Challenge: Video large language models (Vid-LLMs) rely on dense video token representations and require substantial memory and computational overhead in both prefilling and decoding.
Approach: They propose a training-free speculative decoding framework that prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy.
Outcome: The proposed framework achieves 2.68 speedup on LLaVA-OneVision-72B and 2.11 speed up on Qwen2.5-VL-32B.

Similar Papers

PruneVid: Visual Token Pruning for Efficient Video Large Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Existing approaches to video token pruning face significant computational challenges due to the redundancy inherent in video data.
Approach: They propose a training-free visual token pruning method that reduces the redundancy inherent in video data and leverages LLMs’ inherent ability to selectively prune visual tokens irrelevant to specific queries.
Outcome: The proposed method can prune over 80% of tokens while maintaining competitive performance when combined with different video LLMs.
See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for video understanding suffer from autoregressive generation of tokens.
Approach: They propose a training-free loosely SD framework for Video-LLMs that uses visual-relevant tokens to accurately pinpoint the latter.
Outcome: The proposed framework boosts the accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.
Vista-LLM: Decoupled Query-Guided Visual Token Pruning for Efficient Long-Video Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Long-video understanding is bottlenecked by the high cost of processing massive visual tokens.
Approach: They propose a decoupled framework for query-guided visual token pruning . their method reduces visual tokens by 90% and accelerates inference by 98% .
Outcome: The proposed framework reduces visual tokens by 90% and accelerates inference while retaining over 98% of baseline performance on average.
MoPrune: Scene-Guided Motion-Aware Token Pruning for Efficient Video Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Prior approaches to token pruning ignore video dynamics and the fact that different scenes exhibit different redundancy patterns.
Approach: They propose a token pruning framework that is train-free and scene-guided to accelerate VideoLLMs by removing redundant visual information from video frames.
Outcome: MoPrune is a training-free, scene-guided and motion-centric token pruning framework for accelerating VideoLLMs.
Sparse-to-Dense: A Free Lunch for Lossless Acceleration of Video Understanding in LLMs (2025.acl-short)

Copied to clipboard

Challenge: Recent advances in Video Large Language Models (Video-LLMs) have achieved exceptional performance on tasks like video question answering and captioning.
Approach: They propose a decoding strategy that leverages sparse top-K attention and dense full attention to accelerate Video-LLMs without loss.
Outcome: The proposed approach achieves a 1.94 walltime speedup in video processing.
LVPruning: An Effective yet Simple Language-Guided Vision Token Pruning Approach for Multi-modal Large Language Models (2025.findings-naacl)

Copied to clipboard

Challenge: Multi-modal Large Language Models (MLLMs) incur significant computational overhead due to the large number of vision tokens processed, limiting their practicality in resource-constrained environments.
Approach: They propose a language-guided vision token pruning method that can be integrated into existing MLLMs with minimal architectural changes.
Outcome: The proposed method reduces vision tokens by 90% and preserves model performance.
Reducing Token Redundancy in LVLMs: A Systematic Review of Token Pruning Methods (2026.acl-long)

Copied to clipboard

Challenge: Large Vision-Language Models (LVLMs) excel at visual understanding but face severe computational bottlenecks when processing high-resolution images and long videos due to massive visual token counts.
Approach: They propose a taxonomy categorizing methods into vision-side, LLM-side and hybrid paradigms and analyze token selection mechanisms and pruning strategy.
Outcome: The proposed method selectively removes less informative tokens while maintaining performance.
POP: Prefill-Only Pruning for Efficient Large Model Inference (2026.findings-acl)

Copied to clipboard

Challenge: Existing structured pruning methods suffer from significant accuracy degradation . Existing pruning methods are expensive and require specialized hardware and kernels to perform .
Approach: They propose a stage-agnostic pruning approach that overlooks asymmetric roles between prefill and decode stages.
Outcome: The proposed pruning approach achieves 1.37 speedup in prefill latency with minimal performance loss.
Video Compression Commander: Plug-and-Play Inference Acceleration for Video Large Language Models (2025.emnlp-main)

Copied to clipboard

Challenge: Recent studies have shown that Video Large Language Models (Vide-oLLMs) are efficient at video understanding but lack the quadratic complexity of visual tokens.
Approach: They propose a plug-and-play inference acceleration framework for VideoLLM token compression that quantifies each frame’s uniqueness and adaptively adjusts compression intensity across frames.
Outcome: Extensive experiments on video large language models and benchmarks show that the proposed framework can preserve essential information while reducing redundancy in video sequences.
Language-Guided Temporal Token Pruning for Efficient VideoLLM Processing (2025.emnlp-main)

Copied to clipboard

Challenge: Current models struggle with long-form videos due to the quadratic complexity of attention mechanisms.
Approach: They propose a model-agnostic framework that leverages temporal cues from queries to prune video tokens.
Outcome: The proposed framework reduces computation by 65% while preserving 97-99% of original performance.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations