Research map / Large language models / Alignment and preferences
Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning
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.
Introduces FakeVV, a 100K-pair news video misinformation benchmark with entity-level annotations, and Fact-R1, a Qwen2.5-VL-based model trained via long-CoT SFT, DPO, and GRPO with rule-based verifiable rewards.
Problem
Video misinformation detection is limited by small, narrow datasets lacking fine-grained interpretability annotations, and existing MLLM approaches overfit to rigid templates or lack deep reasoning about deceptive content.
Approach
Combine a large news-video misinformation dataset with manipulated-entity labels and a three-stage pipeline (long-CoT tuning, DPO, GRPO) using a rule-based reward that includes an entity-grounding term judged by an LLM.
Key results
- Fact-R1 reaches 81.2 accuracy and 80.3 F1 on FakeVV, above the best listed baseline FakingRec (72.1 / 72.0).
- On FakeSV, Fact-R1 gets 75.6 accuracy and 74.7 F1 versus 69.5 / 70.0 for FakingRec.
- Ablation: removing SFT drops FakeVV accuracy from 81.2 to 66.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Fact-R1 instead of FakingRec | video misinformation detection · FakeSV | accuracy: 69.5 -> 75.6 | Helped |
| use Fact-R1 instead of FakingRec | video misinformation detection · FakeTT | accuracy: 71.0 -> 74.4 | Helped |
| use Fact-R1 instead of FakingRec | video misinformation detection · FakeVV | accuracy: 72.1 -> 81.2 | Helped |
| add misinformation long-CoT instruction tuning (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 71.7 -> 72.7; 66.8 -> 80.3 | Helped |
| add DPO preference alignment (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 72.4 -> 72.7; 79.9 -> 80.3 | Helped |
| add GRPO with the verifiable reward function (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 70.6 -> 72.7; 79.1 -> 80.3 | Helped |
| include audio input (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 72.3 -> 72.7; 77.7 -> 80.3 | Helped |
| include the reasoning keywords reward (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 72.0 -> 72.7; 79.9 -> 80.3 | Helped |
| include the entity grounding reward (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 71.4 -> 72.7; 80.0 -> 80.3 | Helped |
| include the News Image OCR auxiliary task (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 71.8 -> 72.7; 80.2 -> 80.3 | Mixed |
| include the News Video Caption auxiliary task (ablation) | Fact-R1 · FakeTT and FakeVV | F1: 71.6 -> 72.7; 77.7 -> 80.3 | Helped |
Surprising findings
- Text-only DeepSeek-R1 and GPT-o1-mini reach non-trivial accuracy on video misinformation when given captions, and the paper reports strong reasoning from text-only models.
- Removing the DPO stage slightly improves FakeTT accuracy (72.1 vs 74.4 for full model is lower, so DPO helps) but the paper reports w/o DPO at 80.4 on FakeVV, close to full model.
Limitations
- Real-world accuracy remains low, and generalization to open-domain or adversarial settings is untested.
- FakeVV is sourced only from four Western news outlets, which may introduce geographic and cultural bias.
- Entity-identification failures occur when audio and OCR provide no explicit cross-modal clues.
Experimental setup
Datasets: FakeVV FakeSV FakeTT
Benchmarks: FakeVV FakeSV FakeTT
Metrics: Accuracy Precision Recall F1 interpretability accuracy
Baselines: BERT TikTec FANVM SV-FEND FakingRec Gemini2-thinking GPT-4o GPT-o1-mini DeepSeek-R1 Qwen2.5-VL-7B Qwen2.5-VL-72B QVQ-72B-preview
Builds on: DeepSeek-R1 GRPO DPO Qwen2.5-VL
Closest papers by meaning
- Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection (AAAI 2025 · similarity 0.88)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.87)
- Pioneering Explainable Video Fact-Checking with a New Dataset and Multi-role Multimodal Model Approach (AAAI 2025 · similarity 0.87)
- SNIFFER: Multimodal Large Language Model for Explainable Out-of-Context Misinformation Detection (CVPR 2024 · similarity 0.86)
- Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection (NeurIPS 2025 · similarity 0.86)
- PRPO: Paragraph-level Policy Optimization for Vision-Language Deepfake Detection (ICML 2026 · similarity 0.85)
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.