Institution profile

Dankook University

Academic institutionasia · kr
Official website
Research library14linked papers
Opportunities0open roles
Selected work

Representative Papers

Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning

Oct 07, 2026

This study addresses the unclear mechanisms underlying skeleton language selection in multilingual mathematical reasoning and the lack of systematic research beyond English-centric settings. We propose a language-aware skeleton exploration framework that integrates greedy decoding, multi-trajectory evaluation, translation ablation, and cross-benchmark validation to systematically analyze skeleton language effects across varying model scales and linguistic conditions. Our findings reveal that the skeleton language is fundamentally a context-dependent design variable rather than a fixed optimal choice, identifying three distinct patterns of inconsistent language effects. Furthermore, we demonstrate that English skeletons confer only marginal advantages for smaller models operating in low-resource languages and are not universally optimal. This work provides new perspectives for optimizing multilingual reasoning systems by challenging the default assumption favoring English as the skeleton language.

0 citationsRead paper

How Linear Attention Remembers

Sep 27, 2026

This study addresses the memory interference and capacity bottlenecks inherent in the fixed-size recurrent states of linear attention mechanisms. Through analytical decomposition and causal intervention, we systematically investigate the writing, retention, and retrieval dynamics in models such as Gated Linear Attention (GLA), providing the first quantification of cross-fact causal coupling. Our findings reveal that interference stems primarily from overlapping subsequent writes rather than temporal decay, challenging conventional understandings of KV caching. Furthermore, we demonstrate that both recall and editing performance degrade under high cognitive load, and identify that full-attention layers predominantly govern real-time retrieval within hybrid architectures. Ultimately, this work delineates the memory boundaries and selective recall principles of linear attention, offering critical insights for advancing state-space models.

0 citationsRead paper

Metacognitive Selective Ensemble for Mobile Systems

Sep 25, 2026

This study addresses the high execution costs of deep ensemble models in mobile sensing and the additional computational overhead incurred by adaptive selection methods that must evaluate inactive candidates. To overcome these limitations, this work proposes MetaSE (Metacognitive Selective Ensemble), a framework that maintains a small active subset by exploiting the short-term persistence of model reliability. Based on posterior evidence, MetaSE dynamically removes unreliable members and triggers lightweight routing for their replacement. This stateful design enables efficient proactive ensemble inference by accessing the diversity of a large model pool without requiring full-pool evaluation. Experimental results on human activity recognition (HAR) datasets demonstrate that MetaSE achieves accuracy comparable to full ensembles while delivering a 2.7× inference speedup and reducing memory footprint by 69% during edge deployment on a Raspberry Pi 4B.

0 citationsRead paper
Recent publications

Latest Papers

Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning

Oct 07, 2026

This study addresses the unclear mechanisms underlying skeleton language selection in multilingual mathematical reasoning and the lack of systematic research beyond English-centric settings. We propose a language-aware skeleton exploration framework that integrates greedy decoding, multi-trajectory evaluation, translation ablation, and cross-benchmark validation to systematically analyze skeleton language effects across varying model scales and linguistic conditions. Our findings reveal that the skeleton language is fundamentally a context-dependent design variable rather than a fixed optimal choice, identifying three distinct patterns of inconsistent language effects. Furthermore, we demonstrate that English skeletons confer only marginal advantages for smaller models operating in low-resource languages and are not universally optimal. This work provides new perspectives for optimizing multilingual reasoning systems by challenging the default assumption favoring English as the skeleton language.

0 citationsRead paper

How Linear Attention Remembers

Sep 27, 2026

This study addresses the memory interference and capacity bottlenecks inherent in the fixed-size recurrent states of linear attention mechanisms. Through analytical decomposition and causal intervention, we systematically investigate the writing, retention, and retrieval dynamics in models such as Gated Linear Attention (GLA), providing the first quantification of cross-fact causal coupling. Our findings reveal that interference stems primarily from overlapping subsequent writes rather than temporal decay, challenging conventional understandings of KV caching. Furthermore, we demonstrate that both recall and editing performance degrade under high cognitive load, and identify that full-attention layers predominantly govern real-time retrieval within hybrid architectures. Ultimately, this work delineates the memory boundaries and selective recall principles of linear attention, offering critical insights for advancing state-space models.

0 citationsRead paper

Metacognitive Selective Ensemble for Mobile Systems

Sep 25, 2026

This study addresses the high execution costs of deep ensemble models in mobile sensing and the additional computational overhead incurred by adaptive selection methods that must evaluate inactive candidates. To overcome these limitations, this work proposes MetaSE (Metacognitive Selective Ensemble), a framework that maintains a small active subset by exploiting the short-term persistence of model reliability. Based on posterior evidence, MetaSE dynamically removes unreliable members and triggers lightweight routing for their replacement. This stateful design enables efficient proactive ensemble inference by accessing the diversity of a large model pool without requiring full-pool evaluation. Experimental results on human activity recognition (HAR) datasets demonstrate that MetaSE achieves accuracy comparable to full ensembles while delivering a 2.7× inference speedup and reducing memory footprint by 69% during edge deployment on a Raspberry Pi 4B.

0 citationsRead paper