Source-linked AI summary
Self-Reflective Multi-modal Reasoning for Short-Video Fake News Detection
Pinjie Xu, Yuzhou Yang, Zhikai Tan, Qichao Ying, Zaiyang Yu, Ce Li, Zhenxing Qian
TL;DR
Short-video fake-news detectors often lack explicit, self-correcting reasoning and cross-sample verification. SRM-FND addresses this with self-reflective multimodal reasoning, topic-adaptive fine-tuning, and confidence-driven review, outperforming strong baselines in in-domain and cross-dataset evaluations.
Problem
Existing short-video fake-news detectors often lack explicit self-correcting reasoning, leaving them vulnerable to spurious patterns and limiting interpretability.
Method
SRM-FND iteratively constructs reflective rationales through contrastive deliberation and prompt refinement, transfers them via topic-adaptive VLM fine-tuning, and reviews uncertain cases across samples.
Results
76.20 and 72.24 overall accuracy under cross-dataset transfer exceeded the strongest prior baselines by 13.70 and 9.24 points, respectively.
Takeaways & Limitations
SRM-FND provides more reliable and interpretable predictions while improving performance in both in-domain and cross-dataset short-video fake-news detection.
Takeaways & Limitations
Topic-classification accuracy cannot be measured because ground-truth topic labels are unavailable.
Abstract
from arXiv · showhide
Recent fake news detection pipelines increasingly leverage large language models and vision-language models for reasoning-based analysis. However, several challenges remain open: improving reasoning quality through self-reflection without ground-truth chain-of-thought supervision, using improved reasoning to benefit downstream model fine-tuning, and connecting single-sample fraudulent-pattern discovery with cross-sample verification. We propose SRM-FND, a self-reflective multimodal reasoning framework for short-video fake news detection. SRM-FND develops higher-quality reasoning through contrastive deliberation, iterative root-cause diagnosis, and corrective prompt refinement. A Blind Analyst, Counter-Conclusion Reasoner, and Self-Consistency Arbiter collaboratively identify and retain discriminative rationales. The framework also incorporates dual-phase, topic-adaptive vision-language model fine-tuning to improve multimodal grounding and enable lightweight topic specialization. For uncertain cases, it performs confidence-driven cross-sample review by retrieving credible and suspicious co-event examples. Experiments on FakeSV and FakeTT show that SRM-FND outperforms strong baselines, produces more reliable and interpretable predictions, and delivers noticeable improvements in cross-dataset performance.
1 Introduction
SRM-FND addresses limitations of surface-level short-video fake-news detection by combining self-reflective multimodal reasoning, topic-adaptive fine-tuning, and confidence-driven cross-sample review. On FakeSV and FakeTT, it improves overall and cross-dataset accuracy over prior baselines.
- Motivation: Existing supervised multimodal fusion methods capture surface-level correlations but often lack explicit reasoning capability for short-video fake-news detection.These systems typically combine visual, textual, and metadata signals to predict veracity.
- Framework: SRM-FND constructs higher-quality reasoning through contrastive deliberation, iterative root-cause diagnosis, and corrective prompt refinement without requiring ground-truth CoT labels.A Blind Analyst, Counter-Conclusion Reasoner, and Self-Consistency Arbiter collaboratively support the contrastive reasoning process.
- Empirical results: SRM-FND achieves 91.33 and 92.31 overall accuracy on FakeSV and FakeTT, improving over the strongest prior baseline by 1.11 and 3.01 points.These results establish the framework’s in-domain advantage on both datasets.
- Empirical results: Under cross-dataset transfer, SRM-FND achieves 76.20 and 72.24 overall accuracy, outperforming the strongest prior baselines by 13.70 and 9.24 points.The reported transfer results demonstrate substantial gains beyond the training datasets.
- Framework: The framework uses dual-phase topic-adaptive VLM fine-tuning and confidence-driven cross-sample review to connect single-sample analysis with event-level verification for uncertain cases.These components transfer improved reasoning into the detector and provide lightweight topic specialization.
2 Related Works
Prior multimodal fake-news detection methods are organized around content-oriented and evidence-based paradigms, while newer reasoning-augmented approaches use large language models to generate intermediate rationales for verification. However, existing reasoning methods lack self-correction and experience accumulation, motivating self-reflective reasoning for fake-news detection.
- Multimodal Fake-News Detection: Multimodal fake-news detection schemes comprise content-oriented methods that learn veracity predictors from multimodal representations and evidence-based methods.Content-oriented approaches transform news into multimodal latent representations without explicit evidence retrieval.
- Reasoning-Augmented Verification: Reasoning-augmented methods use large language models to generate intermediate rationales and structured reasoning procedures instead of directly mapping inputs to labels.These approaches aim to improve interpretability and decision quality through explicit reasoning processes.
- Research Motivation: Existing reasoning-augmented methods lack self-correction and experience accumulation, limiting robustness on ambiguous or previously unseen cases and motivating self-reflective reasoning for fake-news detection.The stated limitation provides the motivation for exploring self-reflective reasoning in FND.
3 Proposed Method
SRM-FND combines self-reflective reasoning construction, topic-adaptive multimodal fine-tuning, and confidence-driven cross-sample review for short-video fake news detection. It represents each sample through video, text, and metadata, then selectively specializes and verifies predictions across related events.
- Approach Overview: SRM-FND integrates contrastive CoT construction with RCCA, dual-phase topic-adaptive VLM fine-tuning, and confidence-driven cross-sample review.The three components respectively improve reasoning quality, multimodal grounding and topic specialization, and targeted verification of borderline predictions.
- Multimodal Input Preprocessing: Each sample combines keyframes, titles, ASR transcripts, on-screen text, engagement and commenter information, publication time, and publisher profiles.Samples also receive binary fake-news labels and one of four topics: Social Safety, Hard News, Specialized Knowledge, or Life & Trivia.
- Confidence-Driven Cross-Sample Review: Topic routing uses VLM classification without measurable ground-truth accuracy, while uncertain predictions trigger selective retrieval of credible and suspicious co-event witnesses.The same VLM performs topic classification during preparation and inference to preserve systematic bias consistency; uncertainty requires both aggregated real and fake scores to fall below threshold τ.
- Contrastive CoT Construction with RCCA: RCCA contrasts a Blind Analyst’s topic-specific verdict with an opposing Counter-Conclusion Reasoner before an Arbiter removes overlapping arguments and retains discriminative rationales.The Blind Analyst does not receive the ground-truth label, while the Counter-Conclusion Reasoner supports the opposite conclusion using the shared prompt.
- Dual-Phase Topic-Adaptive VLM Fine-Tuning: Dual-phase fine-tuning first reconstructs domain-specific multimodal structure, then trains lightweight topic-routed branches with conclusion-first verdict-and-CoT supervision.The first phase is decoupled from fake-news labels and aligns frames, transcripts, on-screen text, metadata, and publisher signals; LoRA is used for adaptation.
4 Experiments
SRM-FND is evaluated under in-domain and cross-dataset settings on FakeSV and FakeTT, with strong performance across backbones and topics. Iterative refinement reaches good performance after about three iterations, while later iterations mainly improve cross-dataset results.
- Framework Settings: The framework uses a shared underlying model across deliberation roles and the RCCA optimizer, with speech transcribed by Qwen3-ASR before textual incorporation.Backbones include Qwen3-VL-8B-Thinking and InternVL3.5-8B, while analyst roles and RCCA can use GPT-5.1 or Claude-4.5.
- Datasets and Baseline Settings: CoT construction and refinement use only the training split, validation selects checkpoints and hyperparameters, and the test set remains unused during framework preparation.This design balances performance and inference cost while preventing test-set involvement in reasoning generation.
- In-Domain Results: 91.33 overall accuracy on FakeSV and 92.31 on FakeTT with Qwen3-VL surpass FakeSV-VLM by 1.11 and 3.01 points, respectively.With InternVL3.5, SRM-FND achieves 90.04 on FakeSV and 91.30 on FakeTT, outperforming prior baselines.
- In-Domain Results: Accuracy exceeds 88% across all FakeSV topics and 90% across all FakeTT topics with Qwen3-VL, while InternVL3.5 exceeds 86% and 89%, respectively.Table 3 reports these per-topic results for the two backbones, supporting consistent topic-level performance.
- Accuracy Trend over Iterations: After about three iterations, the fine-tuned model reaches good in-domain and cross-dataset performance, with later iterations yielding increasingly marginal gains.When the Blind Analyst peaks, the corresponding fine-tuned model also reaches peak in-domain and cross-dataset performance.
5 Conclusions
SRM-FND is a self-reflective multimodal reasoning framework for short-video fake news detection. It combines reflective reasoning, topic-adaptive optimization, and confidence-driven cross-sample review to improve grounding and inference reliability.
- 5 Conclusions: SRM-FND proposes a self-reflective multimodal reasoning framework for short-video fake news detection.
- 5 Conclusions: Contrastive deliberation, iterative root-cause diagnosis, and corrective prompt refinement produce higher-quality self-reflective chains of thought.
- 5 Conclusions: Dual-phase topic-adaptive LoRA optimization and confidence-driven cross-sample review improve multimodal grounding and inference reliability.
- 5 Conclusions: Extensive experiments demonstrate SRM-FND’s effectiveness, robustness, and interpretability.