C$^2$DLM: Causal Concept-Guided Diffusion Large Language Models

TL;DR

C$^2$DLM introduces a causal concept-guided approach to diffusion language models, enhancing reasoning by explicitly modeling causal relationships, resulting in improved performance and training efficiency.

Contribution

The paper proposes a novel causal concept-guided diffusion language model that explicitly incorporates causal structures into attention mechanisms, improving reasoning capabilities.

Findings

Achieves 12% improvement in COT-OrderPerturb task with 3.2x faster training

Gains an average of 1.31% across six reasoning tasks

Effectively models causal relationships between concepts

Abstract

Autoregressive (AR) language models and Diffusion Language Models (DLMs) constitute the two principal paradigms of large language models. However, both paradigms suffer from insufficient reasoning capabilities. Human reasoning inherently relies on causal knowledge and thought, which are reflected in natural language. But in the AR paradigm, language is modeled as next token prediction (a strictly left-to-right, token-by-token order), whereas natural language itself exhibits more flexible causal structures. In the DLM paradigm, the attention mechanism is fully connected, which entirely disregards causal order. To fill this gap, we propose a \underline{\textbf{C}}ausal \underline{\textbf{C}}oncept-Guided \underline{\textbf{D}}iffusion \underline{\textbf{L}}anguage \underline{\textbf{M}}odel (C$^2$DLM). Starting from DLM's fully connected attention, C$^2$DLM first obtains a concept-level causal graph from the teacher model, and then explicitly guides attention to learn causal relationships between concepts. By focusing on causal relationships and avoiding interference from difficult subgoals involving causal inversion, C$^2$DLM improves 12\% with about 3.2 times training speedup in the COT-OrderPerturb task, and achieves an average gain of 1.31\% across six downstream reasoning tasks. More details in the repository ~\href{https://github.com/Kairong-Han/C-2-DLM}{here}.