dorsal/arxiv
View SchemaDistribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning
| Authors | Shaotian Yan, Kaiyuan Liu, Chen Shen, Bing Wang, Sinan Fan, Jun Zhang, Yue Wu, Zheng Wang, Jieping Ye |
|---|---|
| Categories | |
| ArXiv ID | 2601.09088vv1 |
| URL | https://arxiv.org/abs/2601.09088 |
| License | http://creativecommons.org/licenses/by/4.0/ |
Abstract
In this report, we introduce DASD-4B-Thinking, a lightweight yet highly capable, fully open-source reasoning model. It achieves SOTA performance among open-source models of comparable scale across challenging benchmarks in mathematics, scientific reasoning, and code generation -- even outperforming several larger models. We begin by critically reexamining a widely adopted distillation paradigm in the community: SFT on teacher-generated responses, also known as sequence-level distillation. Although a series of recent works following this scheme have demonstrated remarkable efficiency and strong empirical performance, they are primarily grounded in the SFT perspective. Consequently, these approaches focus predominantly on designing heuristic rules for SFT data filtering, while largely overlooking the core principle of distillation itself -- enabling the student model to learn the teacher's full output distribution so as to inherit its generalization capability. Specifically, we identify three critical limitations in current practice: i) Inadequate representation of the teacher's sequence-level distribution; ii) Misalignment between the teacher's output distribution and the student's learning capacity; and iii) Exposure bias arising from teacher-forced training versus autoregressive inference. In summary, these shortcomings reflect a systemic absence of explicit teacher-student interaction throughout the distillation process, leaving the essence of distillation underexploited. To address these issues, we propose several methodological innovations that collectively form an enhanced sequence-level distillation training pipeline. Remarkably, DASD-4B-Thinking obtains competitive results using only 448K training samples -- an order of magnitude fewer than those employed by most existing open-source efforts. To support community research, we publicly release our models and the training dataset.
{
"annotation_id": "bf56ae39-ac93-46bb-a51a-57a7f7bd7f90",
"date_created": "2026-02-17T05:53:20.413000Z",
"date_modified": "2026-02-17T05:53:20.413000Z",
"file_hash": "bc816cd0f86becf6972351070fb48b9d8e597b6a14ef765ac7fa13eeaab8cc3b",
"private": false,
"record": {
"abstract": "In this report, we introduce DASD-4B-Thinking, a lightweight yet highly capable, fully open-source reasoning model. It achieves SOTA performance among open-source models of comparable scale across challenging benchmarks in mathematics, scientific reasoning, and code generation -- even outperforming several larger models. We begin by critically reexamining a widely adopted distillation paradigm in the community: SFT on teacher-generated responses, also known as sequence-level distillation. Although a series of recent works following this scheme have demonstrated remarkable efficiency and strong empirical performance, they are primarily grounded in the SFT perspective. Consequently, these approaches focus predominantly on designing heuristic rules for SFT data filtering, while largely overlooking the core principle of distillation itself -- enabling the student model to learn the teacher\u0027s full output distribution so as to inherit its generalization capability. Specifically, we identify three critical limitations in current practice: i) Inadequate representation of the teacher\u0027s sequence-level distribution; ii) Misalignment between the teacher\u0027s output distribution and the student\u0027s learning capacity; and iii) Exposure bias arising from teacher-forced training versus autoregressive inference. In summary, these shortcomings reflect a systemic absence of explicit teacher-student interaction throughout the distillation process, leaving the essence of distillation underexploited. To address these issues, we propose several methodological innovations that collectively form an enhanced sequence-level distillation training pipeline. Remarkably, DASD-4B-Thinking obtains competitive results using only 448K training samples -- an order of magnitude fewer than those employed by most existing open-source efforts. To support community research, we publicly release our models and the training dataset.",
"arxiv_id": "2601.09088",
"authors": [
"Shaotian Yan",
"Kaiyuan Liu",
"Chen Shen",
"Bing Wang",
"Sinan Fan",
"Jun Zhang",
"Yue Wu",
"Zheng Wang",
"Jieping Ye"
],
"categories": [
"cs.LG",
"cs.CL"
],
"license": "http://creativecommons.org/licenses/by/4.0/",
"title": "Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning",
"url": "https://arxiv.org/abs/2601.09088",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "15382404-f382-4034-b18d-db4aef495e58",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}