Research map / Video / Video-language understanding
ST-LLM: Large Language Models Are Effective Temporal Learners
Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, Ge Li
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 |
|---|---|---|---|
| feed all spatial-temporal tokens into the LLM instead of mean pooling over time (ablation) | ST-LLM with InstructBLIP, full fine-tuning · MVBench | average accuracy: 47.8 -> 54.9 | Improved |
| feed all spatial-temporal tokens into the LLM instead of mean pooling over time (LoRA) (ablation) | ST-LLM with InstructBLIP, LoRA · MVBench | average accuracy: 45.2 -> 51.9 | Improved |
| use ST tokens with MiniGPT-4 base instead of mean pooling (ablation) | ST-LLM with MiniGPT-4 · MVBench | average accuracy: 46.4 -> 50.5 | Improved |
| add dynamic masking (vs fixed frames or static 50% mask) during training (ablation) | ST-LLM · MVBench, varying test frame counts | accuracy robustness | Improved |
| add static 50% masking instead of dynamic masking (ablation) | ST-LLM · MVBench, varying test frame counts | accuracy robustness | No clear effect |
| add MVM loss on top of dynamic masking (ablation) | ST-LLM · MVBench | accuracy | Improved |
| add global branch via zero-initialized adapter to local input (ablation) | ST-LLM · VideoChatGPT-Bench Temporal | score: 2.75 -> 2.93 | Improved |
| add global branch via zero-initialized adapter to local input (ablation) | ST-LLM · MVBench | average accuracy: 54.9 -> 54.7 | No clear effect |
| add global tokens by simple addition instead of adapter (ablation) | ST-LLM · MVBench | average accuracy: 54.9 -> 50.1 | Hurt |
| add separator tokens between frames or modalities (ablation) | ST-LLM · MVBench | accuracy | No clear effect |
| add absolute positional encodings to visual tokens (ablation) | ST-LLM · MVBench | accuracy | Hurt |
| add image data to video instruction tuning (InstructBLIP base) (ablation) | ST-LLM with InstructBLIP · MVBench | average accuracy: 54.9 -> 53.1 | Hurt |
Later papers that compared with this method
7 later accepted papers report ST-LLM as a baseline in their results tables; 6 of them report a better number on most of the results they share with it.
- IPFormer: Instance Prompt-guided Transformer for Multi-modal Multi-shot Video Understanding (AAAI 2026 · better on 0 of 3 shared results)
- DynFocus: Dynamic Cooperative Network Empowers LLMs with Video Understanding (CVPR 2025 · better on 5 of 7 shared results)
- MASH-VLM: Mitigating Action-Scene Hallucination in Video-LLMs through Disentangled Spatial-Temporal Representations (CVPR 2025 · better on 4 of 7 shared results)
- DynImg: Key Frames with Visual Prompts are Good Representation for Multi-Modal Video Understanding (ICCV 2025 · better on 5 of 5 shared results)
- DisCo: Towards Distinct and Coherent Visual Encapsulation in Video MLLMs (ICCV 2025 · better on 3 of 3 shared results)
- STD-PLM: Understanding Both Spatial and Temporal Properties of Spatial-Temporal Data with PLM (AAAI 2025 · better on 1 of 1 shared result)
- TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations (AAAI 2025 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LLaMA-VID-13B — LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models (ECCV 2024) · better on 2 of 2 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
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.90)
- On the Consistency of Video Large Language Models in Temporal Comprehension (CVPR 2025 · similarity 0.90)
- VideoLLM-online: Online Video Large Language Model for Streaming Video (CVPR 2024 · similarity 0.90)
- Momentor: Advancing Video Large Language Model with Fine-Grained Temporal Reasoning (ICML 2024 · similarity 0.90)
- VTimeLLM: Empower LLM to Grasp Video Moments (CVPR 2024 · similarity 0.90)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (CVPR 2025 · similarity 0.90)
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.
