Selected work

Representative Papers

Efficient Neural Controlled Differential Equations via Attentive Kernel Smoothing

Feb 02, 2026

This work addresses the high computational cost of Neural Controlled Differential Equations (Neural CDEs), which arises from excessively small solver step sizes due to roughness in the control path. To mitigate this, the authors propose a novel path construction method based on kernel functions and Gaussian processes, replacing conventional spline interpolation to effectively suppress high-frequency noise while preserving essential temporal details. Furthermore, they introduce an attention-driven, multi-view CDE architecture that enables controllable modeling of trajectory smoothness and facilitates multi-scale dynamic feature fusion. The resulting model, termed MVC-CDE with GP, achieves state-of-the-art accuracy while significantly reducing the number of function evaluations and inference time, thereby offering an improved balance between model efficiency and performance.

1 citationsRead paper

EvoMem: Memory-Augmented Evolution for Code Optimization

Aug 11, 2026

This work addresses the inefficiency of existing large language model (LLM)-based evolutionary code search methods, which discard effective mutation strategies after each run, hindering knowledge reuse across tasks or iterations and leading to redundant exploration. To overcome this limitation, the paper introduces EvoMem—a persistent memory architecture that, for the first time, integrates a structured, task-aware external memory mechanism into LLM-driven evolutionary search. EvoMem captures successful strategies through knowledge extraction, stores them with source annotations, and enables context-aware retrieval, allowing dynamic reuse in subsequent tasks. Experimental results demonstrate that EvoMem significantly improves search efficiency or target performance across diverse domains, including geometric optimization, multi-hop question answering, and GPU kernel optimization, effectively enabling cross-iteration and cross-task transfer of successful strategies.

0 citationsRead paper

Diffract: Spectral View of LLM Domain Adaptation

Aug 11, 2026

This work investigates efficient adaptation of general-purpose large language models to specialized domains—such as mathematics, instruction following, and code—through continual pretraining, balancing performance and computational efficiency. Spectral analysis via singular value decomposition reveals that adaptation is primarily driven by the evolution of singular vectors, while the singular value spectrum remains remarkably invariant. Leveraging the heterogeneity among attention heads, the authors propose an importance criterion to selectively roll back low-importance heads, reducing up to 60% of updated parameters without compromising model quality and even improving benchmark accuracy by up to 4%. The study further uncovers linearly interpolable domain connectivity across distinct domains and releases Diffract, a large-scale spectral analysis toolkit.

0 citationsRead paper

Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

Aug 11, 2026

This study addresses the unresolved trade-off between batch size and the number of negative samples under fixed memory budgets in training memory-constrained recommender systems. It theoretically and empirically demonstrates, for the first time, that when using sampled Softmax, prioritizing larger batch sizes over a greater number of negative samples yields faster convergence and superior recommendation quality within the same memory constraints. The proposed configuration principle is validated across four real-world sequential recommendation benchmarks—including MovieLens-20M—as well as synthetic data, offering clear guidance for efficient training in resource-limited scenarios.

0 citationsRead paper

MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors

Jul 20, 2026

This work addresses the limitations of existing methods that establish image correspondences at sparse keypoints or pixel levels, which are insufficient for instance-level high-level tasks such as topological navigation. The authors propose cross-view segment matching at the instance segmentation level, leveraging 3D foundation models (e.g., MASt3R, VGGT) to extract mask-level features and introducing three matching heads: a LightGlue-style attention mechanism, multi-scale spatial fusion, and a core multi-view joint self-attention module. The latter recovers transitive correspondences, overcoming the pairwise matching bottleneck. Experiments demonstrate significant improvements: zero-shot AUPRC increases by 4.85 and 25.9 on Replica and Virtual KITTI 2, respectively; in HM3D navigation, multi-view matching raises success rates from 50% to 70%, and the LightGlue head boosts SPL from 45.7 to 59.1—marking the first effective integration of segment-level matching into downstream tasks.

0 citationsRead paper

Recent publications

Latest Papers

EvoMem: Memory-Augmented Evolution for Code Optimization

Aug 11, 2026

This work addresses the inefficiency of existing large language model (LLM)-based evolutionary code search methods, which discard effective mutation strategies after each run, hindering knowledge reuse across tasks or iterations and leading to redundant exploration. To overcome this limitation, the paper introduces EvoMem—a persistent memory architecture that, for the first time, integrates a structured, task-aware external memory mechanism into LLM-driven evolutionary search. EvoMem captures successful strategies through knowledge extraction, stores them with source annotations, and enables context-aware retrieval, allowing dynamic reuse in subsequent tasks. Experimental results demonstrate that EvoMem significantly improves search efficiency or target performance across diverse domains, including geometric optimization, multi-hop question answering, and GPU kernel optimization, effectively enabling cross-iteration and cross-task transfer of successful strategies.

0 citationsRead paper

Diffract: Spectral View of LLM Domain Adaptation

Aug 11, 2026

This work investigates efficient adaptation of general-purpose large language models to specialized domains—such as mathematics, instruction following, and code—through continual pretraining, balancing performance and computational efficiency. Spectral analysis via singular value decomposition reveals that adaptation is primarily driven by the evolution of singular vectors, while the singular value spectrum remains remarkably invariant. Leveraging the heterogeneity among attention heads, the authors propose an importance criterion to selectively roll back low-importance heads, reducing up to 60% of updated parameters without compromising model quality and even improving benchmark accuracy by up to 4%. The study further uncovers linearly interpolable domain connectivity across distinct domains and releases Diffract, a large-scale spectral analysis toolkit.

0 citationsRead paper

Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

Aug 11, 2026

This study addresses the unresolved trade-off between batch size and the number of negative samples under fixed memory budgets in training memory-constrained recommender systems. It theoretically and empirically demonstrates, for the first time, that when using sampled Softmax, prioritizing larger batch sizes over a greater number of negative samples yields faster convergence and superior recommendation quality within the same memory constraints. The proposed configuration principle is validated across four real-world sequential recommendation benchmarks—including MovieLens-20M—as well as synthetic data, offering clear guidance for efficient training in resource-limited scenarios.

0 citationsRead paper

MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors

Jul 20, 2026

This work addresses the limitations of existing methods that establish image correspondences at sparse keypoints or pixel levels, which are insufficient for instance-level high-level tasks such as topological navigation. The authors propose cross-view segment matching at the instance segmentation level, leveraging 3D foundation models (e.g., MASt3R, VGGT) to extract mask-level features and introducing three matching heads: a LightGlue-style attention mechanism, multi-scale spatial fusion, and a core multi-view joint self-attention module. The latter recovers transitive correspondences, overcoming the pairwise matching bottleneck. Experiments demonstrate significant improvements: zero-shot AUPRC increases by 4.85 and 25.9 on Replica and Virtual KITTI 2, respectively; in HM3D navigation, multi-view matching raises success rates from 50% to 70%, and the LightGlue head boosts SPL from 45.7 to 59.1—marking the first effective integration of segment-level matching into downstream tasks.

0 citationsRead paper

ASK-NN: An Asymmetric Nearest-Neighbor Test that detects Distribution Drifts in Natural Language

Jul 17, 2026

This work addresses hallucination and artificial text generation in large language models caused by distributional shifts between prompt and response hidden states, particularly in scenarios with asymmetric reference-query sequence lengths. To tackle this, the authors propose ASK-NN, an asymmetric two-sample test based on directed k-nearest neighbor graphs. The method detects distributional drift by statistically counting how often reference points retain other reference points as nearest neighbors within a mixed sample. ASK-NN is the first to introduce asymmetric nearest neighbor graphs for natural language distribution testing, offering exact conditional mean and variance under finite samples and establishing asymptotic normality and consistency under fixed alternative hypotheses. Experiments demonstrate that ASK-NN matches or outperforms kernel- and graph-based baselines on synthetic benchmarks, artificial text detection, and token-level hidden state hallucination tasks, while remaining computationally efficient and straightforward to implement.

0 citationsRead paper