Depression is a prevalent mental health disorder whose varied and comorbid symptom presentation complicates timely and accurate diagnosis. This study evaluates modern encoder- and decoder-based Large Language Models (LLMs) for automated depression symptom estimation using the DAIC-WOZ dataset. We compare In-Context Learning (ICL) strategies (zero-shot, few-shot, chain-of-thought) against parameter-efficient fine-tuning (PEFT/LoRA) and linear probing techniques. Surprisingly, zero-shot ICL achieves new state-of-the-art results, outperforming fine-tuning approaches and prior benchmarks.