The paper introduces a novel training paradigm—model immunization—where curated, labeled falsehoods are periodically injected into the training of language models, treating them as “vaccine doses” to proactively enhance the model’s resistance to misinformation without degrading its general performance. Specifics below:
Model Immunization Paradigm: Introduces a novel training strategy where LLMs are fine-tuned with a small fraction (5–10%) of explicitly labeled falsehoods, treating them as “vaccine doses” to proactively build resistance against misinformation.
Distinct from Adversarial and RLHF Training: Unlike adversarial training (which defends against perturbed inputs) and RLHF (which uses preference signals), this approach uses supervised falsehood labeling during training to teach models what not to believe or propagate.
Four-Stage Training Pipeline: Consists of (1) data quarantine of curated falsehoods, (2) micro-dosed fine-tuning with corrective supervision, (3) validation against adversarial and factual prompts, and (4) post-deployment monitoring with booster updates and governance oversight.
Improved Truthfulness with Retained Accuracy: Proof-of-concept on GPT-2 XL showed a +18% gain in truthfulness on misinformation prompts (60% → 78%) with only a 1% drop in general QA accuracy, demonstrating robust misinformation resistance without knowledge loss.
Ethically Governed and Scalable: Embeds safeguards for transparency, accountability, and value alignment; designed to be modular and complementary to existing alignment methods (e.g., RLHF, post-hoc filters).