Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO
Authors: Jinyoung Park, Jeehye Na, Jinyoung Kim, Hyunwoo J. Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that our approach significantly improves video reasoning performance across multiple benchmarks. |
| Researcher Affiliation | Academia | 1Korea Advanced Institute of Science and Technology, 2Korea University EMAIL EMAIL |
| Pseudocode | No | The paper includes mathematical equations and derivations for Reg-GRPO loss, but it does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | https://github.com/mlvlab/Deep Video R1 |
| Open Datasets | Yes | To validate the effectiveness of the proposed method, we conduct evaluations on various video benchmarks, including both general video understanding tasks (e.g., SEED-Bench-R1 [56], VSIBench, Video-MMMU, MMVU (mc), MVBench, Temp Compass, Video-MME (wo sub)), long video understanding tasks (e.g., Long Video Bench [25]), and fine-grained spatial-temporal video reasoning tasks (NEx TGQA [37]). More details about datasets are in Appendix E. |
| Dataset Splits | Yes | Table 1 summarizes the performance of various baselines, supervised fine-tuning (SFT), GRPO, and our proposed Deep Video-R1 on the validation splits of the SEED-Bench-R1 (SBR) dataset. For the NEx T-GQA benchmark, ... temporal segment annotations are available only in the validation and test splits, we use the validation and test splits for the model training and model evaluation, respectively. |
| Hardware Specification | Yes | We use NVIDIA A100 GPUs for 3B models and NVIDIA H200 GPUs for 7B models. |
| Software Dependencies | No | We implement our code using the Py Torch library [79]. We also adopt the Hugging Face Transformers library [80] and the TRL library [81] to post-train Video Large Language Models (Video LLMs). For inference and rollout, we use v LLM [82]. The paper mentions these libraries but does not provide specific version numbers for them. |
| Experiment Setup | Yes | For the SEED-Bench-R1 dataset, we apply a KL-divergence regularizer between the model ΟΞΈ and the reference model Οref with coefficient 0.1, following prior GRPO works [12, 56]. We use Qwen2.5-VL as the default base Video LLM, and use Qwen2.5-VL-3B for all analyses. We set the number of generations in the group as 8 for all the settings. For Deep Video-R1, we maintain a reward history using the most recent W = 100 samples, and GRPO does not adopt safeguards based on our empirical study. To train the model on the SEED-Bench-R1 dataset, we limit the maximum number of sampled frames per input video to 16 with a frame resolution of 252 252, and then append the frame indicating the current observation as an additional image input, following SEED-Bench-R1 [56]. |