute: every millisecond of latency on a vehicle chip directly shapes user experience. On the data side, the team used a 280k-plus-hour source pool ranked into nine confidence classes by transcript consistency; all reported runs used class-1 data only. But some cold water is due: benchmark values are single-run best checkpoints with no repeated seeds or confidence intervals, as the paper itself states. The 14-layer model's largest degradation is 0.93 percentage points on Fleurs-en, and eight of ten benchmarks remain above baseline. The release scope is also limited — the repo ships the 14-layer checkpoint, a standalone inference script, and a minimal LoRA finetuning example, while the behavior-probe pipeline, distillation code, and the 280k-hour data pipeline stay private. The license is CC BY-NC 4.0, prohibiting commercial use.\n\nFor teams building on-device speech, the value here is not \"yet another compression method\" but a clearly mapped accuracy-latency frontier for the prune-the-audio-tower path: within two layers of progressive pruning, recovery training can erase or even reverse the accuracy loss; push deeper, and returns start racing risks. As voice interaction moves from phone assistants to vehicles, earbuds, and other compute-constrained scenes, this kind of careful slimming of the ears will land closer to production than blind parameter scaling.","xpeng-x-aut-audio-encoder-pruning","2026-09-12T19:06:47Z","2026-09-12T19:06:52.967158Z","2026-09-12T19:06:52.967167Z",true,"agent",58,[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},"fb97a60d-69a1-4988-8de6-d1540ba63359","2.4B 参数读懂整页 A4:Cohere Labs 把最小的多模态模型挂上了 Apache 2.0","cohere-north-micro-vision-open-vlm","2026-08-18T13:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4244f57a-3afa-465c-aa67-793df6eba5cc","LFM2.5-VL-3B 开源：3.1B 参数让手机读懂屏幕、框住物体、自己调工具","liquid-ai-lfm2-5-vl-3b-edge-vlm","2026-08-14T13:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d331d2b8-94ac-43c1-b53e-d4cb416a08f2","Liquid AI LFM2.5-VL-450M：450M 参数的边缘 VLM，把「结构化视觉」拉进亚秒级","lfm-2-5-vl-450m-liquid-edge-sub-second","2026-06-14T04:14:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"30fca629-bace-4832-9789-b44aa8c8989d","学生团队从零训出开源 7B 模型 ZGCM-1:数学推理硬刚 235B 前沿","zgcm-1-open-7b-foundation-model","2026-09-15T19:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"51c13e24-8072-404c-a8d4-75c40cff05ee","Ling-3.0-flash-VL 开源：124B MoE 只激活 5.5B，视觉塞进 Agent 闭环","ling-3-0-flash-vl-open-weights","2026-09-15T13:18:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"550cee5e-18e8-4236-9304-7207ebc221a8","Agnes 3.0 Flash 开源:72 层仅 18 层带 KV 缓存,33B 单卡跑 262k 上下文","agnes-3-0-flash-preview-open-weights","2026-09-13T15:20:00+00:00"]