Research map / Vision-language and multimodal / Visual question answering
Divide and Conquer: Exploring Language-centric Tree Reasoning for Video Question-Answering
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 a training-free, model-agnostic Language-centric Tree Reasoning (LTR) framework that recursively decomposes VideoQA questions into a logical tree with RAG-based few-shot division, then performs video-aided bottom-up reasoning with answer verification.
Problem
VideoQA models, especially MLLMs, lack controllable, transparent, and traceable reasoning for complex cognitive questions requiring multi-step logic over video content.
Approach
Recursively divide a question into a language-centric logical tree of perceptual leaf questions (using RAG over AGQA-Decomp exemplars and video-grounded checking), answer leaves with the MLLM, then reason bottom-up along the tree with video assistance and in-process verification, yielding a traceable reasoning path.
Key results
- On AGQA-Decomp, LTR raises VideoChat2 compositional cF1 from 67.0 to 77.4 and VideoLLaMA3 from 69.4 to 81.2.
- On NExT-QA, Qwen2-VL improves from 80.4 to 82.7 Acc@All with LTR.
- On Causal-VidQA, Qwen2-VL improves from 68.0 to 70.2 Acc@All with LTR; ablation shows removing answer verification drops it to 69.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply LTR | Qwen2-VL · Causal-VidQA | Acc@All: 68.0 -> 70.2 | Helped |
| Apply LTR | Qwen2-VL · NExT-QA | Acc@All: 80.4 -> 82.7 | Helped |
| Apply LTR | Qwen2-VL · MVBench | Avg: 65.7 -> 67.8 | Helped |
| Apply LTR | Qwen2-VL · AGQA-Decomp | main accuracy: 65.4 -> 69.1 | Helped |
| Add video to logical tree generation (ablation) | LTR with Qwen2-VL · Causal-VidQA | Acc@All: 68.9 -> 70.2 | Helped |
| Add video to logical tree generation (ablation) | LTR with Qwen2-VL · NExT-QA | Acc@All: 81.5 -> 82.7 | Helped |
| Add RAG to logical tree generation (ablation) | LTR with Qwen2-VL · Causal-VidQA | Acc@All: 69.3 -> 70.2 | Helped |
| Add RAG to logical tree generation (ablation) | LTR with Qwen2-VL · NExT-QA | Acc@All: 82.2 -> 82.7 | Helped |
| Add in-process answer verification (ablation) | LTR with Qwen2-VL · Causal-VidQA | Acc@All: 69.7 -> 70.2 | Helped |
| Add in-process answer verification (ablation) | LTR with Qwen2-VL · NExT-QA | Acc@All: 82.0 -> 82.7 | Helped |
| Add perceptual leaf question answering (ablation) | LTR with Qwen2-VL · Causal-VidQA | Acc@All: 66.1 -> 70.2 | Helped |
| Add perceptual leaf question answering (ablation) | LTR with Qwen2-VL · NExT-QA | Acc@All: 79.1 -> 82.7 | Helped |
Surprising findings
- Removing video from the logical reasoning stage drops Causal-VidQA accuracy to 52.6, far below the baseline, indicating leaf answers alone are insufficient.
Limitations
- Errors in leaf perception propagate and can cause unrecoverable wrong answers, as shown in the counting failure case.
- Reliance on GPT-3.5 as judge for open-ended evaluation and on multiple MLLM calls per question for tree construction and verification.
Experimental setup
Datasets: AGQA-Decomp
Benchmarks: MSVD-QA MSRVTT-QA TGIF-QA ActivityNet-QA AGQA-Decomp NExT-QA Causal-VidQA STAR EgoSchema Video-MME MVBench
Metrics: accuracy score cP cR cF1
Baselines: VideoChat2 VideoLLaMA3 LLaVA-OneVision Qwen2-VL VoT Video-LLaVA LLaMA-VID Chat-UniVi
Builds on: AGQA-Decomp Video-of-Thought (VoT) Neural Modular Networks
Closest papers by meaning
- Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning (CVPR 2025 · similarity 0.92)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 2024 · similarity 0.91)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.90)
- Modularized Self-Reflected Video Reasoner for Multimodal LLM with Application to Video Question Answering (ICML 2025 · similarity 0.90)
- Tool-Augmented Spatiotemporal Reasoning for Streamlining Video Question Answering Task (NeurIPS 2025 · similarity 0.89)
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.