RETHINK-MH: Reliability-Triggered Multimodal Rethinking for Mental Health Screening
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Given the sensitive and identifiable nature of clinical interviews, multimodal mental health screening increasingly relies on released, de-identified measurements rather than raw recordings. Most existing systems compress these measurements into a single risk score, leaving no way to inspect which observations support a decision, which contradict it, or how conflicts between them were resolved. We present RETHINK-MH, a reliability-triggered multimodal rethinking framework that turns screening into an auditable two-pass process. A deterministic compiler organizes the acoustic, visual, and transcript measurements of each session into a two-layer evidence store: a coarse segment layer for the first-pass assessment and an indexed atomic layer that remains sealed for re-examination. An audit module estimates the error risk of each first-pass judgment from source reliability and validity-weighted cross-source disagreement, and only flagged cases trigger targeted retrieval of atomic evidence under an explicit access budget. A deterministic validator closes each case as preserved, revised, or unresolved, and unresolved or invalid cases are referred to human review, so every decision carries a checkable trail from outcome back to evidence. The loop is trained on its own cross-fitted errors via cold-start supervised fine-tuning followed by iterative preference alignment. On DAIC-WOZ, E-DAIC, and D-Vlog, RETHINK-MH achieves competitive and state-of-the-art screening performance, while correcting more first-pass errors than it introduces and recovering correct decisions for cases that would otherwise be referred to human review. Code is available at https://anonymous.4open.science/r/rethink-mh-85F3/.