Diffusion Language Models Excel as Training-Free Multi-Label Text Classifiers

This research introduces dLLM-SetScore, a novel approach to multi-label text classification that leverages discrete masked-diffusion language models without requiring task-specific fine-tuning. The method operates by posing a simple yes/no question for each potential label and then comparing the probabilities of the answer tokens at a masked position. This innovative technique avoids the need for extensive training on textual-entailment datasets, relying instead on a small validation set to optimize parameters like thresholds and prompt wording. A key finding addresses a common issue in multi-label classification: the positional bias when multiple labels are included in a single prompt. dLLM-SetScore circumvents this by evaluating each label independently, ensuring predictions are invariant to label order. The models LLaDA-8B and Dream-7B were evaluated against various baselines, demonstrating superior performance, particularly LLaDA-Instruct, which achieved the highest training-free metrics on Reuters and ECtHR datasets.