Streaming Video-LMM. Start with Six Development Paths from 2023 to 2026 for a map of query visibility, memory budgets, retrieval, interaction, and streaming evaluation.
A resource-aware comparison of growing KV archives, bounded online state, and the evidence behavior that future-query memory should preserve.
Six development paths from long-video compression to bounded memory, proactive interaction, and system-level streaming evaluation.
Similarity-gated visual memory and persistent-query retrieval for proactive ultra-long streaming video.
Hierarchical external event memory and agentic on-demand retrieval for streaming video reasoning.
Causal ViT activation reuse and pre-LLM visual token pruning for faster streaming Video-LLMs.
Temporal-order instruction tuning and uncertainty-gated historical KV retrieval for streaming Video-LLMs.
Training-inference aligned real-time understanding for effectively infinite video streams.
A recent-frame baseline for streaming video understanding.
Multi-grained KV-cache compression for long streaming VideoQA.
A model note on Qwen3-VL: long-context vision-language modeling, video understanding, and multimodal agent ability.