Distorted or Fabricated? A Survey on Hallucination in Video LLMs

This project is a survey paper and theoretical framework rather than a software product. While it provides a valuable taxonomy (dynamic distortion vs. content fabrication) for a high-growth area, it lacks a technical moat. The defensibility is low (2) because the value lies in the intellectual categorization which is easily reproducible and likely to be superseded as Video LLM architectures evolve from CLIP-based frame encoders to native spatio-temporal transformers (like Sora or Gemini 1.5 Pro). The 7 forks within 3 days indicate immediate academic interest, but the absence of stars suggests it's currently circulating within research circles. Frontier labs (OpenAI, Google, Meta) are the primary stakeholders in Video LLM development and are actively building internal proprietary evaluation suites that address these exact hallucination types, making the risk of platform domination high. This work is most useful as a reference for researchers building new benchmarks (like Video-MME or MVBench) rather than a standalone tool.