Selected work

Representative Papers

NAE: Normalizing AutoEncoder

Aug 12, 2026

This work addresses a critical inconsistency between the existing loss functions used in flow-based autoencoders and their reconstruction objectives, which leads to suboptimal training dynamics. The paper provides the first theoretical analysis of this misalignment and introduces a novel conditional loss function designed to align the surrogate gradients of the encoder and decoder with the true reconstruction loss. By preserving the established architecture that combines normalizing flows with autoencoders, the proposed method significantly enhances generative performance. It achieves state-of-the-art results across diverse benchmarks—including molecular generation, tabular data modeling, and image synthesis—demonstrating both its effectiveness and broad applicability.

0 citationsRead paper

Two-stage Odd Residual Flows for Mean-Preserving Probabilistic Time Series Forecasting

Aug 11, 2026

Existing probabilistic time series forecasting methods struggle to simultaneously achieve distributional flexibility and accurate mean prediction. This work proposes a Two-stage Odd Residual Flow (TORF) framework that first leverages a pretrained deterministic model to generate high-accuracy point forecasts, then models the residual uncertainty using a constrained normalizing flow composed exclusively of odd functions centered at the predicted mean. TORF is the first approach to decouple mean estimation from uncertainty modeling while guaranteeing exact mean preservation without requiring sampling, thanks to the odd-function constraint. By jointly optimizing negative log-likelihood and Continuous Ranked Probability Score (CRPS), TORF achieves state-of-the-art performance in both deterministic accuracy (measured by NMAE) and probabilistic forecasting quality (measured by CRPS) across short- and long-horizon time series prediction tasks.

0 citationsRead paper

Do Tabular Foundation Models Agree with Themselves?

Aug 06, 2026

This work investigates whether state-of-the-art tabular foundation models (TFMs) adhere to fundamental mathematical properties of joint distributions when performing autoregressive multivariate forecasting. To this end, we introduce two consistency criteria—marginalization consistency and factorization consistency—and develop the first theoretical framework capable of evaluating the internal consistency of TFMs without access to ground-truth posteriors. Empirical evaluations of Transformer-based TFMs across multiple classification and regression datasets reveal that all tested models violate both criteria, producing predictions that cannot correspond to any valid joint distribution. These findings expose a fundamental flaw in current approaches to multivariate prediction with TFMs, highlighting a critical gap between their practical deployment and probabilistic coherence.

0 citationsRead paper

CAMP: A Cycle-Aware Multi-Scale Patch Mixer for Time Series Forecasting

Aug 04, 2026

Existing time series forecasting methods struggle with time-varying periodic patterns or coexisting multiple cycles and typically apply a uniform processing strategy to all patches, ignoring their relative importance with respect to the prediction horizon. To address these limitations, this work proposes CAMP, a novel framework that incorporates an adaptive periodicity learning module to identify dominant frequencies for each input window and generate corresponding historical and future periodic components. It further introduces a horizon-guided Patch Mixer for position-aware context fusion and jointly models de-seasonalized residual dynamics across multiple scales. CAMP is the first method to achieve input-window-level adaptive periodic modeling, position-aware patch refinement, and multi-resolution residual learning in a unified architecture. Experiments demonstrate that CAMP achieves the best average MSE and MAE on six out of seven long-term forecasting benchmarks and attains the highest MSE win rate across 16 settings on four PEMS traffic datasets.

0 citationsRead paper

Challenges in annotations by humans and LLMs: A case study of evaluative language

Jul 30, 2026

This study addresses the challenging and highly subjective task of annotating evaluative language by focusing on the Attitude subsystem within Appraisal Theory, specifically examining expressions of affect, judgment, and appreciation in TED Talks. Through a systematic comparison of annotation behaviors among trained linguists, novice linguists, and large language models (LLMs), it reveals for the first time the performance disparities between human expertise and LLMs in such nuanced tasks. By integrating three prompt engineering strategies with model fine-tuning, the research significantly enhances the LLM’s capacity for automatic classification of Appraisal categories, achieving an F1 score of 0.77 post-fine-tuning—surpassing novice linguists and approaching the performance of expert linguists. These findings demonstrate the promising potential of LLMs as effective assistants in complex theoretical annotation tasks within digital humanities.

0 citationsRead paper

Recent publications

Latest Papers

NAE: Normalizing AutoEncoder

Aug 12, 2026

This work addresses a critical inconsistency between the existing loss functions used in flow-based autoencoders and their reconstruction objectives, which leads to suboptimal training dynamics. The paper provides the first theoretical analysis of this misalignment and introduces a novel conditional loss function designed to align the surrogate gradients of the encoder and decoder with the true reconstruction loss. By preserving the established architecture that combines normalizing flows with autoencoders, the proposed method significantly enhances generative performance. It achieves state-of-the-art results across diverse benchmarks—including molecular generation, tabular data modeling, and image synthesis—demonstrating both its effectiveness and broad applicability.

0 citationsRead paper

Two-stage Odd Residual Flows for Mean-Preserving Probabilistic Time Series Forecasting

Aug 11, 2026

Existing probabilistic time series forecasting methods struggle to simultaneously achieve distributional flexibility and accurate mean prediction. This work proposes a Two-stage Odd Residual Flow (TORF) framework that first leverages a pretrained deterministic model to generate high-accuracy point forecasts, then models the residual uncertainty using a constrained normalizing flow composed exclusively of odd functions centered at the predicted mean. TORF is the first approach to decouple mean estimation from uncertainty modeling while guaranteeing exact mean preservation without requiring sampling, thanks to the odd-function constraint. By jointly optimizing negative log-likelihood and Continuous Ranked Probability Score (CRPS), TORF achieves state-of-the-art performance in both deterministic accuracy (measured by NMAE) and probabilistic forecasting quality (measured by CRPS) across short- and long-horizon time series prediction tasks.

0 citationsRead paper

Do Tabular Foundation Models Agree with Themselves?

Aug 06, 2026

This work investigates whether state-of-the-art tabular foundation models (TFMs) adhere to fundamental mathematical properties of joint distributions when performing autoregressive multivariate forecasting. To this end, we introduce two consistency criteria—marginalization consistency and factorization consistency—and develop the first theoretical framework capable of evaluating the internal consistency of TFMs without access to ground-truth posteriors. Empirical evaluations of Transformer-based TFMs across multiple classification and regression datasets reveal that all tested models violate both criteria, producing predictions that cannot correspond to any valid joint distribution. These findings expose a fundamental flaw in current approaches to multivariate prediction with TFMs, highlighting a critical gap between their practical deployment and probabilistic coherence.

0 citationsRead paper

CAMP: A Cycle-Aware Multi-Scale Patch Mixer for Time Series Forecasting

Aug 04, 2026

Existing time series forecasting methods struggle with time-varying periodic patterns or coexisting multiple cycles and typically apply a uniform processing strategy to all patches, ignoring their relative importance with respect to the prediction horizon. To address these limitations, this work proposes CAMP, a novel framework that incorporates an adaptive periodicity learning module to identify dominant frequencies for each input window and generate corresponding historical and future periodic components. It further introduces a horizon-guided Patch Mixer for position-aware context fusion and jointly models de-seasonalized residual dynamics across multiple scales. CAMP is the first method to achieve input-window-level adaptive periodic modeling, position-aware patch refinement, and multi-resolution residual learning in a unified architecture. Experiments demonstrate that CAMP achieves the best average MSE and MAE on six out of seven long-term forecasting benchmarks and attains the highest MSE win rate across 16 settings on four PEMS traffic datasets.

0 citationsRead paper

Challenges in annotations by humans and LLMs: A case study of evaluative language

Jul 30, 2026

This study addresses the challenging and highly subjective task of annotating evaluative language by focusing on the Attitude subsystem within Appraisal Theory, specifically examining expressions of affect, judgment, and appreciation in TED Talks. Through a systematic comparison of annotation behaviors among trained linguists, novice linguists, and large language models (LLMs), it reveals for the first time the performance disparities between human expertise and LLMs in such nuanced tasks. By integrating three prompt engineering strategies with model fine-tuning, the research significantly enhances the LLM’s capacity for automatic classification of Appraisal categories, achieving an F1 score of 0.77 post-fine-tuning—surpassing novice linguists and approaching the performance of expert linguists. These findings demonstrate the promising potential of LLMs as effective assistants in complex theoretical annotation tasks within digital humanities.

0 citationsRead paper