ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing
Researchers have developed ALIGNBEAM, a novel method for enhancing the safety of large language models without altering their weights. This technique addresses the issue of domain fine-tuning degrading model safety by enabling alignment transfer even between models with different vocabularies. ALIGNBEAM operates at inference time, using a small LLM judge to select the safest continuation from multiple candidates, thereby improving refusal rates on adversarial benchmarks while maintaining task accuracy and practical inference overhead. AI
IMPACT Enables cross-family LLM safety alignment without retraining, potentially improving the security of deployed models.