generalopinionbullishLLMs can capture semantics and produce high-quality code summaries better than task-specific Transformer modelsComputation and Language28 Jul 2026http://arxiv.org/abs/2607.20933v1