Selected work

Representative Papers

Hacking Generative Perplexity: Why Unconditional Text Evaluation Needs Distributional Metrics

Jun 06, 2026

Current non-autoregressive language models commonly rely on generation perplexity (gen-PPL) to evaluate text quality; however, this metric fails to adequately capture grammatical correctness and semantic coherence. This work proposes a zero-parameter naive sampler that achieves state-of-the-art gen-PPL on LM1B and OpenWebText yet produces clearly incoherent text, thereby systematically exposing the fundamental limitations of gen-PPL for the first time. To address this issue, we introduce direct evaluation methods based on distributional divergences—such as KL and Jensen–Shannon divergence—and scoring from pretrained autoregressive models. Our experiments demonstrate that these distribution-based metrics provide a more faithful and effective assessment of generation quality in unconditional text generation, establishing their necessity and superiority over conventional gen-PPL.

0 citationsRead paper

A Reversible Solver for Diffusion SDEs

Feb 12, 2025

This work addresses the problem of achieving exact, algebraically invertible encoding of real data samples into the prior distribution within diffusion models. We propose the first algebraically rigorous invertible stochastic differential equation (SDE) solver. Methodologically, our approach integrates continuous adjoint theory, Lie algebraic structure modeling, and invertible numerical integration; it combines explicit SDE discretization with Jacobian regularization to guarantee strict algebraic invertibility between forward and reverse trajectories. Unlike conventional irreversible numerical solvers, our method overcomes the fundamental distortion bottleneck in inverse mapping. Experiments on image and audio data demonstrate zero-reconstruction-error bidirectional exact mapping. Consequently, guided generation and controllable editing exhibit significantly improved fidelity and consistency.

0 citationsRead paper

Recent publications

Latest Papers

Hacking Generative Perplexity: Why Unconditional Text Evaluation Needs Distributional Metrics

Jun 06, 2026

Current non-autoregressive language models commonly rely on generation perplexity (gen-PPL) to evaluate text quality; however, this metric fails to adequately capture grammatical correctness and semantic coherence. This work proposes a zero-parameter naive sampler that achieves state-of-the-art gen-PPL on LM1B and OpenWebText yet produces clearly incoherent text, thereby systematically exposing the fundamental limitations of gen-PPL for the first time. To address this issue, we introduce direct evaluation methods based on distributional divergences—such as KL and Jensen–Shannon divergence—and scoring from pretrained autoregressive models. Our experiments demonstrate that these distribution-based metrics provide a more faithful and effective assessment of generation quality in unconditional text generation, establishing their necessity and superiority over conventional gen-PPL.

0 citationsRead paper

A Reversible Solver for Diffusion SDEs

Feb 12, 2025

This work addresses the problem of achieving exact, algebraically invertible encoding of real data samples into the prior distribution within diffusion models. We propose the first algebraically rigorous invertible stochastic differential equation (SDE) solver. Methodologically, our approach integrates continuous adjoint theory, Lie algebraic structure modeling, and invertible numerical integration; it combines explicit SDE discretization with Jacobian regularization to guarantee strict algebraic invertibility between forward and reverse trajectories. Unlike conventional irreversible numerical solvers, our method overcomes the fundamental distortion bottleneck in inverse mapping. Experiments on image and audio data demonstrate zero-reconstruction-error bidirectional exact mapping. Consequently, guided generation and controllable editing exhibit significantly improved fidelity and consistency.

0 citationsRead paper