Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Forgetting: A New Mechanism Towards Better Large Language Model Fine-tuning (2025)
Mitigating Catastrophic Forgetting in Large Language Models with Forgetting-aware Pruning (2025)
Select to Know: An Internal-External Knowledge Self-Selection Framework for Domain-Specific Question Answering (2025)
Not All Parameters Are Created Equal: Smart Isolation Boosts Fine-Tuning Performance (2025)
Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning (2025)
Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (2025)
Can Smaller LLMs do better? Unlocking Cross-Domain Potential through Parameter-Efficient Fine-Tuning for Text Summarization (2025)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend

\n","updatedAt":"2025-09-24T01:35:56.743Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":264}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7113507390022278},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2509.16596","authors":[{"_id":"68d205331ca7156988a8ecbd","user":{"_id":"66384be673c2c55f2ded89fa","avatarUrl":"/avatars/1d8721074f0f51fab405f81474f2035f.svg","isPro":false,"fullname":"Junjie Ye","user":"Junjie-Ye","type":"user"},"name":"Junjie Ye","status":"claimed_verified","statusLastChangedAt":"2025-09-23T02:40:25.849Z","hidden":false},{"_id":"68d205331ca7156988a8ecbe","user":{"_id":"655c6b1abfb531437a54c0e6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/V8Md2mMX83hrSowKk6qMS.jpeg","isPro":false,"fullname":"Yuming Yang","user":"Umean","type":"user"},"name":"Yuming Yang","status":"claimed_verified","statusLastChangedAt":"2025-09-23T10:07:02.770Z","hidden":false},{"_id":"68d205331ca7156988a8ecbf","name":"Yang Nan","hidden":false},{"_id":"68d205331ca7156988a8ecc0","name":"Shuo Li","hidden":false},{"_id":"68d205331ca7156988a8ecc1","name":"Qi Zhang","hidden":false},{"_id":"68d205331ca7156988a8ecc2","name":"Tao Gui","hidden":false},{"_id":"68d205331ca7156988a8ecc3","name":"Xuanjing Huang","hidden":false},{"_id":"68d205331ca7156988a8ecc4","name":"Peng Wang","hidden":false},{"_id":"68d205331ca7156988a8ecc5","name":"Zhongchao Shi","hidden":false},{"_id":"68d205331ca7156988a8ecc6","name":"Jianping Fan","hidden":false}],"publishedAt":"2025-09-20T09:40:32.000Z","submittedOnDailyAt":"2025-09-23T00:56:48.151Z","title":"Analyzing the Effects of Supervised Fine-Tuning on Model Knowledge from\n Token and Parameter Levels","submittedOnDailyBy":{"_id":"655c6b1abfb531437a54c0e6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/V8Md2mMX83hrSowKk6qMS.jpeg","isPro":false,"fullname":"Yuming Yang","user":"Umean","type":"user"},"summary":"Large language models (LLMs) acquire substantial world knowledge during\npre-training, which is further shaped by post-training techniques such as\nsupervised fine-tuning (SFT). However, the impact of SFT on a model's knowledge\nremains underexplored, limiting our ability to control knowledge change\nbehavior in fine-tuned models. To address this gap, we evaluate closed-book\nquestion answering (CBQA) performance across five LLMs from the LLaMA-2 and\nLLaMA-3 families. Surprisingly, models fine-tuned on 1,920 samples perform up\nto 14% worse than those fine-tuned on only 240 samples. Furthermore, varying\nthe level of knowledge mastery in the fine-tuning data leads to performance\nfluctuations of over 12%. To investigate these effects, we analyze model\nbehavior at both the token and parameter levels. Our analysis reveals that up\nto 90% of parameter updates during SFT do not contribute to knowledge\nenhancement. Restoring these updates can improve performance on the CBQA task,\ndepending on the characteristics of the fine-tuning data. These insights offer\npractical guidance for developing fine-tuning strategies that more effectively\nstrengthen model knowledge.","upvotes":12,"discussionId":"68d205331ca7156988a8ecc7","ai_summary":"Supervised fine-tuning of large language models can negatively impact closed-book question answering performance, with up to 90% of parameter updates not contributing to knowledge enhancement.","ai_keywords":["large language models","pre-training","supervised fine-tuning","closed-book question answering","token","parameter updates","knowledge enhancement"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"655c6b1abfb531437a54c0e6","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/V8Md2mMX83hrSowKk6qMS.jpeg","isPro":false,"fullname":"Yuming Yang","user":"Umean","type":"user"},{"_id":"66384be673c2c55f2ded89fa","avatarUrl":"/avatars/1d8721074f0f51fab405f81474f2035f.svg","isPro":false,"fullname":"Junjie Ye","user":"Junjie-Ye","type":"user"},{"_id":"66f68f5f9884fa00e9150a2c","avatarUrl":"/avatars/e94fcf78bffba6705cb0300236dfc71e.svg","isPro":false,"fullname":"Nan","user":"Sirius518","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":"65f8dfbc03a21601bf0b3a5c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f8dfbc03a21601bf0b3a5c/sGhdEf266p7wUlT0M0xbv.png","isPro":false,"fullname":"hienhq","user":"wanhin","type":"user"},{"_id":"66d8512c54209e9101811e8e","avatarUrl":"/avatars/62dfd8e6261108f2508efe678d5a2a57.svg","isPro":false,"fullname":"M Saad Salman","user":"MSS444","type":"user"},{"_id":"65c20ee58aedd6edd2b89000","avatarUrl":"/avatars/db1bbf4c8f6a88459da967ec83e9bc08.svg","isPro":false,"fullname":"Chmielewski","user":"Eryk-Chmielewski","type":"user"},{"_id":"648eb1eb59c4e5c87dc116e0","avatarUrl":"/avatars/c636cea39c2c0937f01398c94ead5dad.svg","isPro":false,"fullname":"fdsqefsgergd","user":"T-representer","type":"user"},{"_id":"61b839889f7cfeae618e72c9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/61b839889f7cfeae618e72c9/5kFRCChdqwv7MGM8T_y5v.jpeg","isPro":false,"fullname":"Raja Biswas","user":"rbiswasfc","type":"user"},{"_id":"62447e04f555de1927a9c879","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1648655841478-noauth.png","isPro":false,"fullname":"jasonjiang","user":"mikinyaa","type":"user"},{"_id":"68939f5ddfd414bbd5aafa12","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/_rPJDe2EnQOyySwbEmdJg.png","isPro":false,"fullname":"Stephen Hope","user":"steve-hope","type":"user"},{"_id":"68d23240e133063ab3c58271","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/5eybL8giYD8K_NcEJLbqV.png","isPro":false,"fullname":"Mykola Martych","user":"Mykola5689","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">

Papers

arxiv:2509.16596

Analyzing the Effects of Supervised Fine-Tuning on Model Knowledge from Token and Parameter Levels

Published on Sep 20

· Submitted by

Yuming Yang on Sep 23

Upvote

Authors:

Junjie Ye ,

Yuming Yang ,

Abstract

Supervised fine-tuning of large language models can negatively impact closed-book question answering performance, with up to 90% of parameter updates not contributing to knowledge enhancement.

AI-generated summary

Large language models (LLMs) acquire substantial world knowledge during pre-training, which is further shaped by post-training techniques such as supervised fine-tuning (SFT). However, the impact of SFT on a model's knowledge remains underexplored, limiting our ability to control knowledge change behavior in fine-tuned models. To address this gap, we evaluate closed-book question answering (CBQA) performance across five LLMs from the LLaMA-2 and LLaMA-3 families. Surprisingly, models fine-tuned on 1,920 samples perform up to 14% worse than those fine-tuned on only 240 samples. Furthermore, varying the level of knowledge mastery in the fine-tuning data leads to performance fluctuations of over 12%. To investigate these effects, we analyze model behavior at both the token and parameter levels. Our analysis reveals that up to 90% of parameter updates during SFT do not contribute to knowledge enhancement. Restoring these updates can improve performance on the CBQA task, depending on the characteristics of the fine-tuning data. These insights offer practical guidance for developing fine-tuning strategies that more effectively strengthen model knowledge.

View arXiv page View PDF Add to collection