Selected work

Representative Papers

Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling

Jul 07, 2026

This work addresses the challenges of solving high-dimensional Hamilton-Jacobi-Bellman (HJB) equations within diffusion models, which typically suffer from prolonged training times and high sensitivity to hyperparameters. The paper introduces, for the first time, a functional tensor train (FTT) low-rank structure into this setting, integrating it with backward stochastic differential equations (BSDEs) and a reverse-time iterative algorithm. This combination enables efficient approximation of high-dimensional density functions and facilitates rapid score-based sampling. The proposed method substantially enhances sampling efficiency and stability for complex high-dimensional distributions, significantly reducing training time and diminishing reliance on careful hyperparameter tuning across multiple benchmarks.

0 citationsRead paper

Square-Root Price Impact Is Necessary for Endogenous Manipulation Cycles in Learning-Agent Markets

Jul 06, 2026

This study investigates the emergence of endogenous manipulation cycles in a learning-based trading agent market and their impact on price dynamics. By constructing a minimal agent-based model comprising one institutionally optimized evolutionary agent and twenty thousand retail traders, and leveraging mean-field theory to reduce it to a nonlinear oscillator system, the work reveals—for the first time—that a square-root price impact is a necessary condition for such endogenous cycles. Remarkably, self-sustained limit cycles arise solely from position feedback and square-root impact, even without herding by retail traders, functioning analogously to a Maxwell’s demon–like information controller. Over 2,000 trading days, experiments consistently reproduce 8–11 cycles, yielding an average cumulative return of 37.7% (peaking at 51%), thereby validating a Hopf bifurcation exponent α ≈ 1/2 and demonstrating that linear price impact cannot generate these cycles.

0 citationsRead paper

Order Splitting and Liquidity Replenishment Are Jointly Necessary for the Square-Root Law of Market Impact:

Jul 05, 2026

This study investigates the origin of the exponent δ ≈ 0.5 in the square-root law (SRL) of market impact and evaluates the validity of competing theoretical frameworks, including GGPS, FGLW, and limit order book (LOB) walking models. Leveraging a heterogeneous agent-based LOB model calibrated to empirical data and validated through large-scale simulations across 2,000 independent stocks, the authors employ controlled experiments and counterfactual ablation analyses to demonstrate that order splitting by informed traders and liquidity replenishment by market makers constitute two necessary and synergistic mechanisms underpinning the SRL. The model reproduces δ = 0.539 ± 0.048, closely matching the empirically observed value of 0.489 from the Tokyo Stock Exchange; removing either mechanism causes δ to deviate significantly from 0.5, whereas other perturbations induce deviations of less than 10%.

0 citationsRead paper

Semi-Supervised Sound Event Detection with Conditional Mixup and Embedding-Level Contrastive Loss

Jun 29, 2026

This work addresses the challenge of sound event detection under limited labeled data by proposing a semi-supervised fine-tuning framework that effectively leverages abundant unlabeled data. Building upon a pretrained audio foundation model, the approach integrates pseudo-labeling, a novel conditional mixing strategy that unifies mixup and perturbation-based augmentation, and embedding-level contrastive learning. The conditional mixing mechanism harmonizes the divergent data augmentation requirements of pseudo-label learning and contrastive learning. Evaluated on the DESED validation set, the method achieves state-of-the-art performance with PSDS1 and PSDS2 scores of 0.645 and 0.822, respectively, setting a new benchmark for sound event detection in low-resource settings.

0 citationsRead paper

LeanSearch v2: Global Premise Retrieval for Lean 4 Theorem Proving

May 13, 2026

This work addresses the challenge in Lean 4 theorem proving that existing tools struggle to retrieve the complete set of prerequisite lemmas needed for a full proof in a single pass. The paper introduces LeanSearch v2, the first dual-mode system supporting global premise retrieval: its standard mode leverages a hierarchical informal Mathlib corpus with an embedding-reranking pipeline for efficient one-shot retrieval, while its reasoning mode employs an iterative sketch-retrieve-reflect loop that enhances recall quality without domain-specific fine-tuning. Evaluated on 69 research-level theorems, the reasoning mode retrieves 46.1% of the true premise sets within just 10 candidates, substantially outperforming baselines. When integrated into a prover, it achieves a 20% proof success rate—significantly higher than both no retrieval (4%) and the next-best system (16%).

0 citationsRead paper

Recent publications

Latest Papers

Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling

Jul 07, 2026

This work addresses the challenges of solving high-dimensional Hamilton-Jacobi-Bellman (HJB) equations within diffusion models, which typically suffer from prolonged training times and high sensitivity to hyperparameters. The paper introduces, for the first time, a functional tensor train (FTT) low-rank structure into this setting, integrating it with backward stochastic differential equations (BSDEs) and a reverse-time iterative algorithm. This combination enables efficient approximation of high-dimensional density functions and facilitates rapid score-based sampling. The proposed method substantially enhances sampling efficiency and stability for complex high-dimensional distributions, significantly reducing training time and diminishing reliance on careful hyperparameter tuning across multiple benchmarks.

0 citationsRead paper

Square-Root Price Impact Is Necessary for Endogenous Manipulation Cycles in Learning-Agent Markets

Jul 06, 2026

This study investigates the emergence of endogenous manipulation cycles in a learning-based trading agent market and their impact on price dynamics. By constructing a minimal agent-based model comprising one institutionally optimized evolutionary agent and twenty thousand retail traders, and leveraging mean-field theory to reduce it to a nonlinear oscillator system, the work reveals—for the first time—that a square-root price impact is a necessary condition for such endogenous cycles. Remarkably, self-sustained limit cycles arise solely from position feedback and square-root impact, even without herding by retail traders, functioning analogously to a Maxwell’s demon–like information controller. Over 2,000 trading days, experiments consistently reproduce 8–11 cycles, yielding an average cumulative return of 37.7% (peaking at 51%), thereby validating a Hopf bifurcation exponent α ≈ 1/2 and demonstrating that linear price impact cannot generate these cycles.

0 citationsRead paper

Order Splitting and Liquidity Replenishment Are Jointly Necessary for the Square-Root Law of Market Impact:

Jul 05, 2026

This study investigates the origin of the exponent δ ≈ 0.5 in the square-root law (SRL) of market impact and evaluates the validity of competing theoretical frameworks, including GGPS, FGLW, and limit order book (LOB) walking models. Leveraging a heterogeneous agent-based LOB model calibrated to empirical data and validated through large-scale simulations across 2,000 independent stocks, the authors employ controlled experiments and counterfactual ablation analyses to demonstrate that order splitting by informed traders and liquidity replenishment by market makers constitute two necessary and synergistic mechanisms underpinning the SRL. The model reproduces δ = 0.539 ± 0.048, closely matching the empirically observed value of 0.489 from the Tokyo Stock Exchange; removing either mechanism causes δ to deviate significantly from 0.5, whereas other perturbations induce deviations of less than 10%.

0 citationsRead paper

Semi-Supervised Sound Event Detection with Conditional Mixup and Embedding-Level Contrastive Loss

Jun 29, 2026

This work addresses the challenge of sound event detection under limited labeled data by proposing a semi-supervised fine-tuning framework that effectively leverages abundant unlabeled data. Building upon a pretrained audio foundation model, the approach integrates pseudo-labeling, a novel conditional mixing strategy that unifies mixup and perturbation-based augmentation, and embedding-level contrastive learning. The conditional mixing mechanism harmonizes the divergent data augmentation requirements of pseudo-label learning and contrastive learning. Evaluated on the DESED validation set, the method achieves state-of-the-art performance with PSDS1 and PSDS2 scores of 0.645 and 0.822, respectively, setting a new benchmark for sound event detection in low-resource settings.

0 citationsRead paper

LeanSearch v2: Global Premise Retrieval for Lean 4 Theorem Proving

May 13, 2026

This work addresses the challenge in Lean 4 theorem proving that existing tools struggle to retrieve the complete set of prerequisite lemmas needed for a full proof in a single pass. The paper introduces LeanSearch v2, the first dual-mode system supporting global premise retrieval: its standard mode leverages a hierarchical informal Mathlib corpus with an embedding-reranking pipeline for efficient one-shot retrieval, while its reasoning mode employs an iterative sketch-retrieve-reflect loop that enhances recall quality without domain-specific fine-tuning. Evaluated on 69 research-level theorems, the reasoning mode retrieves 46.1% of the true premise sets within just 10 candidates, substantially outperforming baselines. When integrated into a prover, it achieves a 20% proof success rate—significantly higher than both no retrieval (4%) and the next-best system (16%).

0 citationsRead paper