AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

Salus: Strategic Diagnostic Testing for Complex Diagnosis via Multi-Agent Reinforcement Learning

Shuohao Gao, Xuanzhong Chen, Lingxiao Luo, Zilin Ding, Rong Han, Rui Jiang, Ting Chen

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Salus, a multi-agent reinforcement learning approach that trains LLM-based agents to strategically request diagnostic tests over multiple rounds until reaching a definitive diagnosis of complex diseases.

Problem

LLMs used as clinical assistants struggle with the sequential, iterative process of complex diagnosis, in which clinicians must request multiple rounds of diagnostic tests to differentiate similar diseases. The abstract is truncated, so the specific failure modes beyond premature diagnosis are not s

Approach

Frames strategic diagnostic testing as a multi-agent reinforcement learning problem so that agents learn when and which diagnostic tests to request before committing to a diagnosis. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add multi-agent RL (DR and SC) after SFT (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 55.02 -> 83.64Improved
add RL to the Strategic Controller (in addition to DR) (ablation)Salus-7B · 269 high-complexity casesTop-1 accuracy: 61.71 -> 83.64Improved
add RL for DR only after SFT (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 55.02 -> 61.71Improved
add RL for SC (ablation)Salus-7B · 269 high-complexity casesInformation sufficiency: 65.43 -> 72.86Improved
increase SFT data from 12.5% to 100% (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 44.61 -> 55.02Improved
use 7B backbone instead of 3B for SFT (ablation)Salus-Zero · 269 high-complexity casesTop-1 accuracy: 43.87 -> 55.02Improved
use DeepSeek-V3.1 as teacher instead of stronger DeepSeek-V3.2 (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 43.12 -> 55.02Improved
use DeepSeek-V3.1 teacher instead of GLM-Z1-Flash (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 31.97 -> 55.02Improved
increase SFT data from 50% to 100% (ablation)Salus-Zero-7B · 269 high-complexity casesTop-1 accuracy: 54.28 -> 55.02Improved
apply Salus RL framework to 3B backboneQwen2.5-3B-Instruct · 269 high-complexity casesTop-1 accuracy: 19.33 -> 65.06Improved
apply Salus training to Qwen2.5-7B-InstructQwen2.5-7B-Instruct · 269 high-complexity casesTop-1 accuracy: 27.88 -> 83.64Improved
apply Salus training on routine casesQwen2.5-7B-Instruct · 483 routine casesTop-1 accuracy: 59.63 -> 91.10Improved

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy