@ericyocam6Professor, California Polytechnic State University
Shipped: Adaptive Activation Cancellation for LLM Hallucination Mitigation
Adaptive Activation Cancellation: a method that reduces LLM hallucinations by suppressing the internal activations that cause them.
llmhallucinationinterpretabilityai-safetynlp
Comments1
Suppressing activations at inference is appealing because it avoids retraining, but how do you identify the hallucination-linked directions? If they come from probes, I wonder how well they transfer across domains, and whether cancelling them costs fluency or recall on correct answers.
1 comment
Suppressing activations at inference is appealing because it avoids retraining, but how do you identify the hallucination-linked directions? If they come from probes, I wonder how well they transfer across domains, and whether cancelling them costs fluency or recall on correct answers.