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..
Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning
Authors: Fanrui Zhang, Dian Li, Qiang Zhang, Chenjun, sinbadliu, Junxiong Lin, Jiahong Yan, Jiawei Liu, Zheng-Jun Zha
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the performance of our proposed framework Fact-R1 and several baseline models, we conduct experiments on three real-world short video misinformation datasets: Fake SV, Fake TT, and Fake VV. ... We compare Fact-R1 against 15 competitive baselines... We conduct ablation experiments to verify the contribution of each stage in our proposed three-stage training pipeline, as shown in Table 3. |
| Researcher Affiliation | Collaboration | Fanrui Zhang1,2 , Dian Li3 , Qiang Zhang1 , Jun Chen3, Gang Liu3, Junxiong Lin4, Jiahong Yan3, Jiawei Liu1 , Zheng-Jun Zha1 1Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC 2Shanghai Innovation Institute 3Tencent QQ 4Fudan University |
| Pseudocode | Yes | Algorithm 1: GRPO with Task-Specific Reward Functions |
| Open Source Code | Yes | Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The paper is described in README.md in terms of dataset preparation, the exact commands to be run to reproduce the results and the environment. |
| Open Datasets | Yes | We construct Fake VV, the largest and most comprehensively annotated news-domain video misinformation dataset. ... the Fake VV dataset is provided strictly for non-commercial research purposes and is accessible only to verified academic researchers under a research-specific license agreement. |
| Dataset Splits | Yes | Following the original setup in [24], we adopt a chronological split with a ratio of 70% for training, 15% for validation, and 15% for testing, simulating realistic scenarios where only past data is available for detecting future misinformation. |
| Hardware Specification | Yes | All experiments are conducted using Py Torch on 8 NVIDIA A100 GPUs, with Qwen2.5-VL as the base MLLM and the default 8-frame sampling strategy. |
| Software Dependencies | No | All experiments are conducted using Py Torch on 8 NVIDIA A100 GPUs, with Qwen2.5-VL as the base MLLM and the default 8-frame sampling strategy. ... using Adam W optimizer. |
| Experiment Setup | Yes | Stage 1 (Long-Co T Tuning). We freeze the visual encoder and apply Lo RA (r = 128, α = 256), using learning rates of 2 10 5 (Lo RA) and 2 10 6 (multimodal projector). Training is performed for 2 epochs with batch size 2, using Adam W optimizer. Stage 2 (DPO). We train for 1 epoch on 5k human preference pairs with batch size 4 and a learning rate of 2 10 5. The reference model is frozen, and gradient clipping is applied with a max norm of 1.0. Stage 3 (GRPO). We train for 172 steps using the verifiable reward function, sampling 5 candidate responses per input and applying a learning rate of 1 10 6. |