AI Paper Atlas

Research map / Video / Video-language understanding

AdaAlloc: Adaptive Visual Token Allocation for Long-Video Question Answering

Joungbin An, Kristen Grauman

NeurIPS 2026 · Video › Video-language understanding · approach: Video question answering · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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 AdaAlloc, an adaptive visual-token allocation method for long-video question answering under fixed token budgets; the abstract is truncated, so details are limited.

Problem

Long-video question answering with multimodal LLMs under strict visual-token budgets, where uniform sampling loses fine-grained local cues and caption- or similarity-based retrieval relies on lossy proxies that miss motion and state changes.

Approach

Adaptively allocates the visual-token budget across a video to balance temporal coverage and spatial detail, rather than using uniform sampling or proxy-based retrieval; exact mechanism not stated in the available text.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
allocate tokens between global overview and local evidence with AdaAlloc instead of uniform samplingQwen3-VL-8B-Instruct · LVBench (32K)accuracy: 46.1 -> 56.6Improved
allocate tokens with AdaAlloc instead of uniform samplingQwen3-VL-8B-Instruct · VideoMME w/o sub (16K)accuracy: 67.1 -> 68.4Improved
make the global/local split question-conditioned instead of fixed Balanced (ablation)AdaAlloc · 4 benchmarks (16K)accuracy: LVBench 49.6 -> 51.7Improved
add a fixed global stream to a local-only variant (ablation)AdaAlloc · 4 benchmarks (16K)accuracy: LVBench 48.9 -> 49.6No clear effect
add temporal-memory refinement of retrieved intervals (ablation)AdaAlloc · MLVU-Test (32K)accuracy: 66.5 -> 67.1Improved
add overview-conditioned planning and refinement on top of question-only retrieval (ablation)AdaAlloc · 4 benchmarks (16K)accuracy: LVBench +2.4, MLVU-Test -0.8Mixed
use retrieval-only local clips instead of uniform sampling (ablation)Qwen3-VL-8B-Instruct · MLVU-Test (32K)accuracy: 61.4 -> 59.4Hurt
use five discrete policies instead of directly predicting the ratio (ablation)AdaAlloc · 4 benchmarks (32K)average accuracy: 64.94 -> 65.51Improved
double the planner overview density (ablation)AdaAlloc · 4 benchmarks (16K)accuracy: LVBench 51.7 -> 51.6No clear effect
use UniversalVTG instead of query-frame similarity retrieval (ablation)AdaAlloc · 4 benchmarks (32K)average accuracy: 63.85 -> 65.51Improved
apply AdaAlloc to VideoLLaMA3-7BVideoLLaMA3-7B · 4 benchmarks (32K)average accuracy: 56.35 -> 58.60Improved
use AdaAlloc instead of DIG/AKS frame selection with fewer tokensQwen3-VL-8B-Instruct · VideoMME medium (w/o sub)accuracy: 58.4 -> 68.7Improved

Limitations

Experimental setup

Baselines: uniform sampling caption-based retrieval frame-similarity retrieval

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