Selected work

Representative Papers

Exponential Backoff: Meta-Stability and Implicit Admission Control

Sep 29, 2026

This study investigates the long-term dynamic behavior and channel-sharing mechanisms of exponential backoff algorithms in distributed communication. To characterize the algorithm’s evolution over extended timescales, we analyze its metastable oscillations using stochastic process theory and validate our findings through numerical simulations with a custom-developed open-source Python package. For the first time, we theoretically prove and experimentally demonstrate that this algorithm achieves implicit admission control by restricting effective access sources, thereby exhibiting distinct metastable properties. By providing reproducible theoretical tools and source code, this work offers deep insights into the implicit congestion control mechanisms inherent in network protocols.

0 citationsRead paper

Stream Learning: Partition-Fair Gossip Learning Without Tokens

Aug 07, 2026

This work addresses the challenge of efficient and fair model partition exchange in decentralized collaborative training without relying on a central coordinator, token mechanisms, or neighbor metadata. Inspired by peer-to-peer live video streaming, it pioneers an analogy between model partition scheduling and video chunk transmission, introducing a two-stage selection strategy based on local partition age and designing ten token-free streaming learning protocols. Relying solely on local information, the proposed approach substantially simplifies protocol design while achieving performance comparable to PTGL in fault-free settings. Notably, under heterogeneous conditions with 30% permanent failures among high-performance nodes, it improves accuracy by 5.53% on HAR and 5.41% on MNIST tasks.

0 citationsRead paper

Recent publications

Latest Papers

Exponential Backoff: Meta-Stability and Implicit Admission Control

Sep 29, 2026

This study investigates the long-term dynamic behavior and channel-sharing mechanisms of exponential backoff algorithms in distributed communication. To characterize the algorithm’s evolution over extended timescales, we analyze its metastable oscillations using stochastic process theory and validate our findings through numerical simulations with a custom-developed open-source Python package. For the first time, we theoretically prove and experimentally demonstrate that this algorithm achieves implicit admission control by restricting effective access sources, thereby exhibiting distinct metastable properties. By providing reproducible theoretical tools and source code, this work offers deep insights into the implicit congestion control mechanisms inherent in network protocols.

0 citationsRead paper

Stream Learning: Partition-Fair Gossip Learning Without Tokens

Aug 07, 2026

This work addresses the challenge of efficient and fair model partition exchange in decentralized collaborative training without relying on a central coordinator, token mechanisms, or neighbor metadata. Inspired by peer-to-peer live video streaming, it pioneers an analogy between model partition scheduling and video chunk transmission, introducing a two-stage selection strategy based on local partition age and designing ten token-free streaming learning protocols. Relying solely on local information, the proposed approach substantially simplifies protocol design while achieving performance comparable to PTGL in fault-free settings. Notably, under heterogeneous conditions with 30% permanent failures among high-performance nodes, it improves accuracy by 5.53% on HAR and 5.41% on MNIST tasks.

0 citationsRead paper