[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-language-models":3},{"tag":4,"articles":11,"peer_article_count":85},{"id":5,"name":6,"slug":7,"article_count":8,"description_zh":9,"description_en":10},"3a123b8f-ad63-4875-82f2-2616713263ce","language models","language-models",3,"語言模型是生成式 AI 的核心，涵蓋預訓練、詞彙擴充、對齊與安全評估等議題。這裡會整理模型如何學習語意、處理新 token，以及在 jailbreak 與漏洞測試中暴露的風險。","Language models sit at the core of generative AI, spanning pretraining, token initialization, alignment, and security evaluation. This tag collects work on how LMs learn semantics, absorb new vocabulary, and where jailbreak tests expose failure modes.",[12,21,28,35,42,49,56,63,70,77],{"id":13,"slug":14,"title":15,"summary":16,"category":17,"image_url":18,"cover_image":18,"language":19,"created_at":20},"b398938d-f651-4d91-bfee-d888ba44fe6f","diffusiongemma-transparency-measured-en","DiffusionGemma’s transparency problem, measured","Researchers split diffusion-model transparency into two parts and show DiffusionGemma can be made much more interpretable.","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781848969642-b497.png","en","2026-06-19T06:02:30.672396+00:00",{"id":22,"slug":23,"title":24,"summary":25,"category":17,"image_url":26,"cover_image":26,"language":19,"created_at":27},"01f05d3f-fb22-4194-b211-bfe8e02bd544","language-models-value-axis-en","Language models have a “value axis”","A new paper shows Qwen3-8B internally tracks whether its current path is likely to succeed.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781589776527-cruc.png","2026-06-16T06:02:35.947355+00:00",{"id":29,"slug":30,"title":31,"summary":32,"category":17,"image_url":33,"cover_image":33,"language":19,"created_at":34},"1770f0e4-4b10-459d-bb9b-be13075b1a3d","persona-pruner-lightweight-role-playing-models-en","Persona-Pruner trims models for role-playing","Persona-Pruner prunes language models into persona-specific role-play bots while keeping general capabilities intact.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781505171903-58bv.png","2026-06-15T06:32:25.55966+00:00",{"id":36,"slug":37,"title":38,"summary":39,"category":17,"image_url":40,"cover_image":40,"language":19,"created_at":41},"53ec2203-e127-4bf8-8b3d-2dce8d156a54","causal-learnability-formal-language-tasks-en","Causal methods for measuring task learnability","This paper shows correlational learnability tests can mislead, and proposes causal tools for formal-language tasks.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780987698514-ky8m.png","2026-06-09T06:47:35.103221+00:00",{"id":43,"slug":44,"title":45,"summary":46,"category":17,"image_url":47,"cover_image":47,"language":19,"created_at":48},"480aabe2-9885-456e-8ea0-490f39890389","next-token-models-plan-ahead-en","Why next-token models can plan ahead","This paper argues autoregressive language models can exhibit lookahead behavior despite training only on next-token prediction.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780645687192-whr3.png","2026-06-05T07:47:34.828225+00:00",{"id":50,"slug":51,"title":52,"summary":53,"category":17,"image_url":54,"cover_image":54,"language":19,"created_at":55},"a3c57be7-a302-4666-a308-113cb75f7494","how-to-build-ai-research-foundations-with-deepmind-en","How to Build AI Research Foundations with DeepMind","Follow this guide to build a practical foundation in modern language models and fine-tuning.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779963500137-2j1d.png","2026-05-28T10:17:24.797504+00:00",{"id":57,"slug":58,"title":59,"summary":60,"category":17,"image_url":61,"cover_image":61,"language":19,"created_at":62},"43ff5c87-307c-42f3-9a93-15ba8b239f83","convextok-tokenisation-convex-relaxations-en","ConvexTok Reframes Tokenization as Optimization","ConvexTok turns tokeniser construction into a linear program and gets closer-to-optimal tokenization.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779431162626-69fc.png","2026-05-22T06:25:37.211584+00:00",{"id":64,"slug":65,"title":66,"summary":67,"category":17,"image_url":68,"cover_image":68,"language":19,"created_at":69},"22c43f4e-8be9-4440-bd1b-74a00b60dfa3","llms-implicit-grammar-representations-en","Do LLMs Learn Grammar Beyond Likelihood?","A probe study finds hidden layers in language models encode grammaticality better than string probability, but not plausibility.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778135464967-fzem.png","2026-05-07T06:30:35.804749+00:00",{"id":71,"slug":72,"title":73,"summary":74,"category":17,"image_url":75,"cover_image":75,"language":19,"created_at":76},"b712257f-129d-400a-bc73-5e1c3ab200a4","avise-ai-security-evaluation-framework-en","AVISE tests AI security with modular jailbreak evals","AVISE is an open-source framework for finding AI vulnerabilities, with a 25-case jailbreak test that flagged all nine models as vulnerable.","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776924767358-ocir.png","2026-04-23T06:12:31.125572+00:00",{"id":78,"slug":79,"title":80,"summary":81,"category":82,"image_url":83,"cover_image":83,"language":19,"created_at":84},"e487e7c6-aa22-484d-9555-46261cc7a91d","grounded-token-initialization-new-vocabulary-en","A Better Way to Seed New LM Tokens","GTI grounds new vocabulary tokens before fine-tuning, aiming to preserve distinctions that mean initialization tends to collapse.","blockchain","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775196588405-1a7u.png","2026-04-03T06:09:29.832749+00:00",5]