Understanding Evaluation Illusion in Diffusion Large Language Models
This paper evaluates the consistency and effectiveness of decoding methods for diffusion large language models (dLLMs) across diverse evaluation settings and reveals their sensitivity to prompt templates.
Provides insights into the inconsistent evaluation results of dLLM decoding methods and proposes practical guidelines for their reliable evaluation.
Before reading this…
Applications
- →Natural language processing
- →Machine learning
To understand this paper, make sure you know these concepts first:
- Understanding of large language modelsfind papers →
- Familiarity with decoding methodsfind papers →
Abstract
More Like ThisDespite the capability of parallel decoding, diffusion large language models (dLLMs) require many denoising steps to maintain generation quality, motivating recent research on efficient decoding strategies. However, existing studies have reported inconsistent evaluation results even under seemingly identical evaluation settings, risking biased conclusions about dLLM decoding methods. To understand this evaluation concern, we conduct a rigorous evaluation of current decoding methods for dLLMs across diverse evaluation settings. Surprisingly, our analysis reveals that the ranking of decoding methods is highly sensitive to the choice of prompt templates. Single-template evaluation can lead to an illusion that decoding methods improve inference efficiency without performance degradation. Through comprehensive experiments, we find that current parallel decoding methods consistently underperform the single-token decoding baseline, failing to overcome the speed-quality trade-off. We further identify this evaluation inconsistency as the high sensitivity of parallel decoding methods to minor variations in prompt templates. Our experiments show that an effective prompt template can achieve strong evaluation results even with fewer denoising steps, markedly outperforming the marginal gain from increasing denoising steps. Beyond prompt templates, our experiments indicate that overlooked evaluation settings can also notably affect the assessment of decoding methods. Based on these findings, we propose practical guidelines for the reliable evaluation of decoding methods in dLLMs.