paper-with-me

Papers

Reasoning-preserved Efficient Distillation of Large Language Models via Activation-aware Initialization

2026-05-28 · Junlin He, Yihong Tang, Tong Nie, Guilong Li, Binyu Yang, Jinxiao Du, Lijun Sun, Wei Ma arxiv

Efficient Distillation (EDistill) compresses large language models (LLMs) by structured pruning parameters and tuning lightweight modules with high training efficiency. Although these EDistilled LLMs achieve state-of-the-art (SOTA) performance on general ability benchmarks relative to similarly sized LLMs, we identify a severe degradation in their multi-step reasoning ability, which we term reasoning collapse. We systematically analyze the geometric origins of reasoning collapse and show that the SOTA EDistill method based on width-reducing projection matrices suffers from eRank collapse, in which the effective rank (eRank) of hidden representations drops. We theoretically explain how singular values of randomly initialized projection matrices become unevenly distributed, leading to eRank collapse and thus token indistinguishability. To address this issue, we propose RED (Reasoning-preserved Efficient Distillation) for LLMs, which introduces activation-aware initialization to initialize projection matrices as channel-selection matrices, thus theoretically mitigating eRank collapse. Experiments on Llama and Qwen series demonstrate that RED substantially recovers reasoning while maintaining high training efficiency and SOTA general ability.

📄 PDF Abstract BibTeX arXiv:2605.29327

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

UniComp: A Unified Evaluation of Large Language Model Compression via Pruning, Quantization, and Distillation

2026-02-09 · Jonathan von Rad, Yong Cao, Andreas Geiger arxiv

Model compression is increasingly essential for deploying large language models (LLMs), yet existing comparative studies largely focus on pruning and quantization evaluated primarily on knowledge-centric benchmarks. Thus…

Knowledge DistillationModel Compression

CRISP: Compressed Reasoning via Iterative Self-Policy Distillation

2026-03-05 · Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He 외 arxiv

Reasoning models often generate far more tokens than a task requires, which raises inference cost and can compound errors. We introduce CRISP (Compressed Reasoning via Iterative Self-Policy Distillation), an on-policy se…

ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance

2026-04-14 · Daniil Gurgurov, Tom Röhr, Sebastian von Rohrscheidt, Josef van Genabith 외 arxiv

Despite advances in multilingual capabilities, most large language models (LLMs) remain English-centric in their training and, crucially, in their production of reasoning traces. Even when tasked with non-English problem…

Reinforcement LearningCross-Lingual TransferGeneral Knowledge

Task-agnostic Distillation of Encoder-Decoder Language Models

2023-05-21 · Chen Zhang, Yang Yang, Jingang Wang, Dawei Song

Finetuning pretrained language models (LMs) have enabled appealing performance on a diverse array of tasks. The intriguing task-agnostic property has driven a shifted focus from task-specific to task-agnostic distillatio…

Abstractive Text SummarizationDecoder

Entropy-Preserving Supervised Fine-Tuning via Adaptive Self-Distillation for Large Reasoning Models

2026-02-02 · Hao Wang, Hao Gu, Hongming Piao, Kaixiong Gong 외 arxiv

The standard post-training recipe for large reasoning models, supervised fine-tuning followed by reinforcement learning (SFT-then-RL), may limit the benefits of the RL stage: while SFT imitates expert demonstrations, it …

Reinforcement LearningMathematical Reasoning