Simpler beats complicated</p>\n","updatedAt":"2026-08-31T14:23:31.027Z","author":{"_id":"62f8ea1177b722f186611e8e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1660479589894-noauth.jpeg","fullname":"Alexia Jolicoeur-Martineau","name":"AlexiaJM","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":95,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.9421804547309875},"editors":["AlexiaJM"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1660479589894-noauth.jpeg"],"reactions":[{"reaction":"🔥","users":["travisking"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.28444","authors":[{"_id":"6a95859294b7be1064d0f46a","name":"Alexia Jolicoeur-Martineau","hidden":false},{"_id":"6a95859294b7be1064d0f46b","name":"Rhea Sanjay Sukthanker","hidden":false},{"_id":"6a95859294b7be1064d0f46c","name":"Pashmina Cameron","hidden":false},{"_id":"6a95859294b7be1064d0f46d","name":"Emy Gervais","hidden":false}],"publishedAt":"2026-08-28T00:00:00.000Z","submittedOnDailyAt":"2026-08-31T00:00:00.000Z","title":"Sliding-window beats linear attention","submittedOnDailyBy":{"_id":"62f8ea1177b722f186611e8e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1660479589894-noauth.jpeg","isPro":false,"fullname":"Alexia Jolicoeur-Martineau","user":"AlexiaJM","type":"user","name":"AlexiaJM"},"summary":"Due to the nature of quadratic attention, Large Language Models (LLMs) consume a lot of memory and energy. Every new token costs more than the previous one. For each additional token, the keys and values must be stored in memory indefinitely, which is unsustainable.\n Several alternatives have been proposed to fix the quadratic scaling problem, one of which is retrofitting LLMs to use Linear Attention. This idea has attracted a lot of attention, given its promise to solve the quadratic scaling problem with state-of-the-art performance at low cost. However, this line of research has not been properly compared to simpler baselines.\n In this work, we show that Sliding Window Attention (SWA) with sinks performs as well or better than post-trained Linear Attention models. We observe this across multiple LLMs on various downstream tasks. For long-context reasoning tasks (Needle-in-a-Haystack and BABILong), SWA achieves massively higher performance (2 to 10 times higher than linear attention). SWA requires no post-training, is extremely fast, and requires low memory; therefore, making it an extremely cheap and reliable solution.\n To reduce inference memory cost, we strongly recommend switching to SWA instead of post-training linear models. Linear attention models may have shown some promise, but they likely require to be trained from scratch or extensive post-training in order to even match SWA.","upvotes":9,"discussionId":"6a95859294b7be1064d0f46e","ai_summary":"Sliding window attention with sinks outperforms post-trained linear attention on long-context tasks without requiring retraining, offering a cheaper and more reliable inference solution.","ai_keywords":["Linear Attention","Sliding Window Attention","attention sinks","quadratic attention","Needle-in-a-Haystack","BABILong","long-context reasoning"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"62f8ea1177b722f186611e8e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1660479589894-noauth.jpeg","isPro":false,"fullname":"Alexia Jolicoeur-Martineau","user":"AlexiaJM","type":"user"},{"_id":"677ab8ea25dae52ac487bc7f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/NbmYdSc_xP0bhJzg-leRt.png","isPro":true,"fullname":"Lucas Müller","user":"gustavzudeml","type":"user"},{"_id":"6186ddf6a7717cb375090c01","avatarUrl":"/avatars/716b6a7d1094c8036b2a8a7b9063e8aa.svg","isPro":true,"fullname":"Julien BLANCHON","user":"blanchon","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":"63284f86cbc744f197050300","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63284f86cbc744f197050300/cGbUDe5fn-8A8Jcmz5lre.png","isPro":false,"fullname":"Hoptimizer","user":"bunnycore","type":"user"},{"_id":"661e8e57ebe3616a1b084101","avatarUrl":"/avatars/b72ed568a97b147b54339a5c26185f71.svg","isPro":false,"fullname":"Travis King","user":"travisking","type":"user"},{"_id":"624bebf604abc7ebb01789af","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1649143001781-624bebf604abc7ebb01789af.jpeg","isPro":true,"fullname":"Apolinário from multimodal AI art","user":"multimodalart","type":"user"},{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"5e6485f787403103f9f1055e","name":"microsoft","fullname":"Microsoft","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.28444.md","query":{}}">
Sliding-window beats linear attention
Abstract
Sliding window attention with sinks outperforms post-trained linear attention on long-context tasks without requiring retraining, offering a cheaper and more reliable inference solution.
Due to the nature of quadratic attention, Large Language Models (LLMs) consume a lot of memory and energy. Every new token costs more than the previous one. For each additional token, the keys and values must be stored in memory indefinitely, which is unsustainable.
Several alternatives have been proposed to fix the quadratic scaling problem, one of which is retrofitting LLMs to use Linear Attention. This idea has attracted a lot of attention, given its promise to solve the quadratic scaling problem with state-of-the-art performance at low cost. However, this line of research has not been properly compared to simpler baselines.
In this work, we show that Sliding Window Attention (SWA) with sinks performs as well or better than post-trained Linear Attention models. We observe this across multiple LLMs on various downstream tasks. For long-context reasoning tasks (Needle-in-a-Haystack and BABILong), SWA achieves massively higher performance (2 to 10 times higher than linear attention). SWA requires no post-training, is extremely fast, and requires low memory; therefore, making it an extremely cheap and reliable solution.
To reduce inference memory cost, we strongly recommend switching to SWA instead of post-training linear models. Linear attention models may have shown some promise, but they likely require to be trained from scratch or extensive post-training in order to even match SWA.
Community
Simpler beats complicated
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.28444 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.28444 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.28444 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.