Selected work

Representative Papers

Finding the Heads and the Neurons Responsible for Network Information Retrieval in Language Models

Oct 06, 2026

This study investigates how language models internally represent network infrastructure information, such as hostname-IP pairs. Through causal ablation and intervention experiments, the authors localize and validate specific attention heads and neurons responsible for this task, evaluating their generalizability via correlation-based ranking and cross-dataset transfer tests. The findings demonstrate that attention-head-level causal mechanisms exhibit cross-architecture universality, with full-head detectors achieving 99.5%–100% accuracy across multiple models. Conversely, neuron-level responsibility distributions are shown to be model-specific, limiting the transferability of single-neuron approaches and necessitating per-model validation.

0 citationsRead paper

Perspectives on Sustainable Computational Science and Engineering

Sep 24, 2026

This study addresses the longstanding oversight in Computational Science and Engineering (CSE), which has predominantly been regarded as an enabler of external sustainability while its own internal sustainability remains critically neglected. To bridge this gap, this work establishes sustainability as a core design principle for CSE by constructing a dual analytical framework encompassing both internal and external dimensions, anchored by two pillars: sustainable computing and sustainable software. The methodology integrates systems engineering paradigms, interdisciplinary analysis, and case studies, substantiated through evaluation frameworks from software engineering and high-performance computing. Ultimately, this research advances a theoretical perspective on sustainable CSE, formulates best practice guidelines alongside implementation recommendations for stakeholders, and promotes the development of quantifiable sustainability assessment standards to systematically guide the field toward enduring ecological and operational viability.

0 citationsRead paper

Mental Model Management: An Operator-Based Framework for LLM Memory

Aug 15, 2026

This study addresses the challenge of lacking compact and dynamically evolving conceptual representations in large language models (LLMs) by proposing the 3M framework. This approach models knowledge as compact mental models and introduces innovative operator mechanisms, including chunking, extractive retrieval, consistency verification, and evolution, to enable continuous knowledge integration and dynamic reorganization. The research effectively overcomes memory management bottlenecks in LLMs by achieving dynamic knowledge evolution while maintaining representational compactness. Consequently, this method significantly enhances the model's knowledge reasoning capabilities and establishes a novel paradigm for constructing intelligent systems equipped with adaptive memory.

0 citationsRead paper

Adaptive Search in Collatz Exponent-Code Space via 2-adic and 3-adic Constraints

Jul 10, 2026

This study investigates structural obstacles to orbit convergence in the Collatz conjecture, aiming to characterize essential features of potential counterexamples. It introduces the “2-3-∞ diagnostic framework,” which for the first time jointly incorporates 2-adic initial conditions and 3-adic terminal constraints within a finite-index symbolic space derived from accelerated mappings. By integrating real-valued drift metrics, asymptotic residue theory, and adaptive evolutionary search, the work demonstrates that any counterexample must exhibit near-critical drift and small residue. Moreover, it proves that the asymptotic residue rate of integer-generated codes is zero. Experimental results over orbit lengths from 100 to 400 significantly improve the performance trade-off at finite lengths, with all configurations maintaining a positive residue rate, thereby validating the framework’s effectiveness and novelty.

0 citationsRead paper

WOLF-VLA: Whole-Body Humanoid Optimal Locomotion Framework for Vision-Language-Action Learning

Jun 24, 2026

This work addresses key challenges in applying vision–language–action (VLA) models to whole-body, contact-rich humanoid robot control—namely, data scarcity, dynamically inconsistent demonstrations, and the difficulty of simultaneously ensuring optimality and safety. To overcome these limitations, we propose an end-to-end learning framework that integrates whole-body optimal control with large-scale multimodal data. We introduce, for the first time, a dynamically consistent multimodal dataset of whole-body humanoid motions, enabling direct generation of robust, safe, and high-performance motor policies from natural language instructions. Experimental results demonstrate that the learned policies exhibit strong generalization across diverse tasks and environmental conditions, robustness to variations in initial states, and state-of-the-art performance across multiple metrics, while establishing a reproducible benchmark for future research.

0 citationsRead paper

Recent publications

Latest Papers

Finding the Heads and the Neurons Responsible for Network Information Retrieval in Language Models

Oct 06, 2026

This study investigates how language models internally represent network infrastructure information, such as hostname-IP pairs. Through causal ablation and intervention experiments, the authors localize and validate specific attention heads and neurons responsible for this task, evaluating their generalizability via correlation-based ranking and cross-dataset transfer tests. The findings demonstrate that attention-head-level causal mechanisms exhibit cross-architecture universality, with full-head detectors achieving 99.5%–100% accuracy across multiple models. Conversely, neuron-level responsibility distributions are shown to be model-specific, limiting the transferability of single-neuron approaches and necessitating per-model validation.

0 citationsRead paper

Perspectives on Sustainable Computational Science and Engineering

Sep 24, 2026

This study addresses the longstanding oversight in Computational Science and Engineering (CSE), which has predominantly been regarded as an enabler of external sustainability while its own internal sustainability remains critically neglected. To bridge this gap, this work establishes sustainability as a core design principle for CSE by constructing a dual analytical framework encompassing both internal and external dimensions, anchored by two pillars: sustainable computing and sustainable software. The methodology integrates systems engineering paradigms, interdisciplinary analysis, and case studies, substantiated through evaluation frameworks from software engineering and high-performance computing. Ultimately, this research advances a theoretical perspective on sustainable CSE, formulates best practice guidelines alongside implementation recommendations for stakeholders, and promotes the development of quantifiable sustainability assessment standards to systematically guide the field toward enduring ecological and operational viability.

0 citationsRead paper

Mental Model Management: An Operator-Based Framework for LLM Memory

Aug 15, 2026

This study addresses the challenge of lacking compact and dynamically evolving conceptual representations in large language models (LLMs) by proposing the 3M framework. This approach models knowledge as compact mental models and introduces innovative operator mechanisms, including chunking, extractive retrieval, consistency verification, and evolution, to enable continuous knowledge integration and dynamic reorganization. The research effectively overcomes memory management bottlenecks in LLMs by achieving dynamic knowledge evolution while maintaining representational compactness. Consequently, this method significantly enhances the model's knowledge reasoning capabilities and establishes a novel paradigm for constructing intelligent systems equipped with adaptive memory.

0 citationsRead paper

Adaptive Search in Collatz Exponent-Code Space via 2-adic and 3-adic Constraints

Jul 10, 2026

This study investigates structural obstacles to orbit convergence in the Collatz conjecture, aiming to characterize essential features of potential counterexamples. It introduces the “2-3-∞ diagnostic framework,” which for the first time jointly incorporates 2-adic initial conditions and 3-adic terminal constraints within a finite-index symbolic space derived from accelerated mappings. By integrating real-valued drift metrics, asymptotic residue theory, and adaptive evolutionary search, the work demonstrates that any counterexample must exhibit near-critical drift and small residue. Moreover, it proves that the asymptotic residue rate of integer-generated codes is zero. Experimental results over orbit lengths from 100 to 400 significantly improve the performance trade-off at finite lengths, with all configurations maintaining a positive residue rate, thereby validating the framework’s effectiveness and novelty.

0 citationsRead paper

WOLF-VLA: Whole-Body Humanoid Optimal Locomotion Framework for Vision-Language-Action Learning

Jun 24, 2026

This work addresses key challenges in applying vision–language–action (VLA) models to whole-body, contact-rich humanoid robot control—namely, data scarcity, dynamically inconsistent demonstrations, and the difficulty of simultaneously ensuring optimality and safety. To overcome these limitations, we propose an end-to-end learning framework that integrates whole-body optimal control with large-scale multimodal data. We introduce, for the first time, a dynamically consistent multimodal dataset of whole-body humanoid motions, enabling direct generation of robust, safe, and high-performance motor policies from natural language instructions. Experimental results demonstrate that the learned policies exhibit strong generalization across diverse tasks and environmental conditions, robustness to variations in initial states, and state-of-the-art performance across multiple metrics, while establishing a reproducible benchmark for future research.

0 citationsRead paper