Research map / Video / Video-language understanding
NarrativeTrack: Evaluating Entity-Centric Reasoning for Narrative Understanding
Hyeonjeong Ha, Jinjin Ge, Bo Feng, Kaixin Ma, Gargi Chakraborty
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 changed | Where | Measured | Effect |
|---|---|---|---|
| ensemble Detectron2 and OWLv2 detections via spatial consensus (ablation) | entity detection pipeline · AVA recall | recall: 0.780 / 0.801 -> 0.848 | Improved |
| scale open-source model size within a family | InternVL3 8B to 38B · NarrativeTrack avg | accuracy (%): 48.81 -> 55.47 | Improved |
| scale Video-LLaMA2 from 7B to 72B | Video-LLaMA2 · NarrativeTrack avg | accuracy (%): 43.04 -> 49.70 | Improved |
| scale LLaVA-NeXT-Video from 7B to 34B | LLaVA-NeXT-Video · NarrativeTrack avg | accuracy (%): 42.94 -> 39.36 | Hurt |
| increase GPT-4o frames from 20 to 128 (ablation) | GPT-4o · NarrativeTrack avg | accuracy (%): 70.97 -> 72.27 | Improved |
| increase frame count beyond 20 for open-source models (ablation) | open-source MLLMs · NarrativeTrack avg | accuracy (%) | Hurt |
| remove visual input (ablation) | GPT-4o · NarrativeTrack avg | accuracy (%): -30.52% | Hurt |
| reverse video frames (ablation) | GPT-4o · entity change ordering | accuracy (%): 51.2 -> 6.1 | Hurt |
| backward instead of forward temporal reasoning | MLLMs · NarrativeTrack | accuracy (%): gap of 20.65% (general), 9.96% (video), 17.55% (proprietary) | Hurt |
| synthetic instead of real distractors | video-specialized MLLMs · NarrativeTrack | accuracy (%): up to 9.69% drop | Hurt |
| synthetic instead of real distractors | general-purpose MLLMs · NarrativeTrack | accuracy (%): +0.44% on synthetic | No clear effect |
Closest papers by meaning
- HENASY: Learning to Assemble Scene-Entities for Interpretable Egocentric Video-Language Model (NeurIPS 2024 · similarity 0.89)
- SeriesBench: A Benchmark for Narrative-Driven Drama Series Understanding (CVPR 2025 · similarity 0.88)
- VELOCITI: Benchmarking Video-Language Compositional Reasoning with Strict Entailment (CVPR 2025 · similarity 0.87)
- JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.87)
- OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs (ICLR 2026 · similarity 0.87)
- Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval (NeurIPS 2025 · similarity 0.87)
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.
