BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152640Z
LOCATION:
DTSTART;TZID=America/Los_Angeles:20260728T174800
DTEND;TZID=America/Los_Angeles:20260728T174900
UID:dac_DAC 2026_sess317_RESEARCH2650@linklings.com
SUMMARY:Think Before You Prune: Self-Reflective Structured Pruning for Rea
 soning Language Models
DESCRIPTION:Ziyan Wang (University of North Carolina at Charlotte); Enmao 
 Diao (DreamSoul); Qi Le (University of Minnesota); Pu Wang, Guanchu Wang, 
 and Minwoo Lee (University of North Carolina at Charlotte); Shu-ping Yeh (
 Intel); and Li Yang (University of North Carolina at Charlotte)\n\nReasoni
 ng LLMs (RLMs) such as OpenAI o1, DeepSeek-R1, and Qwen3 deliver strong mu
 lti-step reasoning through chain-of-thought generation, but their large mo
 del sizes and lengthy decode-time outputs make them costly to deploy and u
 nsuitable for resource-constrained settings. To reduce computing and memor
 y cost, pruning offers a promising solution by removing unimportant parame
 ters. However, despite their success on standard LLMs, existing pruning me
 thods severely damage RLMs, as even moderate sparsity (e.g., 20%) can coll
 apse accuracy and completely disrupt the model's reasoning coherence. We b
 egin by analyzing why existing pruning pipelines fail on reasoning LLMs an
 d find that their brittleness largely stems from a mismatch between the ca
 libration data, the pruning objective, and the model's decode-time reasoni
 ng behavior. Our study further shows that the most reliable calibration si
 gnal comes not from human-written labels but from the model's own self-gen
 erated reasoning traces, which more accurately reflect its inference distr
 ibution. Guided by these insights, we introduce RESP, a self-reflective st
 ructured pruning framework that aligns pruning decisions with the model's 
 reasoning dynamics through self-generated calibration, decode-only gradien
 t-based importance estimation, and progressive regeneration that maintains
  calibration fidelity as sparsity increases. Experiments on Qwen3-8B demon
 strate that RESP markedly outperforms existing structured pruning methods 
 on both GSM8K and MathQA, preserving near-dense accuracy at 20–30% sparsit
 y and substantially mitigating performance collapse at higher sparsity lev
 els. At 40% sparsity, RESP attains 81.3% accuracy on GSM8K and 59.6% on Ma
 thQA, surpassing the strongest baselines by 66.87% and 47%, respectively.\
 n\nTrack: EDA7-II. Physical Design and Verification, Design\n\nSession Cha
 irs: Christian Pilato (Politecnico di Milano); Mohsen Imani (University of
  California, Irvine); Flavio Ponzina (San Diego State University); Amin Fa
 rshidi (Cadence Design Systems, Inc.); Yonggan Fu (Georgia Institute of Te
 chnology, NVIDIA); Dimitrios Tsaras (Huawei, Hong Kong University of Scien
 ce and Technology (HKUST)); Anh Nguyen (Fujitsu Limited); Ruiyang Qin (Vil
 lanova University); Jiliang Zhang (Hunan University); Meng Li (Peking Univ
 ersity); and Rassul Bairamkulov (Advanced Micro Devices (AMD))\n\n
END:VEVENT
END:VCALENDAR