[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-convawg-controlled-vawg-dialogue-generation-zh":3,"article-related-convawg-controlled-vawg-dialogue-generation-zh":29,"series-research-1544300b-d8fc-4341-ac8a-530391b179b2":72},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":11},"1544300b-d8fc-4341-ac8a-530391b179b2","convawg-controlled-vawg-dialogue-generation-zh","ConVAWG 讓 VAWG 對話可控生成","\u003Cp data-speakable=\"summary\">6,000+ 筆合成對話事件證明，ConVAWG 能用轉輪毒性控制生成可控的 VAWG 情境。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>研究機構\u003C\u002Fstrong>：arXiv 摘要未明確標註\u003C\u002Fli>\u003Cli>\u003Cstrong>核心數據\u003C\u002Fstrong>：6,000+ 多輪對話事件\u003C\u002Fli>\u003Cli>\u003Cstrong>突破點\u003C\u002Fstrong>：檢索式情境＋毒性控制\u003C\u002Fli>\u003C\u002Ful>\u003Cp>這篇論文要解的，不是單句髒話分類，而是更難的問題：怎麼在不碰真實受害者資料的前提下，做出看起來像真實互動的 VAWG 多輪對話。作者把暴力對女性與女孩的行為，當成一段會隨時間推進的對話流程來建模，而不是只看單一有害句子。\u003C\u002Fp>\u003Cp>這個切法很重要。很多安全研究只抓得到明顯辱罵，卻抓不到更常見的互動型傷害，例如控制、威脅、跟蹤、孤立、施壓，或是一步一步升高的言語暴力。若資料本身只有單句，模型就很難學到「傷害是怎麼展開的」。\u003C\u002Fp>\u003Ch2>這篇在補哪個缺口\u003C\u002Fh2>\u003Cp>根據摘要，既有研究多半聚焦在句子層級的 toxicity。這對偵測粗暴字眼有幫助，但對理解關係脈絡與時間序列就不夠了。真實世界的虐待與騷擾，往往不是一槍斃命，而是透過多輪互動慢慢堆疊。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786516377717-xygt.png\" alt=\"ConVAWG 讓 VAWG 對話可控生成\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>ConVAWG 的定位，是做「可控的情境生成」，不是一般聊天生成。它要產出的，是和 CPS 類情境對齊的合成對話。這讓資料可以用在安全評估、對話分析、或有害對話偵測的測試集建置上。\u003C\u002Fp>\u003Cp>對開發者來說，這種資料的價值在於可用性。真實敏感對話常常不能公開，或者根本不能大規模共享。合成資料如果能保留互動結構，就能補上\u003Ca href=\"\u002Fnews\u002Fsurgical-wam-video-pretraining-robot-control-zh\">訓練\u003C\u002Fa>與評估時最缺的那一塊。\u003C\u002Fp>\u003Ch2>方法怎麼做出來\u003C\u002Fh2>\u003Cp>ConVAWG 的起點不是讓模型自由發揮，而是先做情境定位。摘要提到，它從 persona seeds、英國國家統計局的族群與人口模式、官方犯罪定義，以及檢索到的 Domestic Homicide Review 案例出發，再把這些材料整理成階層式事件時間線。\u003C\u002Fp>\u003Cp>接著，這些時間線會驅動多場景的角色扮演式對話生成。白話一點說，就是先有骨架，再長出對話。不是單純丟 prompt 叫模型編故事，而是先把事件順序、情境脈絡和角色關係固定下來，再讓模型填成多輪互動。\u003C\u002Fp>\u003Cp>摘要裡另一個關鍵詞是 targeted activation-steered toxicity control。雖然摘要沒有公開完整實作細節，但方向很清楚：毒性不是平均撒在整段對話裡，而是只在情境需要的回合被引導出來。這樣比較接近真實互動，也比較能控制輸出。\u003C\u002Fp>\u003Cp>這種設計其實很有工程感。檢索負責 grounding，時間線負責一致性，定向控制負責讓有害內容出現在對的位置。三者合起來，才有機會做出可分析、可重現的敏感對話資料。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>摘要明確提到，作者釋出了 6,000+ 筆多輪對話事件，涵蓋 200 個情境，而且每筆資料都有豐富的 scenario、event、turn 層級 metadata。這點很實用，因為它不只是給一串文字，而是給可以切片分析的結構化資料。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786516379448-04ms.png\" alt=\"ConVAWG 讓 VAWG 對話可控生成\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>作者也說他們做了大量 human evaluation、\u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa>-as-Judge、ablation，還有 downstream tasks，結果顯示對話品質與領域貼合度都不錯。不過，摘要沒有公開完整 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 數字、\u003Ca href=\"\u002Fnews\u002Fbenchmark-scores-dont-predict-your-bill-zh\">分數\u003C\u002Fa>或詳細評測表，所以目前能確定的是方法與資料集層面的成果，不是某個\u003Ca href=\"\u002Fnews\u002Fswe-bench-verified-model-leaderboard-limit-zh\">排行榜\u003C\u002Fa>成績。\u003C\u002Fp>\u003Cp>也就是說，這篇的公開訊息重點不在「比誰高幾個百分點」，而在「怎麼把難以取得的敏感互動資料，做成可控又有結構的合成資料」。對很多應用場景來說，這種成果比單一分數更有落地價值。\u003C\u002Fp>\u003Cul>\u003Cli>檢索式情境建構，把生成綁回外部案例材料。\u003C\u002Fli>\u003Cli>階層式事件時間線，保住多輪敘事結構。\u003C\u002Fli>\u003Cli>定向毒性控制，讓有害內容出現在該出現的回合。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>對開發者有什麼用\u003C\u002Fh2>\u003Cp>如果你在做 moderation、safety evaluation、或 abuse detection，這篇最值得看的地方是：它把「有害對話」當成流程問題，而不是字詞問題。這會直接影響資料設計。你要測的不是模型會不會認出髒話，而是它能不能看懂前後文、關係變化與威脅升高。\u003C\u002Fp>\u003Cp>資料格式也比平面文字更好用。scenario-level、event-level、turn-level 的 metadata，能讓你做分層評估、錯誤分析、課程式訓練，甚至檢查模型在哪個回合開始失真。這對做安全產品的人很實際，因為你不只想知道「有沒有判錯」，還想知道「錯在哪一段」。\u003C\u002Fp>\u003Cp>但限制也要看清楚。摘要沒有說這 200 個情境的多樣性到底有多廣，也沒有公開完整 benchmark 數字，所以不能只憑摘要就推論它在所有任務上都很強。它也沒有說這套方法能不能自然地泛化到 VAWG 以外的敏感領域。\u003C\u002Fp>\u003Cp>另外，任何合成的暴力資料都有一個共同風險：它會反映來源材料與生成流程的假設。換句話說，這類資料適合拿來做分析、測試和輔助訓練，但不應被當成真實世界理解的替代品。對實務團隊來說，最好把它視為一個可控的壓力測試工具。\u003C\u002Fp>\u003Ch2>總結\u003C\u002Fh2>\u003Cp>ConVAWG 證明了一件事：在難以使用真實敏感資料的領域，檢索式情境加上階層式事件建模，可以把 VAWG 對話做成可控的多輪合成資料。它的亮點不是單一分數，而是把「可生成、可控、可檢查」這三件事串起來。\u003C\u002Fp>\u003Cp>對\u003Ca href=\"\u002Ftag\u002F台灣開發者\">台灣開發者\u003C\u002Fa>來說，這篇比較像一個方法模板。若你要做任何敏感對話資料集，重點都不是先追求大量文本，而是先想清楚：情境怎麼來、事件怎麼排、毒性要怎麼精準放進對話裡。ConVAWG 提供的，就是這條路線。\u003C\u002Fp>","ConVAWG 用檢索式情境與轉輪毒性控制，生成 6,000+ 筆多輪 VAWG 對話事件。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.11200",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786516377717-xygt.png","research","zh","605dd415-e62d-455a-bb4b-d1d2fa487c1b",[17,18,19,20,21],"VAWG","synthetic dialogue","toxicity control","retrieval grounding","multi-turn dialogue",[23,24,25],"把有害對話當成多輪流程來建模，比單句偵測更貼近真實互動。","用檢索式情境與階層式事件時間線，讓合成資料更有結構與一致性。","摘要沒有公開完整 benchmark 數字，因此目前可確認的是方法與資料集價值。",1,"2026-08-12T06:32:30.301717+00:00","2026-08-12T06:32:30.28+00:00",{"tags":30,"relatedLang":31,"relatedPosts":35},[],{"id":15,"slug":32,"title":33,"language":34},"convawg-controlled-vawg-dialogue-generation-en","ConVAWG generates controlled VAWG dialogues","en",[36,42,48,54,60,66],{"id":37,"slug":38,"title":39,"cover_image":40,"image_url":40,"created_at":41,"category":13},"2aa54cfd-2caf-4c34-834c-e771e1308747","sparse-autoencoders-set-level-instability-zh","SAE 不是特徵袋","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786518181678-0ycl.png","2026-08-12T07:02:31.647391+00:00",{"id":43,"slug":44,"title":45,"cover_image":46,"image_url":46,"created_at":47,"category":13},"51473b63-b17b-492a-8dc0-b12069a19b49","surgical-wam-video-pretraining-robot-control-zh","Surgical WAM 用影片訓練手術機器人控制","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786514588496-1uqy.png","2026-08-12T06:02:32.223678+00:00",{"id":49,"slug":50,"title":51,"cover_image":52,"image_url":52,"created_at":53,"category":13},"97ecb84d-bd2a-437b-839e-6e8a416d4a94","swe-bench-verified-model-leaderboard-limit-zh","SWE-bench Verified 已不再是乾淨的模型排行榜","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786498365710-5zg7.png","2026-08-12T01:32:19.598349+00:00",{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":13},"0299c84e-cdca-4d9f-821c-437119627dbf","dutch-government-llm-benchmark-values-zh","荷蘭政府 LLM 不能只看準確率","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786431771084-my8i.png","2026-08-11T07:02:25.893246+00:00",{"id":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":13},"b076a7b7-f01a-4438-8d49-548201e9ccec","mmdiff-multimodal-feature-discovery-control-zh","MMDiff：把多模態特徵變成控制旋鈕","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786429980947-qtkn.png","2026-08-11T06:32:30.336974+00:00",{"id":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":13},"f5a1bf22-1f75-4be1-873e-f2bd717c2397","tts-evaluators-miss-more-than-naturalness-zh","TTS 評測不只看自然度","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786428174396-4b56.png","2026-08-11T06:02:28.652133+00:00",[73,78,83,88,93,98,103,108,113,118],{"id":74,"slug":75,"title":76,"created_at":77},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":79,"slug":80,"title":81,"created_at":82},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":84,"slug":85,"title":86,"created_at":87},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":89,"slug":90,"title":91,"created_at":92},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00",{"id":94,"slug":95,"title":96,"created_at":97},"53a0dc54-0371-4e40-8d5e-74e94a73840c","geometry-aware-similarity-metrics-for-neural-representations-zh","超越距離測量：用微分幾何重新理解神經網路","2026-03-31T06:01:01.241968+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"fee7d472-a775-4b1d-bbc2-1e8bca1bbf8b","on-the-fly-repulsion-in-the-contextual-space-for-rich-divers-zh","讓AI繪圖更有創意：用排斥力提升生成多樣性","2026-03-31T06:01:25.439673+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"a9901203-d69b-447b-8854-15d14eab32b4","vision-aided-beam-prediction-cnn-eca-zh","影像輔助波束預測升級 CNN","2026-04-01T10:00:25.8073+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"b55e7dd4-0a24-4b3d-804d-b0309a03f498","triple-band-fss-mimo-antenna-sub-6-ghz-zh","三頻 FSS MIMO 天線瞄準 sub-6 GHz","2026-04-01T13:18:36.857305+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"f68290bd-e7f3-4b30-ba22-dcd4e0130a66","openclaw-1299-repos-eight-weeks-analysis-zh","OpenClaw 1299 個 Repo 的資料解讀","2026-04-02T05:03:45.208411+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"ed9f80eb-eb02-4d35-8ad4-0ddf428751dd","beam-coherence-aware-combining-mmwave-mimo-zh","毫米波 MIMO 的雙階合併法","2026-04-02T05:27:26.897188+00:00"]