[ASCENT: A Benchmark for Evaluating and Advancing Stepwise Diagnostic Reasoning in Large Language Models on Common Clinical Scenarios]

Table of Contents

• [ASCENT: A Benchmark for Evaluating and Advancing Stepwise Diagnostic Reasoning in Large Language Models on Common Clinical Scenarios]
  • Table of Contents
  • Paper-to-Code Mapping
  • Dataset
  • Installation
  • Usage
    • 1. Training
    • 2. Inference
    • 3. Evaluation
    • 4. Post-Processing
  • Reproducibility

---

Paper-to-Code Mapping

Paper Section	Code Location	Description
Experimental Settings - models	sft.py	Training details
Experimental Settings - Inference Settings	inference.py	Inference details
Evaluation Metrics	evaluation.py	LLM-as-a-Judge & details
Evaluation Metrics	postprocessing.py	Post-processing LLM Results

Each main function or class in the code is annotated with the corresponding paper section as a comment.
(e.g., # Load configuration as described in the "Experimental Settings – Models" section of the paper.)

---

Dataset

The ASCENT dataset contains two tasks:

Task	Directory	Description
(1) Impressions (Imp)	data/ascent	Generating impressions only
(2) Impressions + Rationales (Imp + Reason)	data/ascent_w_reason	Generating supporting rationales followed by impressions

---

Installation

pip install -r requirements.txt

Usage

1. Training

python3 sft.py --config sft

2. Inference

python3 inference.py --config inference

3. Evaluation

python3 evaluation.py --config evaluation

4. Post-Processing

python3 postprocessing.py --config evaluation

Reproducibility

• All experiments in the paper can be reproduced using this code and the provided configuration files.
• We fixed the random seed (42) for all runs (see the seed parameters in the YAML files in the config/ directory).
• To reproduce all main results, run:

./sft.sh && python3 evaluation.py && python3 postprocessing.py

License

ASCENT
Copyright (c) 2026-present NAVER Corp.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.