\n\n@librarian-bot\n\t suggest similar articles
\n","updatedAt":"2024-10-02T15:36:19.077Z","author":{"_id":"6494638e0a97fa9dbf33ff81","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/bqQU2ZxMfyo_dT35TzNdM.jpeg","fullname":"Diego Rodrigues","name":"diegomrodrigues","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8893412947654724},"editors":["diegomrodrigues"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/bqQU2ZxMfyo_dT35TzNdM.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2312.06585","authors":[{"_id":"6577f1e4ee33d547aebf898d","name":"Avi Singh","hidden":false},{"_id":"6577f1e4ee33d547aebf898e","name":"John D. Co-Reyes","hidden":false},{"_id":"6577f1e4ee33d547aebf898f","user":{"_id":"647875e37b370854241b89a8","avatarUrl":"/avatars/fd25f7e5960e4e132b9ee689f114820b.svg","isPro":false,"fullname":"Rishabh Agarwal","user":"agarwl","type":"user"},"name":"Rishabh Agarwal","status":"extracted_pending","statusLastChangedAt":"2023-12-12T05:38:44.332Z","hidden":false},{"_id":"6577f1e4ee33d547aebf8990","user":{"_id":"626321a613cc9ec91bd7de48","avatarUrl":"/avatars/79e4f1843a5a8041197b01772f21bc48.svg","isPro":false,"fullname":"Ankesh Anand","user":"ankeshanand","type":"user"},"name":"Ankesh Anand","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:23:09.319Z","hidden":false},{"_id":"6577f1e4ee33d547aebf8991","name":"Piyush Patil","hidden":false},{"_id":"6577f1e4ee33d547aebf8992","user":{"_id":"62ded294dcdc9043efaa4fc3","avatarUrl":"/avatars/1294078b8cd5c1901e4021253d877357.svg","isPro":false,"fullname":"Peter J Liu","user":"peterjliu","type":"user"},"name":"Peter J. Liu","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:23:45.836Z","hidden":false},{"_id":"6577f1e4ee33d547aebf8993","name":"James Harrison","hidden":false},{"_id":"6577f1e4ee33d547aebf8994","name":"Jaehoon Lee","hidden":false},{"_id":"6577f1e4ee33d547aebf8995","name":"Kelvin Xu","hidden":false},{"_id":"6577f1e4ee33d547aebf8996","user":{"_id":"635f493ba81c7f7424a702ae","avatarUrl":"/avatars/6f05d4f76609607966ed4635cb1e0a02.svg","isPro":false,"fullname":"Aaron Parisi","user":"AnotherA","type":"user"},"name":"Aaron Parisi","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:25:14.232Z","hidden":false},{"_id":"6577f1e4ee33d547aebf8997","name":"Abhishek Kumar","hidden":false},{"_id":"6577f1e4ee33d547aebf8998","user":{"_id":"6303ff84fc783bfc74441afc","avatarUrl":"/avatars/bc295e0d94c9c8be0c6a8264587ad4a5.svg","isPro":false,"fullname":"Alex Alemi","user":"knowflake","type":"user"},"name":"Alex Alemi","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:25:41.874Z","hidden":false},{"_id":"6577f1e4ee33d547aebf8999","name":"Alex Rizkowsky","hidden":false},{"_id":"6577f1e4ee33d547aebf899a","name":"Azade Nova","hidden":false},{"_id":"6577f1e4ee33d547aebf899b","name":"Ben Adlam","hidden":false},{"_id":"6577f1e4ee33d547aebf899c","name":"Bernd Bohnet","hidden":false},{"_id":"6577f1e4ee33d547aebf899d","name":"Hanie Sedghi","hidden":false},{"_id":"6577f1e4ee33d547aebf899e","name":"Igor Mordatch","hidden":false},{"_id":"6577f1e4ee33d547aebf899f","name":"Isabelle Simpson","hidden":false},{"_id":"6577f1e4ee33d547aebf89a0","user":{"_id":"64c0ccc19832a455bc2a2069","avatarUrl":"/avatars/9d4fdf13cd62c6c258b5f50d67c8a1b1.svg","isPro":false,"fullname":"IZZEDDIN GUR","user":"izzeddingur","type":"user"},"name":"Izzeddin Gur","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:26:12.768Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89a1","name":"Jasper Snoek","hidden":false},{"_id":"6577f1e4ee33d547aebf89a2","name":"Jeffrey Pennington","hidden":false},{"_id":"6577f1e4ee33d547aebf89a3","user":{"_id":"62af5f9e24488e6adfa86321","avatarUrl":"/avatars/5990e5f656c75530e6e8f01f644f7c8d.svg","isPro":false,"fullname":"Jiri Hron","user":"jirihron","type":"user"},"name":"Jiri Hron","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:26:27.286Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89a4","name":"Kathleen Kenealy","hidden":false},{"_id":"6577f1e4ee33d547aebf89a5","user":{"_id":"630e6ef664f1f8d0c771b758","avatarUrl":"/avatars/1df2bfadb2b6fdf8307189936efc6ef0.svg","isPro":false,"fullname":"Kevin Swersky","user":"kswersky","type":"user"},"name":"Kevin Swersky","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:26:41.648Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89a6","name":"Kshiteej Mahajan","hidden":false},{"_id":"6577f1e4ee33d547aebf89a7","name":"Laura Culp","hidden":false},{"_id":"6577f1e4ee33d547aebf89a8","name":"Lechao Xiao","hidden":false},{"_id":"6577f1e4ee33d547aebf89a9","name":"Maxwell L. Bileschi","hidden":false},{"_id":"6577f1e4ee33d547aebf89aa","name":"Noah Constant","hidden":false},{"_id":"6577f1e4ee33d547aebf89ab","user":{"_id":"63e2769bf0740bec2bfff1a1","avatarUrl":"/avatars/e2841a836319d43035bcddfbfe3344d0.svg","isPro":false,"fullname":"Roman Novak","user":"kulpims","type":"user"},"name":"Roman Novak","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:27:20.964Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89ac","name":"Rosanne Liu","hidden":false},{"_id":"6577f1e4ee33d547aebf89ad","user":{"_id":"64d51be85e5f05485ca50751","avatarUrl":"/avatars/5802fcc5a5f1e089df41232892f8f502.svg","isPro":false,"fullname":"Tris Warkentin","user":"trisfromgoogle","type":"user"},"name":"Tris Warkentin","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:27:44.017Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89ae","name":"Yundi Qian","hidden":false},{"_id":"6577f1e4ee33d547aebf89af","name":"Ethan Dyer","hidden":false},{"_id":"6577f1e4ee33d547aebf89b0","name":"Behnam Neyshabur","hidden":false},{"_id":"6577f1e4ee33d547aebf89b1","user":{"_id":"63d209ae8eaa4831006b3b0d","avatarUrl":"/avatars/51cc0f6dfab2233bddd73bfcd0c8aee3.svg","isPro":false,"fullname":"Jascha Sohl-Dickstein","user":"jaschasd","type":"user"},"name":"Jascha Sohl-Dickstein","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:28:09.903Z","hidden":false},{"_id":"6577f1e4ee33d547aebf89b2","user":{"_id":"62f82ebd870a3f98bbf9e62a","avatarUrl":"/avatars/3b03299e128abf2c8ba82db704855de7.svg","isPro":false,"fullname":"Noah Fiedel","user":"nfiedel","type":"user"},"name":"Noah Fiedel","status":"admin_assigned","statusLastChangedAt":"2023-12-12T08:28:18.023Z","hidden":false}],"publishedAt":"2023-12-11T18:17:43.000Z","submittedOnDailyAt":"2023-12-12T03:08:44.351Z","title":"Beyond Human Data: Scaling Self-Training for Problem-Solving with\n Language Models","submittedOnDailyBy":{"_id":"60f1abe7544c2adfd699860c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674929746905-60f1abe7544c2adfd699860c.jpeg","isPro":false,"fullname":"AK","user":"akhaliq","type":"user"},"summary":"Fine-tuning language models~(LMs) on human-generated data remains a prevalent\npractice. However, the performance of such models is often limited by the\nquantity and diversity of high-quality human data. In this paper, we explore\nwhether we can go beyond human data on tasks where we have access to scalar\nfeedback, for example, on math problems where one can verify correctness. To do\nso, we investigate a simple self-training method based on\nexpectation-maximization, which we call ReST^{EM}, where we (1) generate\nsamples from the model and filter them using binary feedback, (2) fine-tune the\nmodel on these samples, and (3) repeat this process a few times. Testing on\nadvanced MATH reasoning and APPS coding benchmarks using PaLM-2 models, we find\nthat ReST^{EM} scales favorably with model size and significantly surpasses\nfine-tuning only on human data. Overall, our findings suggest self-training\nwith feedback can substantially reduce dependence on human-generated data.","upvotes":29,"discussionId":"6577f1e4ee33d547aebf89c0","ai_summary":"Self-training with scalar feedback using expectation-maximization improves language model performance on tasks with verifiable answers, reducing reliance on human data.","ai_keywords":["self-training","expectation-maximization","scalar feedback","MATH reasoning","APPS coding benchmarks","fine-tuning"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"63f974b0aadb0790bb60cd65","avatarUrl":"/avatars/5ed67e7def00b7d1117a00048d0a6e0f.svg","isPro":false,"fullname":"James Liu","user":"jamesliu23","type":"user"},{"_id":"631e40a4b07bce5f2a7faa38","avatarUrl":"/avatars/999e8f6a820d7145b60a99eb16bd7f35.svg","isPro":false,"fullname":"Ogbogu Kalu","user":"ogkalu","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":"63284f86cbc744f197050300","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63284f86cbc744f197050300/cGbUDe5fn-8A8Jcmz5lre.png","isPro":false,"fullname":"Hoptimizer","user":"bunnycore","type":"user"},{"_id":"612246596d9ce900691744d2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/612246596d9ce900691744d2/9DlHVQDqblKz7QPTA6nDa.jpeg","isPro":false,"fullname":"Edoardo Federici","user":"efederici","type":"user"},{"_id":"6311bca0ae8896941da24e66","avatarUrl":"/avatars/48de64894fc3c9397e26e4d6da3ff537.svg","isPro":false,"fullname":"Fynn Kröger","user":"fynnkroeger","type":"user"},{"_id":"60c8d264224e250fb0178f77","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60c8d264224e250fb0178f77/i8fbkBVcoFeJRmkQ9kYAE.png","isPro":true,"fullname":"Adam Lee","user":"Abecid","type":"user"},{"_id":"65554e5b9518733aca07b9e3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65554e5b9518733aca07b9e3/kkvzuINCuGEuw3VNsfBr7.jpeg","isPro":false,"fullname":"Datorien Anderson","user":"niltheory","type":"user"},{"_id":"630412d57373aacccd88af95","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1670594087059-630412d57373aacccd88af95.jpeg","isPro":true,"fullname":"Yasunori Ozaki","user":"alfredplpl","type":"user"},{"_id":"63c3850f6e6561b339e68f24","avatarUrl":"/avatars/23c06fb21107267a717a833d1f5e2b8e.svg","isPro":false,"fullname":"nyoro","user":"nyoronyoro","type":"user"},{"_id":"6362e3412a84d82a8c92be83","avatarUrl":"/avatars/d26aa485d24d4c915c61029361f401be.svg","isPro":false,"fullname":"Error PDF","user":"Error999","type":"user"},{"_id":"6343f83791049e1bce85373e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1665398834110-noauth.png","isPro":false,"fullname":"Zhang ning","user":"pe65374","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":2}">Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models
Abstract
Self-training with scalar feedback using expectation-maximization improves language model performance on tasks with verifiable answers, reducing reliance on human data.
Fine-tuning language models~(LMs) on human-generated data remains a prevalent practice. However, the performance of such models is often limited by the quantity and diversity of high-quality human data. In this paper, we explore whether we can go beyond human data on tasks where we have access to scalar feedback, for example, on math problems where one can verify correctness. To do so, we investigate a simple self-training method based on expectation-maximization, which we call ReST^{EM}, where we (1) generate samples from the model and filter them using binary feedback, (2) fine-tune the model on these samples, and (3) repeat this process a few times. Testing on advanced MATH reasoning and APPS coding benchmarks using PaLM-2 models, we find that ReST^{EM} scales favorably with model size and significantly surpasses fine-tuning only on human data. Overall, our findings suggest self-training with feedback can substantially reduce dependence on human-generated data.
Community
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper