Researchers have published four new papers proposing distinct approaches to optimize large language model (LLM) performance across medical and general applications.
Medical Visual Question Answering
According to arxiv.org, a noise-aware framework for medical visual question answering (Med-VQA) incorporates a denoising autoencoder before visual embeddings are mapped into an LLM’s input space. The method uses parameter-efficient fine-tuning with low-rank adaptation (LoRA) and was evaluated on SLAKE and PathVQA benchmarks, showing “improved robustness to noisy input embeddings while maintaining competitive clean performance,” according to the abstract.
Telehealth Response Generation
A separate arxiv.org paper introduces a severity-aware multi-model framework that trains five LLMs using a three-stage curriculum learning strategy on mild, moderate, and critical medical cases. During inference, all models generate responses and the most appropriate is selected. According to the researchers, the method achieved 86.71% BERTScore in baseline settings and 90.30% after fine-tuning on the MAQA dataset.
Cache Efficiency and Routing
Two additional papers address LLM efficiency: QCFuse proposes a “compressed-view query-aware selector for RAG cache fusion” that uses “chunk-anchor query probing” to reduce prefill-stage costs in retrieval-augmented generation, according to arxiv.org. Meanwhile, MetaRouter introduces a meta-learning framework for “preference-aware LLM routing” that learns users’ implicit cost-performance preferences through interaction, according to another arxiv.org paper.