Research map / Video / Video-language understanding

EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning

NeurIPS 2026 · Video › Video-language understanding · approach: Event and anomaly reasoning in video · 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 ECP, a training-free event-camera-guided visual token pruning framework for Video-LLMs on first-person dynamic spatial reasoning, and introduces ESR-Real, a real RGB-event benchmark.

Problem

Visual token pruning for Video-LLMs in first-person dynamic spatial reasoning fails to preserve motion continuity and geometric structure, since RGB-only temporal and attention-based spatial pruning signals are fragile under high compression.

Approach

Uses event-camera activity as a training-free, score-level motion and structure prior in three cascaded stages: event-triggered keyframe sampling, event-guided token saliency filtering, and rank-based fusion of event saliency with attention scores inside the LLM at pruning layers.

Key results

• At 20% visual token retention on the 13-task public benchmark average, ECP reaches 37.62% vs. 36.31% for the full-token baseline, with 1.89x speedup and 52% GFLOPs reduction (141.9 GFLOPs).
• On ESR-Real at 50% retention, ECP achieves 54.58% vs. 51.90% for the full-token baseline (+2.68 points).
• At 20% retention, the strongest baseline DEP (event-only density pruning) reaches 29.95%, and attention-based FastV reaches 32.64%.

Surprising findings

• Pruning to 20% retention with ECP surpasses the unpruned full-token model on the public benchmark average.
• Event-only density pruning (DEP) performs worse than RGB-only baselines at high compression, suggesting that the gain comes from the cascaded design rather than event access alone.

Limitations

• Requires synchronized RGB-event inputs, so it is not an RGB-only acceleration method.
• Public benchmarks rely on simulated events from vid2e rather than native event signals.
• ESR-Real is small (over 700 QA pairs) and built by an internal team with Gemini-3Pro-generated questions and manual refinement.

Experimental setup

Datasets: VSI-Bench UrbanVideo-Bench ESR-Real

Benchmarks: VSI-Bench UrbanVideo-Bench ESR-Real

Metrics: accuracy latency GFLOPs peak memory

Baselines: DTD FastV PyramidDrop VTW DivPrune BTP Direct Event Pruning (DEP)

Builds on: FastV PyramidDrop DivPrune DTD

Closest papers by meaning

• Talk2Event: Grounded Understanding of Dynamic Scenes from Event Cameras (NeurIPS 2025 · similarity 0.87)
• IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning (ICLR 2026 · similarity 0.86)
• Motion-Grounded Video Reasoning: Understanding and Perceiving Motion at Pixel Level (CVPR 2025 · similarity 0.85)
• Learning to Evolve Scenes: Reasoning about Human Activities with Scene Graphs (NeurIPS 2026 · similarity 0.85)
• ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.85)
• PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.85)