A Context Adaptive Instruction Tuning Framework for Diverse Clinical NLP Tasks

Abstract

Instruction tuning adapts large language models (LLMs) to downstream tasks by training them to follow naturallanguage instructions. However, current biomedical instructiontuning pipelines rely largely on fixed instruction templates or randomly sampled instruction template pools, lacking contextaware or structured mechanisms for selecting instructions during fine-tuning. These limitations reduce semantic alignment between inputs and instructions and overlook task-specific differences in instruction sensitivity, which affect model performance in clinical NLP. We introduce CAIT (Context-Adaptive Instruction Tuning), a configurable framework that systematically explores instruction selection through three complementary mechanisms: (1) contextadaptive matching using keyword-driven clinical context detection, (2) deterministic round-robin cycling to ensure balanced template exposure, and (3) controlled random sampling to preserve exploration. Evaluation on three representative clinical NLP tasks, Medical Question Answering, Clinical Diagnosis, and Clinical Reasoning, demonstrates significant task-dependent performance variation. Round-robin selection improves diagnostic accuracy by 6.7% over random sampling (0.64 vs. 0.60), while context-adaptive selection increases clinical-reasoning conclusion accuracy by 5.7% (0.37 vs. 0.35). Medical QA exhibits relative insensitivity to selection strategy, indicating that optimal instruction selection is task-specific rather than universal. Future studies should explore semantically grounded approaches, such as embedding-based similarity or contextual bandit formulations for enhanced adaptive instruction selection.

Ähnliche Arbeiten