---
license: odbl
---


**CLIR-Bench** is an evidence-auditable benchmark for **multimodal question answering over irregular clinical time series**, constructed from de-identified ICU records in MIMIC-IV. It contains 6,600 multiple-choice QA instances spanning 11 clinical variables and intervention signals, organized into four core capability dimensions: temporal understanding, reasoning, forecasting, and decision-making. Each question is explicitly linked to timestamp-level evidence and task-specific answer derivation rules, enabling comprehensive evaluation not only of answer accuracy but also of evidence faithfulness, evidence reliance, and causal sensitivity. By focusing on sparse, asynchronous, missing, and long-horizon clinical observations, CLIR-Bench provides a realistic and diagnostic testbed for assessing whether models can reliably identify and reason over patient-specific temporal evidence.
> ⚠️ **Note:** The current version does **not include time-series data and evidence derived from MIMIC-IV**.  
> Access to data involving MIMIC-IV-related tasks requires additional credentialing and approval from PhysioNet.  
> Researchers can apply for access via the official website: [MIMIC-IV](https://physionet.org/content/mimiciv/).


## 📊 CLIR-Bench Overview

CLIR-Bench is an evidence-auditable benchmark for multimodal QA over irregular clinical time series.

![Pipeline](./bench.png)


## 📄 Paper

[Download PDF](https://arxiv.org/pdf/2607.09880)