Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models

Dimensions overall score 9.0

GitHub Code Pulse

No public code linked for this paper yet.

• The 'Think While Watching' framework enables MLLMs to perform continuous video reasoning during live video streaming by creating persistent segment-level memory.method 90%
• The method improves single-round accuracy by 2.6% on StreamingBench when built on Qwen3-VL.result 90%
• The method improves single-round accuracy by 3.79% on OVO-Bench when built on Qwen3-VL.result 90%
• In multi-round settings, the method reduces output tokens by 56% while maintaining performance.result 90%
• The framework uses segment-level streaming causal masks and positional encoding to maintain context and improve inference.technical 85%
• The system's memory management scalability might be limited as input streams grow extensively.limitation 70%
• The system might be limited by the underlying MLLM's ability to generalize across varied video content.limitation 70%
• The 'Think While Watching' framework enables MLLMs to perform continuous video reasoning during live streaming by preserving segment-level memory.method 95%
• The proposed method improves single-round accuracy by 2.6% on StreamingBench.result 90%
• The proposed method improves single-round accuracy by 3.79% on OVO-Bench.result 90%
• In multi-round streaming settings, the method maintains performance while reducing output tokens by 56%.result 90%
• The framework enforces strict causality through a segment-level streaming causal mask and streaming positional encoding.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card