Training-Free Adaptation of New-Generation LLMs using Legacy Clinical Models
Researchers developed CAPT, a method that combines a new general-purpose language model with an older clinical model without requiring retraining. The technique uses contrastive decoding to blend clinical knowledge into the newer model while keeping its reasoning abilities intact, and it outperformed existing ensemble methods by 17-41% on clinical tasks. This approach helps hospitals with limited computing resources adopt newer models without the expensive process of retraining them on clinical data.
- CAPT combines a new general-purpose LLM with an old clinical model at inference time via contrastive decoding, no retraining needed, even across mismatched vocabularies.
- It beat UniTE by 17.6% and standard proxy tuning by 41.4% on average across six clinical tasks.
- Physician review and token-level analysis showed more clinically specific language, fewer context errors, and better actionable recommendations.
- Lets resource-constrained hospitals adopt newer models without expensive retraining cycles on clinical data.
Researchers have developed Cross-Architecture Proxy Tuning (CAPT), a method that lets hospitals use newer, general-purpose language models without having to retrain them on clinical data from scratch. The core problem they're solving is real: every time a better foundation model comes out, healthcare organizations either stick with older clinical models or spend massive resources retraining the new ones on medical text. CAPT sidesteps this by combining a new general-purpose model with an existing older clinical model at inference time, using contrastive decoding to blend their strengths. The trick is handling the fact that these models often have completely different vocabularies—the method routes tokens through both models intelligently, amplifying clinical signals from the older model while keeping the reasoning quality of the newer one.
The results are substantial. Across six clinical tasks (classification and text generation), CAPT beat both individual models and other ensemble approaches by wide margins—17.6% better than UniTE and 41.4% better than standard proxy tuning on average. Token-level analysis and case reviews from physicians showed the method actually produces more clinically specific language, fewer context errors, and better actionable recommendations. This matters most for resource-constrained hospitals that can't afford continuous retraining cycles but want to benefit from advances in general AI. They get the latest model capabilities without the computational overhead.