TY - GEN
T1 - DRLLMS
T2 - 36th Workshop on Network and Operating System Support for Digital Audio and Video, NOSSDAV 2026
AU - Lyu, Tao
AU - Zhang, Cong
AU - Duan, Haihan
AU - Fan, Xiaoyi
AU - Hu, Xiping
AU - Cui, Laizhong
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/4/8
Y1 - 2026/4/8
N2 - In interactive LLM streaming, lengthy chain-of-Thought (CoT) reasoning delays the delivery of user-consumable content, degrading responsiveness, particularly under variable network conditions. Existing methods, however, fail to adapt CoT length to network dynamics. We propose DRLLMS, a network-Adaptive framework that treats CoT length as a control knob to optimize Quality of Experience (QoE). Our analysis, centered on a new metric, Time-To-First-Answer (TTFA), confirms that this pre-Answer reasoning is the primary latency bottleneck. DRLLMS employs a reinforcement learning strategy to enable controllable CoT generation and, at runtime, uses a lightweight optimizer to select the optimal CoT length based on network conditions. Experiments demonstrate that DRLLMS effectively balances responsiveness and answer quality, improving user experience across diverse bandwidth scenarios.
AB - In interactive LLM streaming, lengthy chain-of-Thought (CoT) reasoning delays the delivery of user-consumable content, degrading responsiveness, particularly under variable network conditions. Existing methods, however, fail to adapt CoT length to network dynamics. We propose DRLLMS, a network-Adaptive framework that treats CoT length as a control knob to optimize Quality of Experience (QoE). Our analysis, centered on a new metric, Time-To-First-Answer (TTFA), confirms that this pre-Answer reasoning is the primary latency bottleneck. DRLLMS employs a reinforcement learning strategy to enable controllable CoT generation and, at runtime, uses a lightweight optimizer to select the optimal CoT length based on network conditions. Experiments demonstrate that DRLLMS effectively balances responsiveness and answer quality, improving user experience across diverse bandwidth scenarios.
KW - Chain of Thoughts
KW - Large Language Model
KW - Reinforcement Learning
UR - https://www.scopus.com/pages/publications/105036481150
U2 - 10.1145/3798065.3798072
DO - 10.1145/3798065.3798072
M3 - Conference contribution
AN - SCOPUS:105036481150
T3 - NOSSDAV 2026 - Proceedings of the 36th Workshop on Network and Operating System Support for Digital Audio and Video
SP - 43
EP - 49
BT - NOSSDAV 2026 - Proceedings of the 36th Workshop on Network and Operating System Support for Digital Audio and Video
PB - Association for Computing Machinery, Inc
Y2 - 4 April 2026 through 8 April 2026
ER -