Escaping Confidence Trap: Evolutionary Decoding for Mathematical Reasoning in Diffusion LLMs

2026-08-01

AI Summary

Diffusion-based large language models are prone to erroneous reasoning paths in mathematical tasks due to locally high-confidence generations that compromise global coherence. This work identifies, for the first time, two underlying failure mechanisms: overconfident symbolic generation and repetitive reasoning patterns. To address these issues, the authors propose a training-free evolutionary decoding framework that models diffusion decoding at inference time as an evolving population of candidate reasoning states. The approach iteratively selects promising states by preserving effective symbolic-numerical signals and introduces block-level mutations to generate structured alternative pathways, thereby escaping misleading high-confidence regions. Evaluated across multiple mathematical reasoning benchmarks, this method substantially outperforms conventional confidence-based decoding strategies and significantly enhances the reasoning reliability of the LLaDA 2.0 model.

Abstract

Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive LLMs, offering efficient generation through block-wise progressive unmasking. However, their strong general-purpose performance does not necessarily translate into reliable mathematical reasoning, where correctness depends on preserving coherent numerical-symbolic reasoning trajectories. In this work, we analyze the decoding trajectories of LLaDA 2.0 and identify a recurring diffusion confidence trap: local token confidence can become misaligned with global reasoning correctness during progressive block decoding. Our analysis reveals two representative failure regimes: sampling-sensitive failures, where correct paths exist but are unstable, and sampling-consistent failures, where repeated sampling converges to repetitive high-confidence but incorrect continuations. Motivated by this observation, we propose Evolutionary Decoding, a training-free test-time scaling framework that views diffusion decoding as an evolutionary process over candidate reasoning states. The framework combines step-wise selection, which preserves useful numerical-symbolic signals and suppresses repetitive patterns, with block-wise mutation, which introduces structured alternatives to escape incorrect high-confidence basins. Experiments on multiple benchmarks show that Evolutionary Decoding improves LLaDA 2.0 over confidence-based decoding, leading to more reliable mathematical reasoning.

Problem

• diffusion LLMs
• mathematical reasoning
• confidence trap
• decoding trajectory
• reasoning correctness

Innovation

• Evolutionary Decoding
• Diffusion LLMs
• Mathematical Reasoning
• Confidence Trap
• Test-time Scaling