Selected work

Representative Papers

SEATauBench: Adapting Tool-Agent-User Evaluation Into Low-Resource Southeast Asian Languages

Jun 26, 2026

This work addresses the inadequacy of existing evaluation frameworks in effectively assessing agent capabilities for low-resource languages in Southeast Asia, which fails to reflect the real-world performance of sovereign AI in local contexts. To bridge this gap, we propose SEATauBench—the first multilingual agent benchmark tailored for Southeast Asian sovereign AI—extending the tool-agent-user evaluation paradigm to this region and introducing a reusable, multi-tiered localization pipeline encompassing conversational language, tool descriptions, and task domains. Cross-lingual transfer experiments based on TauBench reveal that while model performance remains relatively stable when only the conversational language is switched, it degrades substantially as localization depth increases, particularly under full-domain adaptation. These findings underscore the severe limitations of evaluations relying solely on English-centric benchmarks.

0 citationsRead paper

Exploring Extrinsic and Intrinsic Properties for Effective Reasoning with Code Interpreter

Jun 15, 2026

This study addresses the lack of systematic understanding of effective behavioral characteristics in code interpreter (CI) reasoning, which limits the potential of large language models (LLMs) in executable computation and iterative verification. The work presents the first systematic characterization of key attributes underlying effective CI reasoning, modeling its mechanisms through both extrinsic critical tokens and intrinsic cognitive behaviors—such as verification, backtracking, and backward chaining. To enhance model capabilities, the authors propose injecting critical tokens during inference and incorporating cognitive behaviors into training. Experiments demonstrate that this approach significantly improves performance across mathematical, sorting, and optimization tasks on multiple mainstream LLMs, while simultaneously increasing token efficiency and reducing unproductive reasoning steps.

0 citationsRead paper

SierpinskiCam: Camera-Controlled Video Retaking with Sierpinski Triangle Pattern Cues

Jun 15, 2026

This work addresses the challenge of novel view synthesis from monocular videos under user-specified camera trajectories, where target views deviating significantly from the source trajectory often lead to failed geometry guidance and sparse or missing newly exposed regions. To tackle this, the paper introduces—for the first time—the Sierpiński triangle dome fractal texture as an auxiliary guide for cross-view feature tracking. Combined with a reference-video conditioning mechanism and a dual-stream token fusion strategy employing negative RoPE indexing, the approach achieves appearance anchoring and high-quality re-rendering without modifying the model architecture or requiring per-video fine-tuning. The proposed method substantially enhances camera controllability, geometric consistency, and visual fidelity in complex re-photography scenarios.

0 citationsRead paper

DuDi: Dual-Signal Distillation with Cross-Lingual Verbalizer

Jun 03, 2026

This work addresses the significant performance degradation of small language models in low-resource multilingual settings, such as those found in Southeast Asia. To mitigate this issue, the authors propose DuDi, a dual-signal distillation framework that uniquely integrates online sequence-level supervision with off-policy/on-policy token-level signals. DuDi further introduces a cross-lingual verbalizer to refine teacher feedback, thereby enhancing the complementarity and transferability of distilled knowledge. Evaluated on the SEA-HELM benchmark, the method consistently outperforms existing distillation approaches across diverse model architectures, scales, and teacher–student configurations, demonstrating both its effectiveness and broad applicability.

0 citationsRead paper

SEA-NLI: Natural Language Inference as a Lens into Southeast Asian Cultural Understanding

Jun 02, 2026

This work addresses the limitation of existing natural language inference (NLI) benchmarks, which predominantly rely on Western contexts or translated data and thus inadequately assess models’ reasoning capabilities in Southeast Asian cultural settings. To bridge this gap, the authors introduce SEA-NLI—the first natively constructed multilingual NLI benchmark spanning eight countries, encompassing both English and local languages, and curated by native speakers with an emphasis on culture-specific knowledge. Evaluations across 17 prominent large language models reveal substantial performance degradation in Southeast Asian cultural contexts, particularly in knowledge-intensive categories. While culturally adapted fine-tuning and culture-aware prompting significantly improve model performance, chain-of-thought (CoT) reasoning yields only marginal gains. This study exposes systematic shortcomings of current models in non-Western cultural reasoning and establishes a new benchmark and pathway for culturally sensitive NLI research.

0 citationsRead paper

Recent publications

Latest Papers

SEATauBench: Adapting Tool-Agent-User Evaluation Into Low-Resource Southeast Asian Languages

Jun 26, 2026

This work addresses the inadequacy of existing evaluation frameworks in effectively assessing agent capabilities for low-resource languages in Southeast Asia, which fails to reflect the real-world performance of sovereign AI in local contexts. To bridge this gap, we propose SEATauBench—the first multilingual agent benchmark tailored for Southeast Asian sovereign AI—extending the tool-agent-user evaluation paradigm to this region and introducing a reusable, multi-tiered localization pipeline encompassing conversational language, tool descriptions, and task domains. Cross-lingual transfer experiments based on TauBench reveal that while model performance remains relatively stable when only the conversational language is switched, it degrades substantially as localization depth increases, particularly under full-domain adaptation. These findings underscore the severe limitations of evaluations relying solely on English-centric benchmarks.

0 citationsRead paper

Exploring Extrinsic and Intrinsic Properties for Effective Reasoning with Code Interpreter

Jun 15, 2026

This study addresses the lack of systematic understanding of effective behavioral characteristics in code interpreter (CI) reasoning, which limits the potential of large language models (LLMs) in executable computation and iterative verification. The work presents the first systematic characterization of key attributes underlying effective CI reasoning, modeling its mechanisms through both extrinsic critical tokens and intrinsic cognitive behaviors—such as verification, backtracking, and backward chaining. To enhance model capabilities, the authors propose injecting critical tokens during inference and incorporating cognitive behaviors into training. Experiments demonstrate that this approach significantly improves performance across mathematical, sorting, and optimization tasks on multiple mainstream LLMs, while simultaneously increasing token efficiency and reducing unproductive reasoning steps.

0 citationsRead paper

SierpinskiCam: Camera-Controlled Video Retaking with Sierpinski Triangle Pattern Cues

Jun 15, 2026

This work addresses the challenge of novel view synthesis from monocular videos under user-specified camera trajectories, where target views deviating significantly from the source trajectory often lead to failed geometry guidance and sparse or missing newly exposed regions. To tackle this, the paper introduces—for the first time—the Sierpiński triangle dome fractal texture as an auxiliary guide for cross-view feature tracking. Combined with a reference-video conditioning mechanism and a dual-stream token fusion strategy employing negative RoPE indexing, the approach achieves appearance anchoring and high-quality re-rendering without modifying the model architecture or requiring per-video fine-tuning. The proposed method substantially enhances camera controllability, geometric consistency, and visual fidelity in complex re-photography scenarios.

0 citationsRead paper

DuDi: Dual-Signal Distillation with Cross-Lingual Verbalizer

Jun 03, 2026

This work addresses the significant performance degradation of small language models in low-resource multilingual settings, such as those found in Southeast Asia. To mitigate this issue, the authors propose DuDi, a dual-signal distillation framework that uniquely integrates online sequence-level supervision with off-policy/on-policy token-level signals. DuDi further introduces a cross-lingual verbalizer to refine teacher feedback, thereby enhancing the complementarity and transferability of distilled knowledge. Evaluated on the SEA-HELM benchmark, the method consistently outperforms existing distillation approaches across diverse model architectures, scales, and teacher–student configurations, demonstrating both its effectiveness and broad applicability.

0 citationsRead paper

SEA-NLI: Natural Language Inference as a Lens into Southeast Asian Cultural Understanding

Jun 02, 2026

This work addresses the limitation of existing natural language inference (NLI) benchmarks, which predominantly rely on Western contexts or translated data and thus inadequately assess models’ reasoning capabilities in Southeast Asian cultural settings. To bridge this gap, the authors introduce SEA-NLI—the first natively constructed multilingual NLI benchmark spanning eight countries, encompassing both English and local languages, and curated by native speakers with an emphasis on culture-specific knowledge. Evaluations across 17 prominent large language models reveal substantial performance degradation in Southeast Asian cultural contexts, particularly in knowledge-intensive categories. While culturally adapted fine-tuning and culture-aware prompting significantly improve model performance, chain-of-thought (CoT) reasoning yields only marginal gains. This study exposes systematic shortcomings of current models in non-Western cultural reasoning and establishes a new benchmark and pathway for culturally sensitive NLI research.

0 citationsRead paper