MERID is a framework that builds depression analysis pipelines via Recursive Self‑Improvement (RSI). It first employs Grounded State Construction (GSC) to align multimodal records such as interview audio, transcripts, and sensor readings with subject‑level depression targets, creating an experience base. Then Coupled Pipeline Exploration (CPE) jointly searches successor pipelines across representation learning, feature fusion, and predictor design, improving both classification and severity estimation. Evidence‑Guided Evolution (EGE) evaluates revisions on small depression cohorts using experimental feedback, verifies gains under uncertainty, and only inherits successful pipelines to the next iteration. Extensive experiments on public depression benchmarks show MERID outperforms traditional multimodal models and agent‑based baselines on multiple tasks, with acoustic and linguistic cues proving especially valuable. The code is publicly available.
Review