analiza aiqbench
Plan-Verify-Fill (PVF) introduces a training-free decoding strategy for Diffusion Language Models that leverages quantitative validation to construct hierarchical text skeletons. This approach significantly reduces the Number of Function Evaluations by up to 65% compared to confidence-based methods while maintaining accuracy, offering a more efficient alternative to standard autoregressive paradigms. The technique effectively addresses the reactive limitations of current decoding strategies by actively utilizing global bidirectional context to optimize generation trajectories.
Plan, Verify and Fill: A Structured Parallel Decoding Approach for Diffusion Language Models
arXiv:2601.12247v3 Announce Type: replace-cross Abstract: Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
Abstract:Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2601.12247 [cs.CL]
(or arXiv:2601.12247v3 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 131 cuvinte
Subiecte conexe
Structured-Diffuser: Diffusion with Task-Conditioned Structured Priors for Motion Planning
arXiv · 28 sept. 2026
Adaptive Item-based Collaborative Structures via Noise Rescheduling in Diffusion for Generative Recommendation
arXiv · 25 aug. 2026
ReDiSC: A Reparameterized Masked Diffusion Model for Scalable Node Classification with Structured Predictions
arXiv · 30 iul. 2026
From Uniform to Learned Graph Priors: Diffusion for Structure Discovery
arXiv · 12 iun. 2026