Tag
This paper investigates authority bias in LLMs using a controlled medical QA setting, revealing that models override correct answers in a graded manner proportional to perceived authority. The effect is localized to a critical late layer where correct answer representations are actively erased.
The paper proposes BiRG-LoRA, a rank-gated LoRA method for medical question answering that uses clinically structured priors to select sparse rank subsets, achieving 69.31% macro-average accuracy across four benchmarks while using fewer parameters than mixture-of-experts approaches.
This paper introduces a multi-agent peer-reviewed reasoning method where multiple LLMs independently generate chain-of-thought reasoning and then evaluate each other's outputs to select the best answer. The method outperforms single-model reasoning and majority voting on medical QA benchmarks.
This paper proposes a Variance-Aware Reward Framework using GRPO to improve LLM performance on heart-focused medical question answering, achieving significant accuracy and F1 gains on a HealthBench subset.
Researchers introduce DoseBench, a benchmark of 81 OTC dosing scenarios to evaluate LLM decision-making under temporal uncertainty for acetaminophen and ibuprofen use. Results show LLMs frequently struggle with rolling-window reasoning and can produce confident but medically unsupported responses.
A tutorial and project demonstrating LoRA fine-tuning of Qwen3-1.7B on AMD MI300X using ROCm for clinical question answering, providing a CUDA-free alternative for medical AI development.