Hugging Face Daily Papers · · 5 min read

Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.</p>\n","updatedAt":"2026-06-19T08:00:45.904Z","author":{"_id":"6426781213a9e5d9675a1d57","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/vR7u0azaShh5Tj_OqdP2k.png","fullname":"ZHONGZHU ZHOU","name":"Zhongzhu","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9008266925811768},"editors":["Zhongzhu"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/vR7u0azaShh5Tj_OqdP2k.png"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.16429","authors":[{"_id":"6a34f72c4c5c5e0d69bf1dc6","name":"Zhongzhu Zhou","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dc7","name":"Qingyang Wu","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dc8","name":"Junxiong Wang","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dc9","name":"Mayank Mishra","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dca","name":"Shuaiwen Leon Song","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dcb","name":"Ben Athiwaratkun","hidden":false},{"_id":"6a34f72c4c5c5e0d69bf1dcc","name":"Chenfeng Xu","hidden":false}],"publishedAt":"2026-06-15T00:00:00.000Z","submittedOnDailyAt":"2026-06-19T00:00:00.000Z","title":"Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation","submittedOnDailyBy":{"_id":"6426781213a9e5d9675a1d57","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/vR7u0azaShh5Tj_OqdP2k.png","isPro":false,"fullname":"ZHONGZHU ZHOU","user":"Zhongzhu","type":"user","name":"Zhongzhu"},"summary":"Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.","upvotes":1,"discussionId":"6a34f72d4c5c5e0d69bf1dcd","githubRepo":"https://github.com/FutureMLS-Lab/Taylor-Calibrate","githubRepoAddedBy":"user","ai_summary":"Hybrid linear attention models can be improved through a novel initialization technique that enhances conversion from pretrained Transformers by leveraging teacher attention statistics and alignment steps.","ai_keywords":["hybrid linear attention models","full softmax attention","Gated DeltaNet","teacher-student learning","Taylor-guided initialization","memory timescale","write gates","output gate","distillation tokens","attention projections"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":1,"organization":{"_id":"632b803bb2dd35f135623cc2","name":"togethercomputer","fullname":"Together","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67be52be391885e452f4f1be/XiCaXuCIs-N_03X_nzopf.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6426781213a9e5d9675a1d57","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/vR7u0azaShh5Tj_OqdP2k.png","isPro":false,"fullname":"ZHONGZHU ZHOU","user":"Zhongzhu","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"632b803bb2dd35f135623cc2","name":"togethercomputer","fullname":"Together","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67be52be391885e452f4f1be/XiCaXuCIs-N_03X_nzopf.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.16429.md","query":{}}">
Papers
arxiv:2606.16429

Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation

Published on Jun 15
· Submitted by
ZHONGZHU ZHOU
on Jun 19
Authors:
,
,
,
,
,
,

Abstract

Hybrid linear attention models can be improved through a novel initialization technique that enhances conversion from pretrained Transformers by leveraging teacher attention statistics and alignment steps.

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.

Community

Paper submitter about 2 hours ago

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images

· Sign up or log in to comment

Get this paper in your agent:

hf papers read 2606.16429
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2606.16429 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2606.16429 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2606.16429 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection 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.

More from Hugging Face Daily Papers