Reproduce “SEASON: Mitigating Temporal Hallucination in Video Large Language Models via Self-Diagnostic Contrastive Decoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chang-Hsun Wu, Kai-Po Chang, Yu-Yang Sheng, Hung-Kai Chung, Kuei-Chun Wang, Yu-Chiang Frank Wang
SEASON introduces a training-free framework that leverages temporally-hard negatives and a token-level self-diagnostic mechanism to reduce temporal and spatial hallucinations in Video Large Language Models. This approach achieves top performance on video hallucination benchmarks among inference-time methods, improving temporal sequence hallucination scores by up to 24.5% while preserving general video understanding.
Key insights
- Creating "temporally-hard negatives" by homogenizing temporal features while preserving spatial details is effective for specifically targeting and mitigating temporal hallucinations.
- A self-diagnostic mechanism leveraging attention map divergence can dynamically identify and adaptively address spatial versus temporal hallucination tendencies at the token level.
- Training-free inference-time decoding strategies can achieve state-of-the-art hallucination mitigation, even outperforming some training-based methods, without incurring significant retraining costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.