💬 NLP

Reason2Decide: Rationale-Driven Multi-Task Learning

Reason2Decide is a two-stage, rationale-driven multi-task learning framework that effectively mitigates exposure bias and aligns explanations with predictions in clinical decision support, achieving superior accuracy and fidelity on medical datasets using models 40 times smaller than contemporary foundation models.

Original authors: H M Quamran Hasan, Housam Khalifa Bashier, Jiayi Dai, Mi-Young Kim, Randy Goebel

Published 2026-03-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: H M Quamran Hasan, Housam Khalifa Bashier, Jiayi Dai, Mi-Young Kim, Randy Goebel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are training a junior doctor to make decisions in a busy emergency room.

The Problem:
Usually, when we teach a computer (an AI) to make medical decisions, we show it a patient's notes and the "correct" answer (e.g., "Send to the ER"). We also show it the "correct" explanation for that answer.

The Trap: The AI learns to memorize the correct answer and the correct explanation separately.
The Reality: When the AI is alone in the ER later, it has to guess the answer first. If it guesses wrong, it then has to write an explanation for its wrong guess. But it was never trained to do that! It's like a student who only ever practiced writing essays about the correct history, but when asked to explain a wrong date they wrote on a test, they freeze or make up nonsense. This is called Exposure Bias.

The Solution: "Reason2Decide"
The authors created a new training method called Reason2Decide. Think of it as a two-semester course for our AI student.

🎓 The Two-Stage Training Plan

Stage 1: The "Explanation Bootcamp"
Before the AI is allowed to make any diagnoses, it spends the first semester just learning how to write good medical explanations.

The Analogy: Imagine a medical student who spends a whole month just reading and rewriting case studies, learning how to argue a point, without worrying about getting the diagnosis right yet. They learn the "language" of medical reasoning.
The Twist: The AI learns this using explanations written by other AI models (LLMs), not just humans. This is a big deal because human-written notes are expensive and rare. The paper shows that AI-written notes are good enough to teach the basics.

Stage 2: The "Real-World Simulation"
Now, the student enters the second semester. They must now do two things at once: Diagnose AND Explain.

The Challenge: In the real world, the AI's diagnosis might be wrong. If it says "Go home" when it should have said "Go to ER," it needs to explain why it thought "Go home" was right.
The Secret Sauce (Scheduled Sampling): This is the paper's clever trick.
- At the start of the semester, the teacher (the computer) helps the student by saying, "Here is the correct diagnosis, now explain it."
- As the semester goes on, the teacher slowly stops helping. They start saying, "Here is your diagnosis (even if it's wrong), now explain that."
- By the end, the student is used to explaining their own mistakes and predictions, not just the perfect ones. This bridges the gap between training and the real world.

🏥 The Results: Why It Matters

The researchers tested this on three different medical tasks:

Triage: Deciding if a patient needs to go home, see a doctor in 24 hours, or go to the ER immediately.
Medical Q&A: Answering complex biology questions.

The Wins:

Better Accuracy: The AI made fewer mistakes than other methods.
Better Explanations: When the AI made a decision, its explanation actually matched that decision. It didn't say "Go to the ER" and then write an explanation for "Go home."
Small but Mighty: The AI model they used was 40 times smaller than the massive, famous AI models everyone talks about. It's like using a smart, compact sedan instead of a giant, fuel-guzzling truck to get the job done. This makes it cheaper and easier to run in hospitals with limited computers.
Robustness: Even when they trained it with AI-generated notes (instead of human notes), it still performed incredibly well. This means we might not need to pay humans to write thousands of explanations to train these systems.

🌟 The Big Picture

Reason2Decide is like teaching a student to drive by first letting them practice steering in a parking lot (Stage 1), and then gradually letting them drive on the highway while they are still learning to navigate traffic (Stage 2).

Instead of just memorizing the map, the AI learns to think about its own decisions. This makes it a safer, more trustworthy tool for doctors, because when it says, "I think this patient needs an ambulance," it can actually tell you why it thinks that, based on its own logic.

1. Problem Statement

The paper addresses a critical gap in Clinical Decision Support Systems (CDSS) using Large Language Models (LLMs): the misalignment between predictive accuracy and explanation fidelity.

Exposure Bias: Current self-rationalization models are trained to generate explanations conditioned on gold-standard (ground truth) labels. However, during inference, they must explain their own predictions, which may be incorrect. This train-test discrepancy leads to models generating plausible-sounding but logically inconsistent explanations for their own errors.
Task Separation: Many existing approaches treat prediction and explanation as separate tasks or secondary outputs, failing to enforce a tight coupling where the explanation is an integral part of the reasoning process for the specific prediction made.
Resource Constraints: High-performance foundation models are often too large for resource-constrained clinical deployments, yet smaller models struggle to provide both accurate predictions and high-quality, aligned explanations.

2. Methodology: Reason2Decide

The authors propose Reason2Decide, a two-stage training framework designed to train a single encoder-decoder model (T5 variants) to jointly predict labels and generate self-consistent rationales.

Core Architecture

Model: Uses standard T5 (Small, Base, Large) without architectural modifications. It leverages the text-to-text formulation to handle both discrete label prediction (e.g., "Go to ED Now") and free-text rationale generation under a single objective.
Input/Output: The model takes a clinical note or question ( $x$ ) and outputs a predicted label ( $\hat{y}$ ) and a rationale ( $\hat{r}$ ).

Stage 1: Rationale Foundation Training

Goal: Teach the model the fundamentals of explanation generation.
Process: The model is trained only on generating rationales conditioned on the gold label ( $y^*$ ).
Loss Function: $L_{stage1} = -\log P_\theta(r^* | \text{explain: } x)$
Outcome: The model learns to generate coherent medical explanations before being tasked with making its own predictions.

Stage 2: Joint Optimization with Task-Level Scheduled Sampling

Goal: Jointly optimize prediction and explanation while mitigating exposure bias.
Mechanism: The model predicts a label ( $\hat{y}$ ) and then generates a rationale conditioned on that prediction.
Task-Level Scheduled Sampling: Unlike traditional token-level scheduled sampling, this method switches the conditioning context from the gold label to the model's predicted label.
- For each training step $t$ $t$ , the conditioning label $\tilde{y}$ $\tilde{y}$ is chosen as:
  - Gold label ( $y^*$ ) with probability $1 - \pi_t$
  - Predicted label ( $\hat{y}$ ) with probability $\pi_t$
- Schedule: $\pi_t$ follows a linear schedule with a warm-up phase ( $w=0.05T$ ) and a transition phase ( $m=0.60T$ ), capping at 0.9 to prevent total reliance on potentially erroneous self-predictions during training.
Loss Function: A weighted sum of prediction loss and explanation loss:
- $L_{total} = \alpha_t L_{pred} + (1 - \alpha_t) L_{expl}$
- $\alpha_t$ is adjusted to prioritize explanation loss initially, then stabilizes to balance both tasks.

Inference

During inference, the model first greedily decodes the label $\hat{y}$ , then generates the rationale $\hat{r}$ conditioned strictly on $\hat{y}$ (self-conditioning), ensuring the explanation justifies the actual prediction made.

3. Key Contributions

Two-Stage Framework: A novel curriculum learning approach that separates rationale learning from joint optimization, allowing the model to master explanation fundamentals before tackling the harder task of self-rationalization.
Task-Level Scheduled Sampling: A specific adaptation of scheduled sampling that transitions the conditioning context from gold labels to predicted labels, directly addressing the exposure bias problem in self-explainable models.
Synthetic Rationale Pretraining: Demonstrated that LLM-generated rationales can effectively substitute for human-authored rationales during the pretraining (Stage 1) phase, significantly reducing reliance on costly human annotations.
Efficiency: Achieves state-of-the-art performance using models 40× smaller (e.g., T5-Large at 800M parameters) than contemporary foundation models (e.g., Qwen-3-32B), making clinical reasoning accessible for resource-constrained environments.

4. Experimental Results

The framework was evaluated on three datasets:

Clinical Triage (Proprietary): Nurse triage notes with 12 disposition classes.
PubMedQA: Biomedical QA (Yes/No/Maybe).
BioASQ: Biomedical QA (Yes/No).

Key Findings:

Predictive Performance: Reason2Decide outperformed standard fine-tuning (SFT), Distilling Step-by-Step (DSS), and zero-shot LLMs across all datasets and model sizes in terms of Macro-F1 and Accuracy.
- Example: On Clinical Triage (T5-Small), Reason2Decide achieved 55.88 F1, beating DSS-F1 (52.09) and SFT (52.69).
Rationale Fidelity: The model generated explanations with higher semantic similarity (BERTScore) and surface overlap (BLEU) to gold rationales compared to baselines.
LLM-as-a-Judge: Evaluated by an expert LLM (Qwen-3-8B) on Correctness, Coverage, and Overlap. Reason2Decide consistently scored highest, particularly in Correctness (alignment between the rationale and the predicted disposition).
Source Robustness: When trained on LLM-generated rationales in Stage 1, the model maintained robust performance across different rationale styles (Nurse-authored, Post-processed, LLM-generated) in Stage 2, showing minimal F1 variation compared to baselines.
Efficiency: T5-Large (800M) outperformed Qwen-3-32B (32B) on all three datasets, narrowing the gap between small and large models.

5. Significance and Impact

Trust and Adoption: By ensuring that explanations are logically consistent with the model's own predictions, the system enhances trust in AI-driven clinical decisions, a prerequisite for real-world adoption.
Cost Reduction: The ability to use LLM-generated rationales for pretraining reduces the dependency on expensive, time-consuming human expert annotations.
Accessibility: The framework proves that high-quality, explainable clinical decision support does not require massive foundation models, enabling deployment in settings with limited computational resources.
Generalizability: The approach addresses fundamental training dynamics (exposure bias) rather than just inference-time strategies, offering a robust solution for self-rationalization in any domain requiring explainable predictions.