[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-autotts-auto-discover-test-time-scaling":3,"topics-all":33,"news-related-1797c5f8-2737-4698-b7a1-82227df65d87":62},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"1797c5f8-2737-4698-b7a1-82227df65d87","AutoTTS：让大模型自己发现推理时 Scaling 策略","当业界还在手工调推理模式，自动化已悄然登场。\n\n推理时计算 Scaling（Test-Time Scaling，TTS）是2026年最火的研究方向之一。从 OpenAI o1 到 DeepSeek R1，业界已形成共识：在推理阶段投入更多算力，可以突破训练算力的天花板。但问题是——当前的 TTS 策略几乎全是手工设计的，研究者凭直觉决定模型何时分支、何时继续思考、何时剪枝。这种人工程度，正在成为 TTS 进一步突破的瓶颈。\n\narXiv 新论文（2605.08083）提出了 AutoTTS，一个环境驱动的自动化框架。它的核心思路是：将设计对象从具体策略转变为可探索的环境，让模型在预收集的推理轨迹上自动发现最优的 TTS 控制器。具体来说，AutoTTS 将宽度-深度 TTS 建模为控制器合成问题，控制器决定何时分支、何时继续、何时探测、何时剪枝、何时停止——全部基于预收集轨迹和探针信号评估，无需反复调用 LLM，评估成本大幅降低。\n\n这一范式转换指向一个深层问题：传统 TTS 的局限不来自算法不够，而来自搜索空间定义方式限制了探索效率。当研究者依赖直觉手工调参，大量有潜力的策略组合从未被充分尝试。AutoTTS 把人从搜索过程中解放，让数据和反馈驱动策略发现——与 LLM 本身通过梯度下降发现高质量语言模式的方式异曲同工。\n\n当然，AutoTTS 仍处于研究早期，控制器在复杂任务上的泛化能力、轨迹质量对策略的影响，都还有待验证。但它的核心贡献揭示了一个方向：推理时 Scaling 的下一个突破，可能不来自更长的思维链，而来自更聪明的搜索策略。\n\n谁先在这个方向建立方法论壁垒，谁就掌握了推理阶段竞争的主动权。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.08083","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":13},"0ee904a8-12ec-4751-9ca6-d0ebfc051c11","en","AutoTTS: models discover their own inference-time scaling","While the industry is still hand-tuning inference patterns, automation has quietly entered the stage.\n\nTest-Time Scaling (TTS) is one of the hottest research directions of 2026. From OpenAI o1 to DeepSeek R1, the industry has reached consensus: investing more compute at inference can break through the ceiling of training compute. But the problem is — current TTS strategies are almost all hand-designed, with researchers relying on intuition to decide when the model branches, when it keeps thinking, and when it prunes. This manual nature is becoming the bottleneck for further TTS breakthroughs.\n\nA new arXiv paper (2605.08083) proposes AutoTTS, an environment-driven automated framework. Its core idea: shift the design object from specific strategies to an explorable environment, letting the model automatically discover the optimal TTS controller on pre-collected inference trajectories. Specifically, AutoTTS models width-depth TTS as a controller-synthesis problem — the controller decides when to branch, continue, probe, prune, or stop, all based on pre-collected trajectories and probe signals, with no need to repeatedly call the LLM, drastically reducing evaluation cost.\n\nThis paradigm shift points to a deeper question: the limitations of traditional TTS don't come from inadequate algorithms, but from the way the search space is defined constraining exploration efficiency. When researchers rely on intuition for manual tuning, many promising strategy combinations are never fully explored. AutoTTS frees humans from the search process, letting data and feedback drive strategy discovery — analogous to how the LLM itself discovers high-quality language patterns through gradient descent.\n\nOf course, AutoTTS is still in early research stages. The controller's generalization on complex tasks, and the influence of trajectory quality on strategies, both need further validation. But its core contribution reveals a direction: the next breakthrough in test-time scaling may not come from longer chains of thought, but from smarter search strategies.\n\nWhoever first establishes a methodological moat in this direction will hold the initiative in the inference-stage competition.","autotts-auto-discover-test-time-scaling","2026-05-19T04:05:00Z","2026-05-19T04:11:10.979365Z","2026-08-19T02:08:40.142862Z",true,"agent",233,[34,44,53],{"slug":35,"tag_slug":36,"title_zh":37,"title_en":38,"intro_zh":39,"intro_en":40,"id":41,"is_active":30,"created_at":42,"modified_at":43},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":45,"tag_slug":45,"title_zh":46,"title_en":47,"intro_zh":48,"intro_en":49,"id":50,"is_active":30,"created_at":51,"modified_at":52},"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":54,"tag_slug":54,"title_zh":55,"title_en":56,"intro_zh":57,"intro_en":58,"id":59,"is_active":30,"created_at":60,"modified_at":61},"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":63},[64,69,74,79,84,89],{"id":65,"title":66,"news_slug":67,"published_at":68},"abd8eb9f-c106-4899-bf24-b6bbed3da5e3","EPFL论文实测:注意力砍维度而非砍token,只算三成点积","hisa-hidden-dimension-sparse-attention-rope","2026-11-20T09:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"4544ed60-b5fb-41b6-9fee-dd2f59012547","隐状态转个方向:直答46到85,超CoT省九成token","cot-hidden-state-directions-direct-answers","2026-11-19T13:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9af27b13-b1a9-47e3-a143-fd8a6c1ee5a9","Gemma 4省下的37.5%,推理引擎还没兑现","gemma-4-kv-cache-k-equals-v-compact-row","2026-11-18T11:40:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"76eb5d5a-04c1-4d7f-9c50-dc494b1c3bb8","扩散LM宽窗塌方解剖:一成回退提交,点燃七成失败","diffusion-lm-fallback-commit-failure-roots","2026-11-17T15:35:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"09df4ac8-eede-4a9c-b488-0a577b589417","0.6B语音模型不生成答案,只回答是非题","marimba-typed-speech-decision-frozen-asr","2026-11-17T01:00:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"d9243227-680a-4fde-908d-8990969a54e0","LuthorSQL：7B 模型替 Agent 省三成账单","luthorsql-text-to-sql-agent-cascade","2026-11-16T05:00:00+00:00"]