AI Paper Atlas

Research map / Video / Video-language understanding

Small Vision-Language Models are Smart Compressors for Long Video Understanding

Junjie Fei, Jun Chen, Zechun Liu, Yunyang Xiong, Chong Zhou, Wei Wen, Junlin Han, Mingchen Zhuge, Saksham Suri, Qi Qian, Shuming Liu, Lemeng Wu and 4 more

ECCV 2026 · Video › Video-language understanding · approach: Long video understanding · 8 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Adaptive Token Allocation (8K) vs none (16K) (ablation)Tempo · LVBenchaccuracy: 51.1 -> 52.3Improved
add Stage 3 long-context SFT (ablation)Tempo · LVBenchaccuracy: 47.3 -> 51.1Improved
use ATA instead of uniform subsampling (ablation)Tempo · LongVideoBenchaccuracy: 61.9 -> 65.1Improved
route to lowest-scoring segments (adversarial) (ablation)Tempo · LongVideoBenchaccuracy: 65.1 -> 50.7Hurt
use ATA instead of hard top-K routing (ablation)Tempo · LVBenchaccuracy: 52.7 -> 52.3No clear effect
keep head tokens instead of tail tokens (dynamic truncation) (ablation)Tempo · LongVideoBenchaccuracy: 61.9 -> 65.1Improved
use head truncation instead of token merging (ablation)Tempo · LVBenchaccuracy: 53.0 -> 52.3 on LVBench; 66.3 -> 67.7 on Video-MMEMixed
score relevance with Tempo SVLM and explicit prompt instead of external reranker (ablation)Tempo · LVBenchaccuracy: 51.8 -> 52.3No clear effect
use Tempo SVLM explicit prompt vs standard prompt for scoring (ablation)Tempo · LVBenchaccuracy: 53.4 -> 52.3 on LVBench; 64.1 -> 65.1 on LongVideoBenchMixed
enforce minimal temporal anchors k_min=4 instead of hard pruning (ablation)Tempo · LongVideoBenchaccuracy: 63.9 -> 65.1Improved
raise budget from 4K to 8KTempo · LVBenchaccuracy: 52.7 -> 52.3No clear effect
scale to 2048 frames with 12K budgetTempo · LVBenchaccuracy: 52.3 -> 53.7Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy