Selected work

Representative Papers

AlphaPADI: Formulaic Alpha Discovery via Pool-Aware Hierarchical Discrete Diffusion

Oct 04, 2026

This study addresses the limitations of formulaic alpha discovery, including the neglect of pool context, inadequate preservation of structural hierarchy, and the non-differentiability of pool-level rewards. To overcome these challenges, we propose AlphaPADI, a novel framework that introduces a pool-aware hierarchical discrete diffusion mechanism. By integrating syntax-constrained initialization, multi-scale structural reconstruction, and preference learning, AlphaPADI transcends the bottleneck of conventional item-wise generation, which fails to exploit complementary information within the pool, thereby enabling the efficient synthesis of complementary alpha pools. Empirical evaluations on Chinese and U.S. stock markets demonstrate that the proposed method significantly outperforms baseline approaches in both predictive performance and portfolio returns, validating the effectiveness of pool-aware generation for financial signal discovery.

0 citationsRead paper

Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?

Sep 27, 2026

This study investigates whether Kullback-Leibler (KL) divergence is necessary in online policy distillation for large language models. Through theoretical analysis, we demonstrate that KL divergence is dispensable and that the update directions of critical divergent tokens solely determine distillation efficacy. Accordingly, we propose Consensus Multi-Teacher Online Policy Distillation (C-MOPD), which employs an update-direction-based reward mechanism alongside a multi-teacher collaborative supervision architecture to effectively resolve capability conflicts among teachers. The primary contribution of this work lies in establishing the first KL-divergence-free consensus multi-teacher distillation paradigm. Extensive evaluations on mathematical and code generation benchmarks demonstrate that C-MOPD consistently outperforms conventional multi-teacher distillation approaches, validating both its theoretical soundness and practical superiority.

0 citationsRead paper

The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting

Feb 03, 2026

This study addresses a fundamental limitation in traditional financial forecasting, where supervised labels are assumed to strictly align with the prediction target, thereby constraining model generalization. The authors introduce the concept of the “label temporal paradox,” demonstrating that optimal supervisory signals need not coincide with the target horizon. To resolve this, they propose a bilevel optimization framework that dynamically balances signal-to-noise ratios to automatically discover the optimal intermediate temporal proxy label within a single training pass. This approach eliminates the need for multiple rounds of training while significantly enhancing predictive performance. Extensive experiments on large-scale financial datasets show that the method consistently outperforms existing baselines, underscoring the critical role of supervisory label design in improving model generalization.

0 citationsRead paper

Recent publications

Latest Papers

AlphaPADI: Formulaic Alpha Discovery via Pool-Aware Hierarchical Discrete Diffusion

Oct 04, 2026

This study addresses the limitations of formulaic alpha discovery, including the neglect of pool context, inadequate preservation of structural hierarchy, and the non-differentiability of pool-level rewards. To overcome these challenges, we propose AlphaPADI, a novel framework that introduces a pool-aware hierarchical discrete diffusion mechanism. By integrating syntax-constrained initialization, multi-scale structural reconstruction, and preference learning, AlphaPADI transcends the bottleneck of conventional item-wise generation, which fails to exploit complementary information within the pool, thereby enabling the efficient synthesis of complementary alpha pools. Empirical evaluations on Chinese and U.S. stock markets demonstrate that the proposed method significantly outperforms baseline approaches in both predictive performance and portfolio returns, validating the effectiveness of pool-aware generation for financial signal discovery.

0 citationsRead paper

Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?

Sep 27, 2026

This study investigates whether Kullback-Leibler (KL) divergence is necessary in online policy distillation for large language models. Through theoretical analysis, we demonstrate that KL divergence is dispensable and that the update directions of critical divergent tokens solely determine distillation efficacy. Accordingly, we propose Consensus Multi-Teacher Online Policy Distillation (C-MOPD), which employs an update-direction-based reward mechanism alongside a multi-teacher collaborative supervision architecture to effectively resolve capability conflicts among teachers. The primary contribution of this work lies in establishing the first KL-divergence-free consensus multi-teacher distillation paradigm. Extensive evaluations on mathematical and code generation benchmarks demonstrate that C-MOPD consistently outperforms conventional multi-teacher distillation approaches, validating both its theoretical soundness and practical superiority.

0 citationsRead paper

The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting

Feb 03, 2026

This study addresses a fundamental limitation in traditional financial forecasting, where supervised labels are assumed to strictly align with the prediction target, thereby constraining model generalization. The authors introduce the concept of the “label temporal paradox,” demonstrating that optimal supervisory signals need not coincide with the target horizon. To resolve this, they propose a bilevel optimization framework that dynamically balances signal-to-noise ratios to automatically discover the optimal intermediate temporal proxy label within a single training pass. This approach eliminates the need for multiple rounds of training while significantly enhancing predictive performance. Extensive experiments on large-scale financial datasets show that the method consistently outperforms existing baselines, underscoring the critical role of supervisory label design in improving model generalization.

0 citationsRead paper