Research map / Video / Video-language understanding

EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs

NeurIPS 2026 · Video › Video-language understanding · approach: Long video understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EchoPrune, a training-free, query-aware visual token pruning method for VideoLLMs that scores tokens by crossmodal relevance minus temporal reconstruction error, letting models see up to 20x more frames under a fixed LLM token budget.

Problem

Long-video VideoLLMs face a trade-off: dense frame sampling creates massive visual tokens, while sparse sampling misses critical temporal evidence and induces hallucination. Existing training-free token reduction methods either treat video like images or rely on segment-level merging heuristics that

Approach

Interprets redundant video tokens as temporal echoes: tokens well reconstructed from the previous frame (via correspondence and softmax-weighted echo matching over a neighborhood) are pruned, while query-relevant (max-aggregated crossmodal similarity) and poorly reconstructed tokens are kept, combined in an MMR-style score with a single global Top-K before the LLM.

Key results

• On Qwen2.5-VL-7B with F320 (20x frames, 5% tokens), EchoPrune reaches 55.6 average vs 52.4 for the random baseline and up to 8.6% gain over the F16 baseline setting reported.
• On LLaVA-OV-7B at F320 with 10% tokens, EchoPrune scores 54.5 average (+4.6%), outperforming FastVID (-7.5%) and HoliTom (+2.9%), while UniComp reaches 54.0.
• On LLaVA-OV-7B under F160/20%, EchoPrune achieves 5.6x prefilling speedup and 2.1x TTFT speedup relative to F160 baseline, with score 60.4 vs 58.6.

Surprising findings

• Query-only relevance pruning performs worse than random pruning on short videos for Qwen2.5-VL.
• Increasing frames to F160 without pruning does not improve LLaVA-OV-7B (58.8 to 58.6), while pruning to the F32 budget improves it to 60.4.

Limitations

• Pruning is done once in the projection layer without ViT acceleration, so ViT cost is not reduced.
• Lacks RoPE-aware design for positional handling of retained tokens, per the authors.
• Only vision modality is considered; audio-guided pruning left for future work.

Experimental setup

Datasets: VideoMME EgoSchema LongVideoBench MLVU VideoMMMU LVBench GQA ScienceQA-Img TextVQA

Benchmarks: VideoMME EgoSchema LongVideoBench MLVU VideoMMMU LVBench GQA ScienceQA-Img TextVQA

Metrics: accuracy average score prefilling time time-to-first-token compression time

Baselines: MI-Pruner VisionZip FastVID HoliTom UniComp FlashVID ApET Random sampling Query-only pruning

Builds on: Maximum Marginal Relevance VisionZip FastVID HoliTom FlashVID KiToke

Closest papers by meaning

• FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
• HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
• V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.89)
• One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
• LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.88)
• ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.88)