acceptodds
Desk rejected at ICLR 2027

STETHO: Synchronized Time-series and Evidence Tracking via Heterogeneous Orchestration in Multi-Agent Medical Reasoning

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal medical reasoning requires the seamless integration of heterogeneous clinical data, yet monolithic Multimodal Large Language Models (MLLMs) fre- quently suffer from modality dominance and cognitive overload, failing to simulate the rigorous, cross-examined diagnostic workflows of human experts. We propose STETHO (Synchronized Time-series and Evidence Tracking via Heterogeneous Orchestration), a novel training-free multi-agent framework designed to explicitly decouple the reasoning processes across diverse modalities. STETHO introduces an Independent Blind Review mechanism to prevent early anchoring bias, fol- lowed by a Cross-Modal Examination Protocol where a Chief Agent acts as an active adjudicator to detect logical conflicts and trigger Selective Replay of raw data for targeted reflection. Finally, a Meta-Cognitive Reasoning Refiner distills multi-agent interactions into a standardized Four-Stage Clinical Paradigm, ensuring logical transparency and grounding. Extensive evaluations on challeng- ing benchmarks, including MedFrameQA and OmniMedVQA, demonstrate that STETHO, utilizing only lightweight 7B/8B backbones, significantly outperforms massive monolithic behemoths (e.g., GPT-4o) and reinforcement, learning-based models (e.g., Med-R1). Results based on M3CoTBench metrics further confirm that STETHO achieves superior logical correctness and consistency, providing an auditable diagnostic trail that effectively bridges the gap between automated reasoning and expert-level clinical consultation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi‑agent systems, multimodal medical reasoning, vision‑language models, time‑series reasoning, clinical decision support, evidence‑grounded reasoning, hallucination mitigation, chain‑of‑thought

Related papers

Discussion (0)

Sign in to comment.