Abstract

Real-time applications—including interactive video streaming, cloud gaming, and multimodal
large language model (MLLM) assistants—require low-latency, reliable delivery over highly variable
wireless networks. Yet current transport and application mechanisms are often optimized in
isolation: HTTP-based streaming introduces segment-level delays, generic WebRTC and QUIC
configurations lack application awareness, and MLLM pipelines frequently wait for full video upload
before inference begins. This thesis examines whether adaptive, application-informed transport
can reduce end-to-end latency and improve reliability across these workloads.
The first part of the thesis presents RabbitLLM, an end-to-end system for real-time MLLM
inference. RabbitLLM combines WebRTC-based frame-wise video streaming with chunked KVcache
prefilling and progressive decoding, enabling LLM processing to begin before full video
transmission completes. Compared to conventional TCP pipelines, it achieves lower transmission
delay, reduced end-to-end latency, and more efficient bandwidth usage while maintaining semantic
output quality.
The second part introduces TAROT, an optimization-driven adaptive forward error correction
(FEC) framework for QUIC-based video streaming. TAROT selects redundancy parameters dynamically
based on network conditions and timing constraints, yielding higher perceptual video
quality and more stable playback than static RaptorQ or Reed–Solomon strategies across diverse
cellular traces.
Together, these results show that cross-layer adaptive transport—integrating WebRTC, QUIC,
and application-level execution dynamics can significantly improve latency, efficiency, and perceptual
quality for next-generation real-time multimodal and video streaming systems.