Unverified50% confidenceFactExact time
When targeted binary questions are replaced with generic prompts of the form "is this video physically plausible?", the VLM judge's accuracy drops to the level of random guessing.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMultimodal large models have extremely low factual recall accuracy in video question answering, which is difficult to improve even with supervised fine-tuning.69% similarUnverifiedThe study found that models' accuracy on semantically plausible distorted statements is actually higher than their performance on meaningless random replacements.66% similarUnverifiedThe author hypothesizes that the hybrid approach produces fewer false positives and more actionable findings than 'asking the LLM whether it looks right', but this hypothesis has not yet been measured.66% similarUnverifiedThe wording used to ask a vision model has a huge impact on accuracy: when asked 'Are they the same type of product?', 7 out of 11 different chargers were let through; after switching to a stricter phrasing, only 1 out of 11 passed.66% similarUnverifiedAnswer balancing requires that the yes/no labels of each question type be roughly matched in distribution, so that the expected accuracy of random guessing stays around 50%.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1038895API
curl https://kongchang.com/api/v1/knowledge/claims/1038895MCP
get_claim(id=1038895)