Deprecated: The each() function is deprecated. This message will be suppressed on further calls in /home/zhenxiangba/zhenxiangba.com/public_html/phproxy-improved-master/index.php on line 456 Paper page - GPT-4V(ision) is a Generalist Web Agent, if Grounded
\n","updatedAt":"2024-06-08T18:51:27.406Z","author":{"_id":"6186ddf6a7717cb375090c01","avatarUrl":"/avatars/716b6a7d1094c8036b2a8a7b9063e8aa.svg","fullname":"Julien BLANCHON","name":"blanchon","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":176,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.5016820430755615},"editors":["blanchon"],"editorAvatarUrls":["/avatars/716b6a7d1094c8036b2a8a7b9063e8aa.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2401.01614","authors":[{"_id":"65960e11b02e572eb0bfd00c","user":{"_id":"631a95cfa66151e36e54a905","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/631a95cfa66151e36e54a905/jM_4vy9_Kd8zomFVYtsTo.png","isPro":false,"fullname":"Boyuan Zheng","user":"boyuanzheng010","type":"user"},"name":"Boyuan Zheng","status":"claimed_verified","statusLastChangedAt":"2024-03-18T08:37:29.919Z","hidden":false},{"_id":"65960e11b02e572eb0bfd00d","user":{"_id":"6500870f1e14749e84f8f887","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6500870f1e14749e84f8f887/wfvx4BZvh2OyW-vpq5jEy.jpeg","isPro":false,"fullname":"Boyu Gou","user":"BoyuNLP","type":"user"},"name":"Boyu Gou","status":"claimed_verified","statusLastChangedAt":"2024-01-04T09:35:11.800Z","hidden":false},{"_id":"65960e11b02e572eb0bfd00e","name":"Jihyung Kil","hidden":false},{"_id":"65960e11b02e572eb0bfd00f","name":"Huan Sun","hidden":false},{"_id":"65960e11b02e572eb0bfd010","user":{"_id":"6477a323dbc2a416f8b852b3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6477a323dbc2a416f8b852b3/mRKW5kT9GASORT4YnaZz0.jpeg","isPro":false,"fullname":"Yu Su","user":"ysu-nlp","type":"user"},"name":"Yu Su","status":"claimed_verified","statusLastChangedAt":"2024-01-04T09:35:14.490Z","hidden":false}],"publishedAt":"2024-01-03T08:33:09.000Z","submittedOnDailyAt":"2024-01-03T23:17:00.535Z","title":"GPT-4V(ision) is a Generalist Web Agent, if Grounded","submittedOnDailyBy":{"_id":"60f1abe7544c2adfd699860c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674929746905-60f1abe7544c2adfd699860c.jpeg","isPro":false,"fullname":"AK","user":"akhaliq","type":"user"},"summary":"The recent development on large multimodal models (LMMs), especially\nGPT-4V(ision) and Gemini, has been quickly expanding the capability boundaries\nof multimodal models beyond traditional tasks like image captioning and visual\nquestion answering. In this work, we explore the potential of LMMs like GPT-4V\nas a generalist web agent that can follow natural language instructions to\ncomplete tasks on any given website. We propose SEEACT, a generalist web agent\nthat harnesses the power of LMMs for integrated visual understanding and acting\non the web. We evaluate on the recent MIND2WEB benchmark. In addition to\nstandard offline evaluation on cached websites, we enable a new online\nevaluation setting by developing a tool that allows running web agents on live\nwebsites. We show that GPT-4V presents a great potential for web agents - it\ncan successfully complete 50% of the tasks on live websites if we manually\nground its textual plans into actions on the websites. This substantially\noutperforms text-only LLMs like GPT-4 or smaller models (FLAN-T5 and BLIP-2)\nspecifically fine-tuned for web agents. However, grounding still remains a\nmajor challenge. Existing LMM grounding strategies like set-of-mark prompting\nturns out not effective for web agents, and the best grounding strategy we\ndevelop in this paper leverages both the HTML text and visuals. Yet, there is\nstill a substantial gap with oracle grounding, leaving ample room for further\nimprovement.","upvotes":22,"discussionId":"65960e14b02e572eb0bfd09e","githubRepo":"https://github.com/osu-nlp-group/seeact","githubRepoAddedBy":"auto","ai_summary":"LMMs like GPT-4V demonstrate potential as generalist web agents by following natural language instructions to complete tasks on live websites, outperforming text-only LLMs and smaller models with a combination of HTML text and visuals for grounding, though challenges remain.","ai_keywords":["GPT-4V","multimodal models","SEEACT","MIND2WEB benchmark","web agents","visual understanding","GPT-4","FLAN-T5","BLIP-2","set-of-mark prompting","oracle grounding"],"githubStars":824},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6477a323dbc2a416f8b852b3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6477a323dbc2a416f8b852b3/mRKW5kT9GASORT4YnaZz0.jpeg","isPro":false,"fullname":"Yu Su","user":"ysu-nlp","type":"user"},{"_id":"6500870f1e14749e84f8f887","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6500870f1e14749e84f8f887/wfvx4BZvh2OyW-vpq5jEy.jpeg","isPro":false,"fullname":"Boyu Gou","user":"BoyuNLP","type":"user"},{"_id":"6409915b5d82bf3a2624e09c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678865109748-6409915b5d82bf3a2624e09c.jpeg","isPro":false,"fullname":"HanWang","user":"eseedo","type":"user"},{"_id":"631d760344503b7227837242","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/631d760344503b7227837242/3b6JRusFX6GKJpsN9ZdeJ.png","isPro":false,"fullname":"Max Ku","user":"vinesmsuic","type":"user"},{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"650b95e72ec41a31f3155cf6","avatarUrl":"/avatars/7d87dccf270925f42538ccf33433d799.svg","isPro":false,"fullname":"Ryo Kodama","user":"rkodama","type":"user"},{"_id":"64747f7e33192631bacd8831","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64747f7e33192631bacd8831/dstkZJ4sHJSeqLesV5cOC.jpeg","isPro":false,"fullname":"Taufiq Dwi Purnomo","user":"taufiqdp","type":"user"},{"_id":"6269e6ea124f03fb2e0c15d9","avatarUrl":"/avatars/4e7ac8f92d7fdbb129e0b1fbe2f24cb7.svg","isPro":false,"fullname":"Vincent Tu","user":"alckasoc","type":"user"},{"_id":"65676a0a461af93fca9f2329","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65676a0a461af93fca9f2329/-CB4C1C6yLM4gRU2K5gsS.jpeg","isPro":false,"fullname":"Juan Delgadillo","user":"juandelgadillo","type":"user"},{"_id":"6270d2ddbcef985363d774fa","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6270d2ddbcef985363d774fa/HOKAxx_FKVRF-87WpGQbF.png","isPro":true,"fullname":"jiakai","user":"real-jiakai","type":"user"},{"_id":"639c379cdb7c5f35004066cb","avatarUrl":"/avatars/3e435506ee85aa7d2d0ec2174a07462f.svg","isPro":false,"fullname":"Zhenran Xu","user":"imryanxu","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":1}">
LMMs like GPT-4V demonstrate potential as generalist web agents by following natural language instructions to complete tasks on live websites, outperforming text-only LLMs and smaller models with a combination of HTML text and visuals for grounding, though challenges remain.
AI-generated summary
The recent development on large multimodal models (LMMs), especially
GPT-4V(ision) and Gemini, has been quickly expanding the capability boundaries
of multimodal models beyond traditional tasks like image captioning and visual
question answering. In this work, we explore the potential of LMMs like GPT-4V
as a generalist web agent that can follow natural language instructions to
complete tasks on any given website. We propose SEEACT, a generalist web agent
that harnesses the power of LMMs for integrated visual understanding and acting
on the web. We evaluate on the recent MIND2WEB benchmark. In addition to
standard offline evaluation on cached websites, we enable a new online
evaluation setting by developing a tool that allows running web agents on live
websites. We show that GPT-4V presents a great potential for web agents - it
can successfully complete 50% of the tasks on live websites if we manually
ground its textual plans into actions on the websites. This substantially
outperforms text-only LLMs like GPT-4 or smaller models (FLAN-T5 and BLIP-2)
specifically fine-tuned for web agents. However, grounding still remains a
major challenge. Existing LMM grounding strategies like set-of-mark prompting
turns out not effective for web agents, and the best grounding strategy we
develop in this paper leverages both the HTML text and visuals. Yet, there is
still a substantial gap with oracle grounding, leaving ample room for further
improvement.