Google Research Blog

Google Research Blog 是 Google Research 的官方技術部落格，發布 Google 在機器學習、自然語言處理、AI 系統架構等領域的最新研究成果，也涵蓋 agent memory、RAG 與 model governance 類主題。

基本資訊

• 類型：官方研究部落格
• 營運方：Google Research / Google DeepMind
• 網站：https://research.google/blog/
• 語言：英文

相關頁面

• empirical-research-assistance — ERA 與其 follow-up 應用案例
• academic-workflow-agents — PaperVizAgent / ScholarPeer 的學術工作流自動化
• chain-of-evidence-autonomous-research — Chain-of-Evidence、Science One 與研究 artifact 的可驗證性
• test-time-compute-evaluation — TTD-DR 的 draft-first、retrieval denoising 與 test-time budget 評估
• agentic-rag-cross-corpus — Google 提出的 Agentic RAG 架構（Sufficient Context Agent + Cross-Corpus Retrieval）
• machine-unlearning-audit — Google Research 的 machine unlearning 稽核方法
• reasoningbank-agent-memory — ReasoningBank：把成功與失敗經驗壓成可重用的 agent memory
• reasoning-for-factual-recall — reasoning trace 透過 computational buffer / factual priming 擴大 factual recall 邊界
• knowledge-profiling-factuality — 用 WikiProfile 拆解 encoding、recall、recognition 與 thinking recovery
• frozen-multi-token-prediction — Gemini Nano frozen MTP 手機端推論加速
• groundsource-news-to-data — Gemini 把多語新聞轉成結構化事件資料的 Groundsource 方法
• tabular-foundation-models — TabFM 與表格資料的 zero-shot foundation model 工作流
• time-series-foundation-models — TimesFM-3 與多變量時間序列、covariate、quantile forecast
• wearable-health-foundation-models — SensorFM 與 wearable sensor foundation model、agentic classroom、Personal Health Agent grounding
• world-models — 從靜態標籤與文字表徵延伸到 physical world 的狀態、功能與動態節律
• eval-is-spec — SymptomAI 的真實使用者、主動追問與延遲 ground truth 評測案例
• data-science-agents — DS-STAR 與資料科學 agent 的異質資料 / verifier loop 模式
• planetary-prediction-engine — PPE：LLM 編排 geospatial data selection、curation、AutoML 與 prediction
• test-time-memorization — Titans + MIRAS 的模型層長期記憶與 surprise-based update
• nested-learning-continual-learning — Nested Learning 的 continual learning / nested optimization 設計
• bayesian-teaching-llm-reasoning — 用 Bayesian assistant traces 訓練 LLM 的偏好推理與信念更新
• on-device-intent-extraction — 用小型 multimodal model 從 UI 操作軌跡抽取使用者意圖
• generative-ui — 模型依 prompt 生成完整互動介面的產品形態
• diffusion-model-creativity — score smoothing、data manifold 與 diffusion model 的生成新穎性
• sequential-attention-subset-selection — 用 attention score 做大型模型 subset selection 與 structured pruning
• privacy-preserving-chatbot-analytics — 用 differential privacy 做 chatbot 使用趨勢與 trace 聚合分析
• gist-smart-sampling — 用 Greedy Independent Set Thresholding 做訓練資料 smart sampling
• multilingual-model-scaling-laws — ATLAS 多語言模型 scaling laws 與 cross-lingual transfer 判準
• google-ai — Google 官方 X 帳號與相關發布來源

2026-09-29 新文：Diffusion Controller

Google Research 介紹 Diffusion Controller（DiffCon）：把 reverse diffusion denoising 視為連續控制問題，在 frozen image-generation backbone 上接輕量 side network，透過 policy-gradient／PPO 或 reward-weighted regression 做 prompt／preference alignment。文章報告 Stable Diffusion v1.4、HPS-v2 與 white-box／gray-box 比較結果；90% win rate、LoRA 對照與「不破壞穩定性」均保留為 Google Research／arXiv 研究 attribution，本輪未讀取完整論文、實驗表格、模型權重或重跑 benchmark。相關 raw evidence：[[raw/articles/google-research-diffusion-controller-2026-09-29]]；可與 diffusion-model-creativity、generative-media-agent-workflow 交叉閱讀。本輪只新增 raw evidence，不更新 compiled concept。

2026-09-24 新文：Automating coherent long-form video generation

Google Research 介紹 AI video co-director、CANVAS、A²RD 與 VQQA 四個互相連接的研究框架，把長篇影片生成拆成全域創意搜尋、persistent visual／video memory、segment-level retrieve-synthesize-refine-update 與 VLM closed-loop selection。文章把目標放在降低 semantic drift、cascading failures、feature drift 與 content collapse，並以 GenAD-Bench、HardContinuityBench、VBench-Long 等 benchmark 報告研究框架下的改善；AI Ark 只保留 Google Research 的來源陳述，未讀取連結的 ArXiv paper、benchmark、code 或模型，未做獨立重跑。相關 raw evidence：[[raw/articles/google-research-coherent-long-form-video-generation-2026-09-24]]；可與 generative-media-agent-workflow、dynamic-agent-workflows、agent-trace-observability 與 eval-is-spec 交叉閱讀。

2026-09-15 新文：Retrieve-for-Train（R4T）

Google Research 發布 R4T（Retrieve-for-Train），一個以 RL 作為一次性 objective transducer 的 set-valued retrieval framework：先以 Soft-GRPO + composite set-level reward（groundedness、Vendi Score diversity、alignment 三項互為 counter-anchor）訓練 Gemma3-4B／Qwen3-4B fan-out LLM，離線合成 (query → target-set) supervision pairs，再 distill 成 53.9M-parameter diffusion retriever，在 continuous embedding space 以單一 non-autoregressive pass 生成完整 sub-query set。來源在 fashion（CLIP）與 music playlist（MuLan）benchmark 報告優於 single-query、zero-shot expansion 與 Best-of-N baseline，並宣稱 12–20× latency speedup over autoregressive fan-out；這些是 ICML 2026 論文（arXiv:2603.06397）研究 protocol 下的來源陳述，不是 AI Ark 的獨立重現或跨 domain／production 證據。相關 raw evidence：[[raw/articles/google-research-retrieve-for-train-rl-compiled-diffusion-fan-out-retrieval-2026-09-15]]；可與 agentic-rag-cross-corpus、document-parsing-first-rag、test-time-compute-evaluation 交叉閱讀。

2026-09-10 新文：ToolGrad

Google Research 發布 ToolGrad，提出 answer-first tool-use dataset generation：先以 textual gradients 引導 API workflow 的提案、執行、選擇與更新，再生成相符的 user query。來源以 ToolBench 與 BFCL 報告較高的資料生成 pass rate、較低成本與 ToolGrad-12B 的 tool-use benchmark 結果；這些是研究 protocol 與 benchmark snapshot 下的來源陳述，不是 AI Ark 的獨立重現或跨版本模型排名。相關 raw evidence：[[raw/articles/google-research-toolgrad-tool-use-dataset-generation-2026-09-10]]；可與 agent-post-training-workflow、dynamic-agent-workflows、eval-is-spec 交叉閱讀。

2026-09-03 新文：跨族群 genomic prediction

Google Research 介紹以 UK Biobank 歐洲樣本與 Biobank Japan 日本樣本評估 polygenic risk score（PRS）transfer learning 的研究，跨八個 trait 改變 source／target sample size，比較 UKB discovery + elastic net、cross-population meta-analysis 與 PRS-CSx。來源報告小型 target cohort 可受外部資料 pooling 幫助，但 target cohort 增大後，尤其是較 population-specific 的 trait，target-only training 可能表現更好；這些 sample-size crossover 是特定研究 protocol 的結果，不是臨床或其他族群的固定門檻。相關 raw evidence：[[raw/articles/google-research-transfer-learning-genomic-prediction-2026-09-03]]；可與 wearable-health-foundation-models、data-science-agents 和 eval-is-spec 交叉閱讀。

2026-08-26 新文：GlucoFM

Google Research 發布 GlucoFM，一個針對 continuous glucose monitoring（CGM）的 lightweight self-supervised foundation model。它用 dual-stream encoder 分開較慢的 glucose trend 與短期 deviation，保留 time-of-day、missingness 與 sensor variation，並在四個 cohort、七項臨床預測任務、PPGR forecasting、跨 cohort transfer 與 few-shot 設定中評估 frozen representation；這些結果屬研究資料與來源 protocol 下的 attribution，不等於臨床診斷或部署安全 benchmark。相關知識連結至 wearable-health-foundation-models、data-science-agents 與 eval-is-spec。

2026-08-27 新文：Planetary Prediction Engine（PPE）

Google Research 發布 Planetary Prediction Engine（PPE），把自然語言 geospatial query 交給 LLM 編排的 intelligent data selection、dataset curation、AutoML & prediction 流程，並以 opaque handles 在階段間傳遞大型 data artifacts。來源在美國公共衛生、奈及利亞 food security 與 DRC Ebola nowcasting 報告多項 benchmark 改善；這些結果仍屬 Google Research 的 early-stage experimental capability 與來源 attribution，不視為獨立重現或可直接部署的證據。相關知識連結至 planetary-prediction-engine、data-science-agents、world-models 與 empirical-research-assistance。

2026-08-31 新文：TimesFM-3

Google Research 發布 TimesFM-3，主打原生 multivariate time-series forecasting、multiple targets、past／past-future covariates、single-forward-pass 解碼與 point／quantile outputs。來源稱模型以 330M parameters、超過 1 trillion time points 的 real-world／synthetic corpus 預訓練，並在 Gift-Eval、FEV-Bench 與 TIME 報告最高的 foundation-model average rank；官方 GitHub README 另列 FEV-Bench、TIME 與 GIFT-Eval 的 rank #1 說法。這些是 Google Research 與官方 repo 的研究／benchmark attribution，不是 AI Ark 的獨立重現或 production SOTA。相關知識連結至 time-series-foundation-models、tabular-foundation-models、data-science-agents 與 eval-is-spec。

2026-09-01 新文：MAPL-EMIT

Google Research 發布 MAPL-EMIT，使用 Swin-S vision transformer、EMIT hyperspectral radiance 與 spatial context，自動進行 methane plume detection、enhancement quantification、delineation 與 source localization。文章與 arXiv 摘要報告在 1,084 個 EMIT granules 上捕捉 84% NASA L2B expert-annotated plume complexes，並找到約 1.5 倍 human analysts 的 plausible plumes；這些數字屬研究 protocol 下的來源陳述，不是獨立重現或減排成效 benchmark。相關知識連結至 world-models。

2026-08-25 新文：AgentHands

Google Research 發布 AgentHands XR prototype：以 scene registry、GestureEvents、文字／語音時間戳與本地 headset parser，把 LLM 回應轉成空間定位、動作示範與安全提示。來源報告 N = 12 的 within-subjects study，相較 speech-only baseline 在物件定位、複雜動作理解、警告注意力與認知負荷上有改善；這些結果屬 Google Research 的小型研究原型與特定任務研究，不視為通用 usability 或部署安全 benchmark。相關知識連結至 agent-experience、generative-ui、natively-adaptive-interfaces 與 agent-event-streaming-format。

Raw Sources

• raw/articles/google-research-millemiglia-middle-mile-logistics-instance-generator-2026-09-18.md — MilleMiglia：middle-mile logistics instance generator（C++／Protocol Buffers、space-time graph multi-commodity flow、fixed schedules／throughput／synchronization constraints）
• raw/articles/google-research-agentic-rag-2026-06-05.md — Unlocking Dependable Responses with Gemini Enterprise Agent Platform’s Agentic RAG
• raw/articles/google-research-machine-unlearning-audit-2026-06-10.md — New framework for auditing machine unlearning
• raw/articles/google-research-reasoningbank-agent-memory-2026-04-21.md — ReasoningBank: Enabling agents to learn from experience
• raw/articles/google-research-era-scientist-usage-2026-04-29.md — Four ways Google Research scientists have been using Empirical Research Assistance
• raw/articles/google-research-alphaevolve-theoretical-cs-2025-09-30.md — AI as a research partner: Advancing theoretical computer science with AlphaEvolve
• raw/articles/google-research-science-one-chain-of-evidence-2026-07-30.md — Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence
• raw/articles/google-research-deep-researcher-test-time-diffusion-2025-09-19.md — Deep researcher with test-time diffusion
• raw/articles/google-research-thinking-to-recall-parametric-knowledge-2026-06-24.md — Thinking to recall: How reasoning unlocks parametric knowledge in LLMs
• raw/articles/google-research-gemini-nano-frozen-mtp-2026-06-26.md — Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
• raw/articles/google-research-groundsource-news-to-data-gemini-2026-03-12.md — Introducing Groundsource: Turning news reports into data with Gemini
• raw/articles/google-research-tabfm-zero-shot-tabular-data-2026-06-30.md — Introducing TabFM: A zero-shot foundation model for tabular data
• raw/articles/google-research-sensorfm-wearable-health-data-2026-07-09.md — SensorFM: Towards a general intelligence and interface for wearable health data
• raw/articles/google-research-symptomai-everyday-symptom-assessment-2026-07-22.md — SymptomAI: Towards a conversational AI agent for everyday symptom assessment
• raw/articles/google-research-amie-audio-visual-clinical-consultations-2026-08-11.md — AMIE (Video)：real-time audio-visual clinical consultation、三 agent 架構與 randomized OSCE
• raw/articles/google-research-knowledge-profiling-parametric-factuality-2026-08-12.md — Empty shelves or lost keys? Recall is the bottleneck for parametric factuality
• raw/articles/google-research-photoscan-cardiometabolic-risk-2026-08-17.md — PhotoScan：用 smartphone imagery 估計 body composition 與 cardiometabolic risk
• raw/articles/google-research-biomarker-discovery-framework-2026-08-21.md — Biomarker Discovery Framework：wearable biomarker prioritization、adversarial multi-agent loop 與 human expert evaluation
• raw/articles/google-research-mobility-embedded-pois-2026-08-21.md — Mobility-Embedded POIs、文字與匿名 mobility 融合、場所功能節律與 unseen-place evaluation
• raw/articles/google-research-agenthands-spatially-grounded-agent-conversations-xr-2026-08-25.md — AgentHands：XR 空間定位、GestureEvents、同步手勢與 N=12 使用者研究
• raw/articles/google-research-glucofm-continuous-glucose-monitoring-2026-08-26.md — GlucoFM：CGM dual-stream foundation model、latent prediction、跨 cohort transfer 與 few-shot evaluation
• raw/articles/google-research-planetary-prediction-engine-earth-ai-2026-08-27.md — Planetary Prediction Engine：geospatial data selection、dataset curation、AutoML、multimodal fusion 與 prediction benchmarks
• raw/articles/google-research-timesfm-3-multivariate-forecasting-2026-08-31.md — TimesFM-3：多變量時間序列、covariates、single-pass quantile forecasting 與官方 benchmark／授權邊界
• raw/articles/google-research-mapl-emit-methane-deep-learning-2026-09-01.md — MAPL-EMIT：hyperspectral methane plume 偵測、source localization、synthetic training、benchmark 與信心／物理單位邊界
• raw/articles/google-research-ds-star-data-science-agent-2025-11-06.md — DS-STAR: A state-of-the-art versatile data science agent
• raw/articles/google-research-dialoglab-human-ai-group-conversations-2026-02-10.md — Beyond one-on-one: Authoring, simulating, and testing dynamic human-AI group conversations
• raw/articles/google-research-titans-miras-long-term-memory-2025-12-04.md — Titans + MIRAS: Helping AI have long-term memory
• raw/articles/google-research-bayesian-teaching-llm-reasoning-2026-03-04.md — Teaching LLMs to reason like Bayesians
• raw/articles/google-research-nested-learning-continual-learning-2025-11-07.md — Introducing Nested Learning: A new ML paradigm for continual learning
• raw/articles/google-research-small-models-intent-extraction-decomposition-2026-01-22.md — Small models, big results: Achieving superior intent extraction through decomposition
• raw/articles/google-research-generative-ui-2025-11-18.md — Generative UI: A rich, custom, visual interactive user experience for any prompt
• raw/articles/google-research-sequential-attention-model-efficiency-2026-02-04.md — Sequential Attention: Making AI models leaner and faster without sacrificing accuracy
• raw/articles/google-research-differentially-private-chatbot-use-2025-12-10.md — A differentially private framework for gaining insights into AI chatbot use
• raw/articles/google-research-gist-smart-sampling-2026-01-23.md — Introducing GIST: The next stage in smart sampling
• raw/articles/google-research-atlas-multilingual-scaling-laws-2026-01-27.md — ATLAS: Practical scaling laws for multilingual models
• raw/articles/google-research-toolgrad-tool-use-dataset-generation-2026-09-10.md — ToolGrad: Efficient tool-use dataset generation with textual “gradients”
• raw/articles/google-research-retrieve-for-train-rl-compiled-diffusion-fan-out-retrieval-2026-09-15.md — Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train（R4T）