Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders
Multiple research papers explore methods for detecting and mitigating hallucinations in AI systems, particularly in safety-critical applications like medical imaging and document analysis. One study proposes a cross-modality framework for medical AI, highlighting that general-purpose models can outperform specialized ones in hallucination benchmarks. Another paper introduces SafeLLM, which uses extraction rather than rewriting for retrieval-augmented generation to improve safety and reduce hallucinations. Additionally, research is being done on zero-source hallucination detection using human-like criteria probing and on utilizing optimal transport and causal recurrent labelers for quicker detection of hallucination onset in various AI tasks. AI
IMPACT Developments in hallucination detection and mitigation are crucial for the safe and reliable deployment of AI in critical domains like healthcare and compliance.
- GPT-4
- GPT-3.5
- Llama-2
- Mistral-7B
- OpenHalDet
- Evidence Graph Consistency
- Whisper
- Sparse AutoEncoder
- Large Language Models
- Token-Level Visual-Sensitivity Steering
- Constrained Paraphrase Consistency
- Flan-T5
- Visual Language Models
- DeBERTa
- Diffusion Models
- SafeLLM
- retrieval-augmented generation
- FDA
- medical imaging
- Optimal Transport
- hallucination
- NICE
- Fairseq
- Human-like Criteria Probing for Hallucination Detection
- AI