Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor probability mass into the target model’s vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the mixing strength can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on structured, front-loaded adversarial benchmarks, with a 4.6-6.6 x inference slowdown inside the band of existing inference-time defenses. The results show that early-token refusal behavior can be transferred between model families at inference time, without touching either model’s weights.

Lay summarization aims to make biomedical research accessible to non-experts, but most approaches assume a uniform audience, overlooking variation in medical literacy and information needs. We present MAPS (Multi-Agent Persona-based Summarization), a framework that generates persona-specific summaries through iterative cross-agent feedback. Human evaluation shows MAPS improves quality over single-agent baselines, while automatic metrics fail to capture these gains. LLM-based judges also exhibit limited sensitivity, assigning inflated scores and misdetecting errors. These findings highlight the need for improved evaluation methods for persona-based summarization.