The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models

TL;DR

This paper challenges the assumption that arbitrary order generation in diffusion language models enhances reasoning, showing it often narrows reasoning capabilities and proposing a simpler, more effective training approach called JustGRPO.

Contribution

The paper reveals that arbitrary order generation in dLLMs can limit reasoning and introduces JustGRPO, a straightforward training method that improves reasoning without sacrificing parallel decoding.

Findings

Arbitrary order generation tends to bypass high-uncertainty tokens, reducing reasoning potential.

JustGRPO achieves 89.1% accuracy on GSM8K, outperforming some existing methods.

Simpler training with JustGRPO is more effective than complex RL approaches for reasoning in dLLMs.

Abstract

Diffusion Large Language Models (dLLMs) break the rigid left-to-right constraint of traditional LLMs, enabling token generation in arbitrary orders. Intuitively, this flexibility implies a solution space that strictly supersets the fixed autoregressive trajectory, theoretically unlocking superior reasoning potential for general tasks like mathematics and coding. Consequently, numerous works have leveraged reinforcement learning (RL) to elicit the reasoning capability of dLLMs. In this paper, we reveal a counter-intuitive reality: arbitrary order generation, in its current form, narrows rather than expands the reasoning boundary of dLLMs. We find that dLLMs tend to exploit this order flexibility to bypass high-uncertainty tokens that are crucial for exploration, leading to a premature collapse of the solution space. This observation motivates a rethink of RL approaches for dLLMs, where considerable complexities, such as handling combinatorial trajectories and intractable likelihoods, are often devoted to preserving this flexibility. We demonstrate that effective reasoning can be better elicited by intentionally forgoing arbitrary order and applying standard Group Relative Policy Optimization (GRPO) instead. Our approach, JustGRPO, is minimalist yet surprisingly effective (e.g., 89.1% accuracy on GSM8K) while fully retaining the parallel decoding ability of dLLMs. Project page: https://nzl-thu.github.io/the-flexibility-trap