Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models

Researchers introduce McDiffuSE, an MCTS-based framework that optimizes slot-filling order in Masked Diffusion Models to improve performance on mathematical and code reasoning tasks. The approach achieves 3.2% improvement over autoregressive baselines and up to 19.5% gains on specific benchmarks by strategically exploring generation orderings rather than following sequential patterns.

McDiffuSE addresses a fundamental limitation in plan-and-infill decoding for diffusion language models: the extreme sensitivity to slot-filling order. While Masked Diffusion Models show theoretical promise for complex reasoning tasks, their practical performance has been constrained by high variance depending on how slots are sequentially completed. This research frames slot selection as a decision-making problem amenable to Monte Carlo Tree Search, enabling the model to evaluate partial completions through look-ahead simulations before committing to generation paths.

The work builds on growing recognition that diffusion-based language generation requires careful orchestration beyond simple sequential completion. Previous approaches treated slot ordering as arbitrary, but McDiffuSE demonstrates that strategic planning through tree search substantially improves outcomes. The 8.0% improvement over baseline plan-and-infill methods and particularly strong gains on code-related benchmarks (19.5% on MBPP) suggest the technique addresses real bottlenecks in reasoning tasks.

From an AI development perspective, this research has implications for how generative models balance exploration versus exploitation during decoding. The finding that larger exploration constants outperform increased simulations reveals important insights about overcoming model confidence biases—a challenge relevant across multiple decoding strategies. For AI practitioners building reasoning systems, McDiffuSE offers a concrete method to improve output quality without architectural changes.

Looking ahead, this work validates MCTS as a viable planning mechanism for language generation beyond traditional tree search applications. The unexpected importance of non-sequential generation orders opens questions about optimal decoding strategies and suggests further refinement could yield additional gains. Broader adoption would depend on computational overhead relative to inference speed requirements.

• →McDiffuSE uses Monte Carlo Tree Search to optimize slot-filling order in diffusion models, improving performance 3.2-19.5% depending on task
• →Larger exploration constants prove more effective than increased simulations for discovering optimal generation orderings
• →Non-sequential slot completion generates better results than expected, contradicting assumptions about optimal decoding
• →Framework demonstrates MCTS effectiveness for planning in generative language models beyond traditional applications
• →Strong benchmark improvements on MBPP (19.5%) and MATH500 (4.9%) indicate particular value for code and mathematical reasoning