Enhancing LLM Decision-Making with Probabilistic Outputs and Verbalized Explanations

Why It Matters

In design practice, the ability to understand not just *what* a system predicts, but also *how confident* it is and *why*, is crucial for building trust and enabling effective human-AI collaboration. This research offers a pathway to more transparent and reliable AI-driven decision support tools.

Key Finding

The CLSGen framework successfully enables LLMs to be more accurate in classification and produce explanations that are both relevant to the prediction and easy to understand.

Key Findings

• Models fine-tuned with CLSGen outperform existing baselines in classification metrics (AUROC and F1-score).
• CLSGen-tuned models demonstrate strong alignment between predicted labels and generated justifications.
• The generated explanations exhibit high readability.

Research Evidence

Aim: How can Large Language Models be fine-tuned to simultaneously provide accurate probabilistic classifications and generate meaningful verbalized explanations without compromising either capability?

Method: Framework Development and Empirical Evaluation

Procedure: The CLSGen framework was developed, incorporating a new model architecture, training methodology, and data construction strategy. This framework was then used to fine-tune LLMs for binary classification tasks. The performance of these fine-tuned models was evaluated against existing baselines on benchmark datasets, assessing both classification metrics and the quality of generated explanations.

Context: Natural Language Processing, Artificial Intelligence, Machine Learning

Design Principle

AI systems should strive for both predictive accuracy and explainability, ensuring that users can understand the basis and confidence level of AI-generated outputs.

How to Apply

When integrating LLMs into a design project that requires decision-making, explore fine-tuning techniques that explicitly aim to preserve or enhance explanation capabilities alongside predictive performance.

Limitations

The framework is currently focused on binary classification tasks and its performance on more complex, multi-class problems or different modalities is not detailed. The potential for 'catastrophic forgetting' in very long fine-tuning processes might still exist.

Student Guide (IB Design Technology)

Simple Explanation: This research shows a new way to train AI language models so they can give a score for how sure they are about an answer and also explain why they gave that answer, without messing up either part.

Why This Matters: It's important for design projects that use AI to be understandable and trustworthy. This research provides a method to make AI outputs clearer, which is key for user acceptance and effective use.

Critical Thinking: To what extent can the 'verbalized explanation' generated by an LLM truly reflect its internal decision-making process, or is it merely a plausible narrative constructed post-hoc?

IA-Ready Paragraph: The development of AI-driven decision support systems necessitates a focus on both predictive accuracy and interpretability. Research such as CLSGen (Yoon et al., 2026) presents a framework for fine-tuning Large Language Models to provide reliable probabilistic outputs alongside coherent verbalized explanations, addressing a critical gap in current AI deployment where models often sacrifice one capability for the other.

Project Tips

• Consider how your design project could benefit from AI that not only makes a prediction but also explains its confidence.
• If using LLMs, investigate fine-tuning methods that balance accuracy with interpretability.

How to Use in IA

• Reference this research when discussing the development of AI components in your design project, particularly if interpretability or probabilistic outputs are a focus.

Examiner Tips

• Demonstrate an understanding of the trade-offs between model performance and interpretability when discussing AI integration in your design.

Independent Variable: Fine-tuning framework (CLSGen vs. traditional methods)

Dependent Variable: Classification metrics (AUROC, F1-score), Explanation quality (alignment, readability)

Controlled Variables: LLM architecture, Benchmark datasets, Training data characteristics

Strengths

• Addresses a significant practical challenge in LLM deployment.
• Provides empirical evidence of improved performance in both classification and explanation.
• Introduces a novel, integrated framework.

Critical Questions

• How does the CLSGen framework generalize to different types of classification problems beyond binary classification?
• What are the computational costs associated with this dual-head fine-tuning approach compared to single-objective methods?

Extended Essay Application

• An Extended Essay could investigate the impact of CLSGen-derived explanations on user trust and decision-making in a specific domain, such as medical diagnosis or financial risk assessment.

Source

CLSGen: A Dual-Head Fine-Tuning Framework for Joint Probabilistic Classification and Verbalized Explanation · arXiv preprint · 2026