Selected work

Representative Papers

Topic-to-Timestamp Alignment by Constrained Evidence Selection

Jun 18, 2026

This work addresses the challenge of users struggling to pinpoint specific moments in meeting discussions based solely on content. To overcome this, the paper proposes a novel approach that reframes timestamp prediction as a constrained candidate selection task. Instead of directly generating timestamps, large language models such as Mistral-7B-Instruct are guided to select the most relevant segment from a set of retrieved, timestamped meeting excerpts, thereby avoiding unsupported or invalid predictions. Integrating retrieval-augmented generation (RAG) with a constrained selection mechanism, the method demonstrates significant improvements on a dataset of 200 municipal meetings and 420 queries: Recall@5 increases from 31.9% to 50.0%, mean absolute error decreases to 761 seconds, and the number of valid outputs rises from 373 to 419, substantially enhancing both accuracy and reliability in temporal localization.

0 citationsRead paper

Doubly Stochastic Mean-Shift Clustering

Feb 17, 2026

This work addresses the sensitivity of the traditional Mean-Shift algorithm to the bandwidth hyperparameter, which often leads to over-segmentation and spurious clusters in sparse data regimes. To mitigate this limitation, the authors propose Dual Stochastic Mean-Shift (DSMS), a novel variant that, for the first time, models the bandwidth as a random variable. In each iteration, DSMS jointly samples both data points and kernel radii stochastically, thereby introducing an implicit regularization mechanism that enhances exploration of the underlying density landscape. Theoretical analysis confirms that the proposed method preserves convergence guarantees. Experimental results on synthetic Gaussian mixture datasets demonstrate that DSMS significantly outperforms both standard Mean-Shift and existing stochastic variants, effectively suppressing over-segmentation without sacrificing clustering performance.

0 citationsRead paper

Stochastic Mean-Shift Clustering

Nov 12, 2025

To address the slow convergence and high computational cost of traditional mean shift clustering, this paper proposes Randomized Mean Shift (RMS): a stochastic variant that iteratively samples data point sequences and performs partial gradient ascent updates on the Gaussian kernel density estimate objective—bypassing full-batch iterations. By incorporating a stochastic gradient ascent mechanism, RMS significantly improves convergence speed and clustering accuracy while preserving mode-seeking capability. On synthetic 2D data generated from Gaussian mixtures, RMS achieves an average 2.3× speedup and a 5.7% gain in clustering accuracy over standard mean shift. Applied to speaker embedding clustering on VoxCeleb1, it attains an 89.4% diarization error rate (DER), outperforming mainstream unsupervised baselines. The core contribution is the first systematic integration of stochastic optimization principles into the mean shift framework, achieving a favorable trade-off among efficiency, scalability, and empirical performance.

0 citationsRead paper

Recent publications

Latest Papers

Topic-to-Timestamp Alignment by Constrained Evidence Selection

Jun 18, 2026

This work addresses the challenge of users struggling to pinpoint specific moments in meeting discussions based solely on content. To overcome this, the paper proposes a novel approach that reframes timestamp prediction as a constrained candidate selection task. Instead of directly generating timestamps, large language models such as Mistral-7B-Instruct are guided to select the most relevant segment from a set of retrieved, timestamped meeting excerpts, thereby avoiding unsupported or invalid predictions. Integrating retrieval-augmented generation (RAG) with a constrained selection mechanism, the method demonstrates significant improvements on a dataset of 200 municipal meetings and 420 queries: Recall@5 increases from 31.9% to 50.0%, mean absolute error decreases to 761 seconds, and the number of valid outputs rises from 373 to 419, substantially enhancing both accuracy and reliability in temporal localization.

0 citationsRead paper

Doubly Stochastic Mean-Shift Clustering

Feb 17, 2026

This work addresses the sensitivity of the traditional Mean-Shift algorithm to the bandwidth hyperparameter, which often leads to over-segmentation and spurious clusters in sparse data regimes. To mitigate this limitation, the authors propose Dual Stochastic Mean-Shift (DSMS), a novel variant that, for the first time, models the bandwidth as a random variable. In each iteration, DSMS jointly samples both data points and kernel radii stochastically, thereby introducing an implicit regularization mechanism that enhances exploration of the underlying density landscape. Theoretical analysis confirms that the proposed method preserves convergence guarantees. Experimental results on synthetic Gaussian mixture datasets demonstrate that DSMS significantly outperforms both standard Mean-Shift and existing stochastic variants, effectively suppressing over-segmentation without sacrificing clustering performance.

0 citationsRead paper

Stochastic Mean-Shift Clustering

Nov 12, 2025

To address the slow convergence and high computational cost of traditional mean shift clustering, this paper proposes Randomized Mean Shift (RMS): a stochastic variant that iteratively samples data point sequences and performs partial gradient ascent updates on the Gaussian kernel density estimate objective—bypassing full-batch iterations. By incorporating a stochastic gradient ascent mechanism, RMS significantly improves convergence speed and clustering accuracy while preserving mode-seeking capability. On synthetic 2D data generated from Gaussian mixtures, RMS achieves an average 2.3× speedup and a 5.7% gain in clustering accuracy over standard mean shift. Applied to speaker embedding clustering on VoxCeleb1, it attains an 89.4% diarization error rate (DER), outperforming mainstream unsupervised baselines. The core contribution is the first systematic integration of stochastic optimization principles into the mean shift framework, achieving a favorable trade-off among efficiency, scalability, and empirical performance.

0 citationsRead paper