Publications - PhusRoyal

2026

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

Phu Gia Hoang*, Anwoy Chatterjee*, Tanmoy Chakraborty, Iryna Gurevych, Subhabrata Dutta (* equal contribution)

Submitted to Journal of Machine Learning Research (JMLR) 2026

Can an interpretable SAE feature also be used as a stable steering direction? Our new preprint shows why the answer is often no.

[Project] [Paper] [Code]

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Feature 33539 · TopK Directed ray One shared removal direction

Feature 50431 · Matryoshka 2D directional subspace One shared, meaningfully used plane

Feature 19224 · ReLU Low-D residual Low-dimensional variation around the mean

Feature 43136 · TopK Multi-mode directional Several coherent directional modes

Feature 59154 · ReLU 1D diffuse evidence One-dimensional span without a ray or axis

Feature 13617 · Matryoshka Axis / antipodal One unsigned axis with both signs

Feature 32542 · ReLU kD directional subspace One shared low-dimensional span

Feature 58808 · TopK Unresolved high-D / diffuse High-dimensional, diffuse, or long-tail evidence

Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

Phu Gia Hoang*, Anwoy Chatterjee*, Tanmoy Chakraborty, Iryna Gurevych, Subhabrata Dutta (* equal contribution)

Submitted to Journal of Machine Learning Research (JMLR) 2026

Can an interpretable SAE feature also be used as a stable steering direction? Our new preprint shows why the answer is often no.

[Project] [Paper] [Code]

2024

NAACL

VLUE: A New Benchmark and Multi-task Knowledge Transfer Learning for Vietnamese Natural Language Understanding

VLUE: A New Benchmark and Multi-task Knowledge Transfer Learning for Vietnamese Natural Language Understanding

Phong Nguyen-Thuan Do, Son Quoc Tran, Phu Gia Hoang, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen

Findings of North American Chapter of the Association for Computational Linguistics (NAACL) 2024

We introduce VLUE, the first standardized Vietnamese NLU benchmark spanning five tasks and evaluate seven SOTA models on it. Our model CafeBERT, a Vietnamese-adapted version of XLM-R, sets a new state-of-the-art across all VLUE tasks and is released for public research use.

[Paper] [Code] 18 citations

VLUE: A New Benchmark and Multi-task Knowledge Transfer Learning for Vietnamese Natural Language Understanding

Phong Nguyen-Thuan Do, Son Quoc Tran, Phu Gia Hoang, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen

Findings of North American Chapter of the Association for Computational Linguistics (NAACL) 2024

We introduce VLUE, the first standardized Vietnamese NLU benchmark spanning five tasks and evaluate seven SOTA models on it. Our model CafeBERT, a Vietnamese-adapted version of XLM-R, sets a new state-of-the-art across all VLUE tasks and is released for public research use.

[Paper] [Code] 18 citations

2023

EACL

ViHOS: Hate Speech Spans Detection for Vietnamese

ViHOS: Hate Speech Spans Detection for Vietnamese

Phu Gia Hoang, Canh Duc Luu, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen

European Chapter of the Association for Computational Linguistics (EACL) 2023

Proposed a novel task for Vietnamese NLP, ViHOS - Vietnamese Hate and Offensive Spans, which consists of 12k online comments with advanced annotations on what make a sentence offensive and benchmark experiments.

[Paper] [Code] 43 citations

ViHOS: Hate Speech Spans Detection for Vietnamese

Phu Gia Hoang, Canh Duc Luu, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen

European Chapter of the Association for Computational Linguistics (EACL) 2023

Proposed a novel task for Vietnamese NLP, ViHOS - Vietnamese Hate and Offensive Spans, which consists of 12k online comments with advanced annotations on what make a sentence offensive and benchmark experiments.

[Paper] [Code] 43 citations

SemEval

VBD_NLP at SemEval-2023 Task 2: Named Entity Recognition Systems Enhanced by BabelNet and Wikipediae

VBD_NLP at SemEval-2023 Task 2: Named Entity Recognition Systems Enhanced by BabelNet and Wikipediae

Phu Gia Hoang, Le Thanh, Hai-Long Trieu

International Workshop on Semantic Evaluation (SemEval) 2023

We enhanced fine-grained NER performance for English and Bangla in SemEval-2023 by augmenting limited training data using BabelNet and Wikipedia redirections, and training mDeBERTa models on this enriched data. Our approach significantly outperformed baseline systems, achieving notable macro-F1 improvements in both languages.

[Paper] 1 citation

VBD_NLP at SemEval-2023 Task 2: Named Entity Recognition Systems Enhanced by BabelNet and Wikipediae

Phu Gia Hoang, Le Thanh, Hai-Long Trieu

International Workshop on Semantic Evaluation (SemEval) 2023

We enhanced fine-grained NER performance for English and Bangla in SemEval-2023 by augmenting limited training data using BabelNet and Wikipedia redirections, and training mDeBERTa models on this enriched data. Our approach significantly outperformed baseline systems, achieving notable macro-F1 improvements in both languages.

[Paper] 1 citation

NCA

Vietnamese hate and offensive detection using PhoBERT-CNN and social media streaming data

Vietnamese hate and offensive detection using PhoBERT-CNN and social media streaming data

Khanh Quoc Tran, An Trong Nguyen, Phu Gia Hoang, Canh Duc Luu, Trong-Hop Do, Kiet Van Nguyen

Neural Computing and Applications Journal 2023

An end-to-end Vietnamese hate speech detection system combining PhoBERT with Text-CNN, enhanced by tailored preprocessing and data augmentation. Our model achieves state-of-the-art results on benchmark datasets and is deployed in a real-time streaming application, demonstrating both high performance and practical utility.

[Paper] [Code] 45 citations

Vietnamese hate and offensive detection using PhoBERT-CNN and social media streaming data

Khanh Quoc Tran, An Trong Nguyen, Phu Gia Hoang, Canh Duc Luu, Trong-Hop Do, Kiet Van Nguyen

Neural Computing and Applications Journal 2023

An end-to-end Vietnamese hate speech detection system combining PhoBERT with Text-CNN, enhanced by tailored preprocessing and data augmentation. Our model achieves state-of-the-art results on benchmark datasets and is deployed in a real-time streaming application, demonstrating both high performance and practical utility.

[Paper] [Code] 45 citations

2021

SemEval

UIT-E10dot3 at SemEval-2021 Task 5: Toxic Spans Detection with Named Entity Recognition and Question-Answering Approaches

UIT-E10dot3 at SemEval-2021 Task 5: Toxic Spans Detection with Named Entity Recognition and Question-Answering Approaches

Phu Gia Hoang, Luan Thanh Nguyen, Kiet Nguyen

International Workshop on Semantic Evaluation (SemEval) 2021

To tackle toxic language detection in SemEval-2021 Task 5, we explored both NER and QA-based approaches. Our spaCy-based NER model achieved the best F1-score of 66.99% in identifying toxic spans from text.

[Paper] 8 citations

UIT-E10dot3 at SemEval-2021 Task 5: Toxic Spans Detection with Named Entity Recognition and Question-Answering Approaches

Phu Gia Hoang, Luan Thanh Nguyen, Kiet Nguyen

International Workshop on Semantic Evaluation (SemEval) 2021

To tackle toxic language detection in SemEval-2021 Task 5, we explored both NER and QA-based approaches. Our spaCy-based NER model achieved the best F1-score of 66.99% in identifying toxic spans from text.

[Paper] 8 citations