Planner Aware Path Learning in Diffusion Language Models Training
Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Michael M. Bronstein, Anru Zhang, Joey Bose, Alexander Tong
We propose Planner Aware Path Learning (PAPL), a simple planner-aligned training method for Diffusion Language Models that resolves the training–inference mismatch and consistently improves generation quality.
Abstract
Diffusion language models have emerged as a powerful alternative to autoregressive models, enabling fast inference through more flexible and parallel generation paths. This flexibility of sampling is unlocked by new engineered sampling strategies, or *planners*, that select more favorable generation paths by iteratively planning---versus uniformly at random---where to denoise along the sequence. However, by modifying the reverse paths via planning, planners create an irrevocable mismatch between the uniformly random denoising paths during training and planning-based inference. In this paper, we systematically investigate the mismatch of discrete diffusion training and inference under planning and theoretically prove that the standard discrete diffusion training evidence lower bound (ELBO) does not accurately describe a denoiser that uses a non-uniform planner. To address this gap, we derive a new planned evidence lower bound (P-ELBO) that incorporates planner-based reverse dynamics directly into the training objective. Using the P-ELBO, we introduce *Planner Aware Path Learning* (PAPL), a novel training scheme that aligns training and inference under a planned denoiser. PAPL is implemented as a simple yet effective modification to the standard masked discrete diffusion loss, making it widely applicable and easy to adopt. Empirically, we show PAPL delivers consistent gains across domains, including a 40\% relative improvement in protein sequences, improved text generation with up to a $4\times$ relative MAUVE gain, and 23\% relative improvement in code generation HumanEval pass@10.
Theoretical characterization shows MDMs are expressively equivalent to padded looped transformers, more efficient for parallel problems.
- Establishes equivalence between masked diffusion models and polynomially-padded looped transformers in finite-precision log-width setting
- Shows MDMs can solve all problems CoT-augmented transformers can with identification of expressivity gaps
- Demonstrates MDMs inherently more efficient than CoT transformers on highly-parallelizable problems like regular languages
- Theoretical complexity analysis
- Chain of thought reasoning
- Transformer expressivity theory
- Circuit complexity
Authors did not state explicit limitations.
Investigate hybrid models combining autoregressive generation with non-autoregressive infilling within blocks
from the paper
Author keywords
- Diffusion Language Models
- Discrete Diffusion
- Diffusion Models
- code generation
- protein generation
- text generation
Related orals
Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)
RealUID provides universal distillation for matching models without GANs, incorporating real data into one-step generator training.
GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion Models
GLASS Flows samples Markov transitions via inner flow matching models to improve inference-time reward alignment in flow and diffusion models.
Neon: Negative Extrapolation From Self-Training Improves Image Generation
Neon inverts model degradation from self-training by extrapolating away from it, improving generative models with minimal compute.
Generative Human Geometry Distribution
Introduces distribution-over-distribution model combining geometry distributions with two-stage flow matching for human 3D generation.
Cross-Domain Lossy Compression via Rate- and Classification-Constrained Optimal Transport
Cross-domain lossy compression unifies rate and classification constraints via optimal transport framework.