Selected work

Representative Papers

Diffusion Transformers are Provably Optimal In-context Generators

Oct 04, 2026

This study addresses the issue that task uncertainty under few-shot demonstrations can readily induce distributional bias in generative modeling. To mitigate this, the authors employ a Diffusion Transformer (DiT) architecture that integrates score estimation with an attention-based aggregation mechanism for diffusion sampling. Theoretically, they demonstrate that DiT is capable of capturing residual task uncertainty and learning a predictive distribution that faithfully reflects such uncertainty, rather than merely estimating a single task output. As a key contribution, this work elucidates the intrinsic mechanisms by which DiT handles task uncertainty and establishes minimax optimal convergence rates on test tasks within Hölder classes. These findings provide a rigorous theoretical foundation for few-shot generative modeling.

0 citationsRead paper

LTLDiff: Finite Linear Temporal Logic-Guided Data Generation and Diffusion Policies for Multi-agent Robotic Manipulation

Sep 09, 2026

针对多机器人协作操作中的同步、动作顺序和协调问题，提出LTLDiff框架，通过有限线性时序逻辑指导数据生成和扩散策略学习，提高任务成功率。

0 citationsRead paper

Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models

Aug 12, 2026

This work addresses the inefficiency of existing parallel decoding strategies in diffusion language models, which overlook the potential of early deterministic decisions to enhance global decoding efficiency. The authors propose a training-free active parallel decoding method that, for the first time, identifies and leverages a “ripple effect” during decoding: by detecting medium-entropy “pivot” positions, prospectively evaluating their impact on downstream uncertainty, and dynamically scheduling optimal decoding paths using KV cache management. Evaluated across three diffusion language models and four benchmarks spanning reasoning and code generation, the approach achieves 4–10× end-to-end speedup (up to 18× in peak cases) while preserving generation quality and consistently outperforming prior state-of-the-art baselines by up to 5.49% in accuracy across most settings.

0 citationsRead paper

Recent publications

Latest Papers

Diffusion Transformers are Provably Optimal In-context Generators

Oct 04, 2026

This study addresses the issue that task uncertainty under few-shot demonstrations can readily induce distributional bias in generative modeling. To mitigate this, the authors employ a Diffusion Transformer (DiT) architecture that integrates score estimation with an attention-based aggregation mechanism for diffusion sampling. Theoretically, they demonstrate that DiT is capable of capturing residual task uncertainty and learning a predictive distribution that faithfully reflects such uncertainty, rather than merely estimating a single task output. As a key contribution, this work elucidates the intrinsic mechanisms by which DiT handles task uncertainty and establishes minimax optimal convergence rates on test tasks within Hölder classes. These findings provide a rigorous theoretical foundation for few-shot generative modeling.

0 citationsRead paper

LTLDiff: Finite Linear Temporal Logic-Guided Data Generation and Diffusion Policies for Multi-agent Robotic Manipulation

Sep 09, 2026

针对多机器人协作操作中的同步、动作顺序和协调问题，提出LTLDiff框架，通过有限线性时序逻辑指导数据生成和扩散策略学习，提高任务成功率。

0 citationsRead paper

Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models

Aug 12, 2026

This work addresses the inefficiency of existing parallel decoding strategies in diffusion language models, which overlook the potential of early deterministic decisions to enhance global decoding efficiency. The authors propose a training-free active parallel decoding method that, for the first time, identifies and leverages a “ripple effect” during decoding: by detecting medium-entropy “pivot” positions, prospectively evaluating their impact on downstream uncertainty, and dynamically scheduling optimal decoding paths using KV cache management. Evaluated across three diffusion language models and four benchmarks spanning reasoning and code generation, the approach achieves 4–10× end-to-end speedup (up to 18× in peak cases) while preserving generation quality and consistently outperforming prior state-of-the-art baselines by up to 5.49% in accuracy across most settings.

0 citationsRead paper