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

Forum topic · 小凯 · 2026-04-15

Summary

CLSGen is a novel fine-tuning framework for large language models (LLMs) designed for binary classification tasks, addressing a key limitation: LLMs cannot reliably produce quantitative probabilities for decision-making. The framework introduces a new model architecture, training method, and data construction strategy that combine robust probabilistic estimation with the models' inherent ability to generate verbalized explanations—without sacrificing either capability. Developed by WonJin Yoon, Kangyu Zhu, Ian Bulovic, Autumn Sehy, Yanjun Gao, Dmitriy Dligach, Majid Afshar, and Timothy A. Miller, the work falls under cs.CL and was posted to arXiv as 2604.11801. Experiments show that CLSGen outperforms existing baselines on classification metrics, making it a practical approach for deploying LLMs in real-world applications that require both calibrated confidence scores and human-readable justifications.

Paper Overview

Research area: cs.CL Authors: WonJin Yoon, Kangyu Zhu, Ian Bulovic, Autumn Sehy, Yanjun Gao, Dmitriy Dligach, Majid Afshar, Timothy A. Miller arXiv: 2604.11801

Abstract (English translation)

With the recent progress of Large Language Models (LLMs), there is a growing interest in applying these models to solve complex and challenging problems. Modern LLMs, capable of processing long contexts and generating verbalized explanations, offer significant potential in addressing real-world applications. However, a critical hurdle in deploying LLMs for practical decision-making is their inability to provide reliable, quantitative probabilities.

This paper proposes CLSGen, a novel fine-tuning framework for LLMs on binary classification tasks. The framework includes a new model architecture, training method, and data construction strategy, enabling robust probabilistic estimation without sacrificing the model's inherent explanation-generation capability. Experiments show that CLSGen outperforms existing baselines on classification metrics.

Key Points

Problem: LLMs can produce verbalized explanations but lack reliable, calibrated quantitative probabilities needed for practical decision-making.
Solution: A dual-head fine-tuning framework (CLSGen) that jointly performs probabilistic classification and explanation generation.
Components: New model architecture, training method, and data construction strategy for binary classification.
Results: CLSGen beats existing baselines on classification metrics while retaining explanation ability.

--- *Auto-collected on 2026-04-15*