Selected work

Representative Papers

Synchronizing Probabilities in Model-Driven Lossless Compression

Jan 15, 2026

This work addresses a critical yet previously unformalized issue in model-driven lossless compression: decoding failures caused by inconsistencies between the probability predictions of the encoder and decoder. To resolve this prediction mismatch problem, the authors propose PMATIC, a model-agnostic fault-tolerant encoding algorithm that introduces a bounded prediction bias tolerance mechanism. This approach maintains theoretical correctness while substantially reducing computational and compression overhead. PMATIC is designed as a drop-in replacement for conventional arithmetic coders and is compatible with any probabilistic prediction model, including deep neural networks. Empirical evaluations on textual data demonstrate that PMATIC achieves superior compression ratios compared to state-of-the-art tools, even in the presence of non-negligible prediction discrepancies.

1 citationsRead paper

VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment

Jul 28, 2026

This work addresses the demand for high-accuracy, low-latency voice activity detection (VAD) on resource-constrained edge devices by proposing kiloVAD, a lightweight causal model that relies solely on standard Mel-spectrogram features and a pure convolutional neural network (CNN) architecture. By integrating layer-wise structured pruning, self-distillation, and angle-based quantization-aware training (AQAT), the method significantly enhances post-compression performance without resorting to non-standard or hardware-specific components. The resulting model contains only 2.1k parameters, operates with a 200ms context window, and achieves an AUC of 0.850 on the AVA-Speech benchmark, establishing a new state of the art for deployable causal VAD systems.

0 citationsRead paper

Industrial Dexterity Benchmark: A Hardware-Software Benchmarking Platform for Industrial Dexterous Manipulation

Jul 15, 2026

Industrial dexterous manipulation tasks—such as cable routing and connector insertion—remain heavily reliant on human labor due to their automation challenges. This work proposes an end-to-end multimodal imitation learning framework featuring three key contributions: the establishment of a benchmark platform for industrial dexterous manipulation (IDB), the design of a scalable DAG-ROS-based imitation learning architecture, and the development of a diffusion policy, AG-iDP3, which fuses multi-view RGB images, point clouds, joint states, and wrist torque measurements. Leveraging the R3M encoder, the system enables end-to-end training and achieves a combined success rate of 78% in grasping and inserting cables in a data center setting using only approximately 100 human demonstrations—substantially outperforming a single-camera baseline that attains just 36% success.

0 citationsRead paper

A Kleene theorem for free many-sorted algebras

Jun 29, 2026

This work extends the classical Kleene theorem to a multi-sorted setting of free algebras, resolving the equivalence between recognizable and regular languages in this generalized framework. Under suitable finiteness assumptions, the study integrates multi-sorted algebraic structures, formal language theory, and automata-theoretic techniques to establish, for the first time, a necessary and sufficient condition that a language over a multi-sorted free algebra is recognizable if and only if it is regular. This result unifies and generalizes existing theories of language recognizability across diverse multi-sorted structures, thereby providing a broader foundation for algebraic automata theory.

0 citationsRead paper

Forgotten Words: Benchmarking NeoBERT for Dementia Detection in Low-Resource Conversational Filipino and English Speech

May 25, 2026

This study addresses the limitations of existing neurolinguistic dementia detection approaches, which are predominantly English-centric and ill-suited for low-resource multilingual clinical settings involving Tagalog–English code-switching. The authors construct the first Tagalog–English parallel dementia dialogue dataset comprising 4,000 human-translated utterances and systematically evaluate models—including TF-IDF with logistic regression, BERT, NeoBERT, XLM-R, and RoBERTa-Tagalog—across monolingual, zero-shot cross-lingual, and bilingual fine-tuning setups. Notably, NeoBERT is introduced for the first time in clinical NLP, and the work establishes the first systematic benchmark for dementia detection in Tagalog. Experimental results demonstrate that bilingual fine-tuning effectively mitigates cross-lingual performance degradation, enabling all Transformer-based models to achieve Macro-F1 scores of 0.969–0.973 on Tagalog, thereby underscoring that linguistic coverage outweighs model size or architecture in determining performance in multilingual clinical NLP.

0 citationsRead paper

Recent publications

Latest Papers

VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment

Jul 28, 2026

This work addresses the demand for high-accuracy, low-latency voice activity detection (VAD) on resource-constrained edge devices by proposing kiloVAD, a lightweight causal model that relies solely on standard Mel-spectrogram features and a pure convolutional neural network (CNN) architecture. By integrating layer-wise structured pruning, self-distillation, and angle-based quantization-aware training (AQAT), the method significantly enhances post-compression performance without resorting to non-standard or hardware-specific components. The resulting model contains only 2.1k parameters, operates with a 200ms context window, and achieves an AUC of 0.850 on the AVA-Speech benchmark, establishing a new state of the art for deployable causal VAD systems.

0 citationsRead paper

Industrial Dexterity Benchmark: A Hardware-Software Benchmarking Platform for Industrial Dexterous Manipulation

Jul 15, 2026

Industrial dexterous manipulation tasks—such as cable routing and connector insertion—remain heavily reliant on human labor due to their automation challenges. This work proposes an end-to-end multimodal imitation learning framework featuring three key contributions: the establishment of a benchmark platform for industrial dexterous manipulation (IDB), the design of a scalable DAG-ROS-based imitation learning architecture, and the development of a diffusion policy, AG-iDP3, which fuses multi-view RGB images, point clouds, joint states, and wrist torque measurements. Leveraging the R3M encoder, the system enables end-to-end training and achieves a combined success rate of 78% in grasping and inserting cables in a data center setting using only approximately 100 human demonstrations—substantially outperforming a single-camera baseline that attains just 36% success.

0 citationsRead paper

A Kleene theorem for free many-sorted algebras

Jun 29, 2026

This work extends the classical Kleene theorem to a multi-sorted setting of free algebras, resolving the equivalence between recognizable and regular languages in this generalized framework. Under suitable finiteness assumptions, the study integrates multi-sorted algebraic structures, formal language theory, and automata-theoretic techniques to establish, for the first time, a necessary and sufficient condition that a language over a multi-sorted free algebra is recognizable if and only if it is regular. This result unifies and generalizes existing theories of language recognizability across diverse multi-sorted structures, thereby providing a broader foundation for algebraic automata theory.

0 citationsRead paper

Forgotten Words: Benchmarking NeoBERT for Dementia Detection in Low-Resource Conversational Filipino and English Speech

May 25, 2026

This study addresses the limitations of existing neurolinguistic dementia detection approaches, which are predominantly English-centric and ill-suited for low-resource multilingual clinical settings involving Tagalog–English code-switching. The authors construct the first Tagalog–English parallel dementia dialogue dataset comprising 4,000 human-translated utterances and systematically evaluate models—including TF-IDF with logistic regression, BERT, NeoBERT, XLM-R, and RoBERTa-Tagalog—across monolingual, zero-shot cross-lingual, and bilingual fine-tuning setups. Notably, NeoBERT is introduced for the first time in clinical NLP, and the work establishes the first systematic benchmark for dementia detection in Tagalog. Experimental results demonstrate that bilingual fine-tuning effectively mitigates cross-lingual performance degradation, enabling all Transformer-based models to achieve Macro-F1 scores of 0.969–0.973 on Tagalog, thereby underscoring that linguistic coverage outweighs model size or architecture in determining performance in multilingual clinical NLP.

0 citationsRead paper

Symmetry-Aware Fusion of Vision and Tactile Sensing via Bilateral Force Priors for Robotic Manipulation

Feb 14, 2026

This work addresses the challenge that vision alone struggles to capture fine-grained contact interactions in robotic insertion tasks, and existing vision–tactile fusion methods often exhibit unstable performance. To this end, the authors propose a Cross-Modal Transformer (CMT) that effectively integrates wrist-mounted visual and tactile signals through structured self-attention and cross-attention mechanisms. Notably, they introduce, for the first time, a physics-based bilateral force equilibrium prior as a regularization term to stabilize tactile embeddings and enhance symmetry perception. Evaluated on the TacSL benchmark, the proposed method achieves a 96.59% insertion success rate, substantially outperforming existing baselines and approaching the performance of an oracle system with access to ideal contact force information (96.09%), thereby demonstrating the efficacy of integrating multimodal perception with physical priors.

0 citationsRead paper