TRIAD-TS: Explainable Multi-Agent Orchestration for Joint Task–Style Alignment on On-Device Mental Health LLMs

ACL ARR 2026 January Submission10860 Authors

Keywords: On-device LLM, Personalization, LoRA, Multi-Agent

Abstract: On-device LLM personalization must satisfy two constraints simultaneously: (i) task correctness and (ii) user-specified response style, under strict resource limits. We present TRIAD-TS, a multi-agent orchestration framework that (1) decomposes a request into task and style intents, (2) retrieves evidence in an explainable latent space using precomputed intent centroids, (3) blends style adapters via an explicit weight vector, and (4) applies a learned quality-driven abstention policy that avoids unreliable outputs without conflating uncertainty with device feasibility. We also introduce TriadBench-TS, a benchmark spanning 8 task categories and 12 therapeutic communication styles, with verification that filters invalid rewrites. Using ELO-based pairwise evaluation across three dimensions (task, style, joint), TRIAD-TS achieves 3.7% higher joint ELO rating (+41 points) than state-of-the-art adapter composition methods and reduces style drift by 52% compared to instant adapter blending (CRAYON), while providing transparent rationales through centroid matches and adapter-weight explanations. Code and data are available at https://anonymous.4open.science/r/TRAID-02C1.

Paper Type: Long

Research Area: Low-resource Methods for NLP

Research Area Keywords: LLM agents, multi-agent systems, agent communication, privacy, mental health

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Data resources

Languages Studied: English

Submission Number: 10860