Selected work

Representative Papers

VulContextBench: A Benchmark for Security Context Retrieval in Coding Agents

Sep 26, 2026

This study addresses the limitation of existing vulnerability detection benchmarks, which evaluate only final verdicts and thus cannot distinguish data-flow reasoning from reliance on pretraining memorization. We construct a benchmark comprising 111 real-world vulnerability-introducing commits, manually audited to eliminate label noise and annotate gold-standard contexts. Uniquely, we propose scoring models based on retrieved evidence rather than final conclusions, introducing a multi-granularity evaluation framework at the file, chunk, and line levels. Experiments reveal that while state-of-the-art coding agents can explore most critical code, their final citation rates fall 37 to 73 percentage points below their viewing rates. This discrepancy highlights a significant gap between models’ ability to discover evidence and their capacity to explicitly cite it in their reasoning.

0 citationsRead paper

CoSA: Context-Aware Severity Assessment via Context Analysis with Large Language Models

Aug 13, 2026

This study addresses the challenges of missing repository-level evidence and noise interference in automated vulnerability severity assessment by proposing CoSA. The method constructs a Code Property Graph and employs a two-stage pruning strategy combined with an explicit metric-guided LLM retrieval mechanism to obtain precise contextual summaries. A lightweight Transformer is then utilized to predict CVSS metrics. Evaluated on a newly constructed high-quality dataset, experimental results demonstrate that CoSA achieves a 14.4% improvement in accuracy and a 15.3% increase in Macro-F1 score compared to existing baselines. These findings confirm that CoSA significantly outperforms current state-of-the-art methods, effectively enabling accurate repository-level vulnerability severity assessment.

0 citationsRead paper

Privacy-Aware Synthetic Video Benchmarking and Relational Evaluation for Worker-Under-Suspended-Load Detection

Jul 17, 2026

This work addresses the lack of publicly shareable video benchmarks for construction sites, particularly for rare, hazardous, and privacy-sensitive relational risks such as “workers under suspended loads.” To bridge this gap, the authors introduce SynthSite, a synthetic video benchmark comprising 55 clips that encompass diverse load configurations and surveillance conditions. They propose a structure-preserving blurring strategy that effectively suppresses worker identity while retaining critical geometric and spatiotemporal relationships, thereby balancing privacy protection with hazard recognition. Experimental results demonstrate that this approach significantly outperforms appearance-smoothing baselines, maintaining high-risk detection performance across five privacy-preserving conditions. Furthermore, the study reveals that preserving only raw visual appearance is insufficient to ensure alignment with human annotations, advocating for a paradigm shift in privacy evaluation from mere appearance obfuscation toward semantic structure preservation.

0 citationsRead paper

Prompt Attack Detection with LLM-as-a-Judge and Mixture-of-Models

Mar 26, 2026

This work addresses the challenge of simultaneously achieving real-time performance and high accuracy in detecting prompt-based attacks—such as jailbreaking and prompt injection—in low-latency production environments. To this end, the authors propose a lightweight, general-purpose large language model (LLM)-based security adjudicator architecture. The system employs a structured reasoning pipeline comprising intent decomposition, safety signal verification, harm assessment, and self-reflection to enable efficient threat detection. Deployed as a centralized protective service within a public-sector chatbot in Singapore, this approach demonstrates for the first time that a lightweight general-purpose LLM (e.g., gemini-2.0-flash-lite-001) can meet stringent low-latency security requirements under real-world production constraints. Additionally, the study investigates hybrid multi-model mechanisms, revealing only marginal performance gains over the single lightweight model.

0 citationsRead paper

Safe at the Margins: A General Approach to Safety Alignment in Low-Resource English Languages -- A Singlish Case Study

Feb 18, 2025

Mainstream large language models (LLMs) exhibit safety misalignment on low-resource English varieties—e.g., Singlish—due to overreliance on Western-centric English data. Method: We propose the first safety alignment framework tailored to non-standard English, introducing KTO-S, a stabilized variant of Kahneman–Tversky Optimization (KTO), the first application of KTO to safety fine-tuning for low-resource languages. We theoretically show that DPO implicitly enforces only weak safety objectives and empirically demonstrate that SFT+KTO significantly outperforms DPO. Results: Applied to SEA-Lion-v2.1-Instruct (a Llama3-8B variant), our method reduces toxicity by 99% on the Singlish safety benchmark, maintains strong generalization to TOXIGEN, and preserves full performance on standard academic benchmarks (MMLU, BBH), with no accuracy degradation.

0 citationsRead paper

Recent publications

Latest Papers

VulContextBench: A Benchmark for Security Context Retrieval in Coding Agents

Sep 26, 2026

This study addresses the limitation of existing vulnerability detection benchmarks, which evaluate only final verdicts and thus cannot distinguish data-flow reasoning from reliance on pretraining memorization. We construct a benchmark comprising 111 real-world vulnerability-introducing commits, manually audited to eliminate label noise and annotate gold-standard contexts. Uniquely, we propose scoring models based on retrieved evidence rather than final conclusions, introducing a multi-granularity evaluation framework at the file, chunk, and line levels. Experiments reveal that while state-of-the-art coding agents can explore most critical code, their final citation rates fall 37 to 73 percentage points below their viewing rates. This discrepancy highlights a significant gap between models’ ability to discover evidence and their capacity to explicitly cite it in their reasoning.

0 citationsRead paper

CoSA: Context-Aware Severity Assessment via Context Analysis with Large Language Models

Aug 13, 2026

This study addresses the challenges of missing repository-level evidence and noise interference in automated vulnerability severity assessment by proposing CoSA. The method constructs a Code Property Graph and employs a two-stage pruning strategy combined with an explicit metric-guided LLM retrieval mechanism to obtain precise contextual summaries. A lightweight Transformer is then utilized to predict CVSS metrics. Evaluated on a newly constructed high-quality dataset, experimental results demonstrate that CoSA achieves a 14.4% improvement in accuracy and a 15.3% increase in Macro-F1 score compared to existing baselines. These findings confirm that CoSA significantly outperforms current state-of-the-art methods, effectively enabling accurate repository-level vulnerability severity assessment.

0 citationsRead paper

Privacy-Aware Synthetic Video Benchmarking and Relational Evaluation for Worker-Under-Suspended-Load Detection

Jul 17, 2026

This work addresses the lack of publicly shareable video benchmarks for construction sites, particularly for rare, hazardous, and privacy-sensitive relational risks such as “workers under suspended loads.” To bridge this gap, the authors introduce SynthSite, a synthetic video benchmark comprising 55 clips that encompass diverse load configurations and surveillance conditions. They propose a structure-preserving blurring strategy that effectively suppresses worker identity while retaining critical geometric and spatiotemporal relationships, thereby balancing privacy protection with hazard recognition. Experimental results demonstrate that this approach significantly outperforms appearance-smoothing baselines, maintaining high-risk detection performance across five privacy-preserving conditions. Furthermore, the study reveals that preserving only raw visual appearance is insufficient to ensure alignment with human annotations, advocating for a paradigm shift in privacy evaluation from mere appearance obfuscation toward semantic structure preservation.

0 citationsRead paper

Prompt Attack Detection with LLM-as-a-Judge and Mixture-of-Models

Mar 26, 2026

This work addresses the challenge of simultaneously achieving real-time performance and high accuracy in detecting prompt-based attacks—such as jailbreaking and prompt injection—in low-latency production environments. To this end, the authors propose a lightweight, general-purpose large language model (LLM)-based security adjudicator architecture. The system employs a structured reasoning pipeline comprising intent decomposition, safety signal verification, harm assessment, and self-reflection to enable efficient threat detection. Deployed as a centralized protective service within a public-sector chatbot in Singapore, this approach demonstrates for the first time that a lightweight general-purpose LLM (e.g., gemini-2.0-flash-lite-001) can meet stringent low-latency security requirements under real-world production constraints. Additionally, the study investigates hybrid multi-model mechanisms, revealing only marginal performance gains over the single lightweight model.

0 citationsRead paper

Safe at the Margins: A General Approach to Safety Alignment in Low-Resource English Languages -- A Singlish Case Study

Feb 18, 2025

Mainstream large language models (LLMs) exhibit safety misalignment on low-resource English varieties—e.g., Singlish—due to overreliance on Western-centric English data. Method: We propose the first safety alignment framework tailored to non-standard English, introducing KTO-S, a stabilized variant of Kahneman–Tversky Optimization (KTO), the first application of KTO to safety fine-tuning for low-resource languages. We theoretically show that DPO implicitly enforces only weak safety objectives and empirically demonstrate that SFT+KTO significantly outperforms DPO. Results: Applied to SEA-Lion-v2.1-Instruct (a Llama3-8B variant), our method reduces toxicity by 99% on the Singlish safety benchmark, maintains strong generalization to TOXIGEN, and preserves full performance on standard academic benchmarks (MMLU, BBH), with no accuracy degradation.

0 citationsRead paper