Hallucination
Woodpecker
- 기존: 추가 데이터
- 5 stages
- inconsistent with the input image
- non-existent objects
- fail to describe the attribute
- 긴 text에서 심함
- LRV-Instruction으로 학습하면 limit length
- VIGC
- multistep generation
Holistic Analysis of Hallucination in GPT-4V
- Bingo benchmark
- bias and interference
- popular: self-correction, cot는 효과적이지 않음
- 190 failure instances in GPT-4V
- region bias, OCR bias, factual bias
- image-to-image interference, text-to-text interference
- Self-correction
- Your answer is wrong. Review your previous answer and find problems with your answer. Answer me again.
OPERA
- over trust penalty and retrospection allocation
- hallucination은 knowledge aggregation patterns을 self-attention matrix에서 가깝게 한다.
- MLLM은 few summary tokens에 집중함. 모든 토큰이 아니라
- over trust issue를 줄이기 위한 penalty term