Selected work

Representative Papers

Simple Methods Defend RAG Systems Well Against Real-World Attacks

Aug 04, 2025

Retrieval-augmented generation (RAG) systems often generate erroneous responses to out-of-distribution (OOD) queries in safety-critical applications. Method: We propose a lightweight, robust OOD detection framework that jointly models retrieval and generation representations via two complementary dimensionality reduction and feature separation strategies—Principal Component Analysis (PCA) and Neural Collapse (NC)—and integrates GPT-4o with regression models for efficient detection. Response quality is rigorously validated through dual-track evaluation using large language models (LLMs) and human annotators. Contribution/Results: Experiments on standard benchmarks and a real-world COVID-19 vaccine chatbot demonstrate significant improvements in OOD detection accuracy and response relevance. Our results underscore the critical role of external OOD detectors in enhancing RAG safety. Notably, the simple PCA-based strategy outperforms complex baselines under realistic adversarial conditions, highlighting its deployment efficiency and practical utility.

0 citationsRead paper

Conventional Contrastive Learning Often Falls Short: Improving Dense Retrieval with Cross-Encoder Listwise Distillation and Synthetic Data

May 25, 2025

Traditional contrastive learning (e.g., InfoNCE) often degrades the performance of state-of-the-art dense retrieval models during corpus-level fine-tuning. This paper identifies and addresses this issue by proposing a robust cross-encoder listwise knowledge distillation framework. Instead of relying on biased human-authored queries, it leverages diverse synthetic queries—declarative, keyword-based, and question-form—generated by large language models. By operating at the ranking-list level rather than sample-wise contrastive matching, the method mitigates instability inherent in pairwise or pointwise contrastive objectives and preserves holistic ranking structure from the teacher model. Evaluated on multiple standard benchmarks, the distilled BERT-based embedding models achieve new state-of-the-art results, outperforming all previously published models in retrieval effectiveness. The code and pretrained models are publicly released.

0 citationsRead paper

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

May 07, 2025

This work addresses the inaccurate hallucination detection by large language models (LLMs) in summarization tasks—particularly the weak faithfulness evaluation in retrieval-augmented generation (RAG) settings. We propose FaithJudge, an LLM-as-a-judge evaluation framework guided by few-shot human annotations. FaithJudge integrates human-annotated priors with self-supervised hallucination modeling to significantly improve automated hallucination detection accuracy. Furthermore, it introduces a dual-track evolutionary hallucination leaderboard that enables dynamic, fine-grained, and task-adaptive quantification and ranking of faithfulness. Extensive experiments demonstrate that FaithJudge outperforms state-of-the-art methods—including HHEM and the Vectara Leaderboard—across multiple benchmarks. By providing a more reliable, scalable, and principled benchmarking infrastructure, FaithJudge advances rigorous, interpretable, and application-aware faithfulness assessment for LLM-generated summaries.

0 citationsRead paper

Teaching Dense Retrieval Models to Specialize with Listwise Distillation and LLM Data Augmentation

Feb 27, 2025

Fine-tuning domain-specialized dense retrieval models often degrades performance—even with hard negative mining and denoising techniques—because the standard pointwise InfoNCE loss impairs generalization. Method: We identify this failure mode and propose listwise knowledge distillation to replace pointwise learning, augmented with high-quality synthetic queries generated by large language models (LLMs); we further employ a cross-encoder teacher to enhance supervision quality. Contribution/Results: Our approach achieves consistent improvements across multiple domain-specific benchmarks. Empirical analysis confirms that LLM-generated queries match human-authored queries in effectiveness. We also find that the cross-encoder teacher’s capacity constitutes the current performance bottleneck. All code and training scripts are publicly released.

0 citationsRead paper

Recent publications

Latest Papers

Simple Methods Defend RAG Systems Well Against Real-World Attacks

Aug 04, 2025

Retrieval-augmented generation (RAG) systems often generate erroneous responses to out-of-distribution (OOD) queries in safety-critical applications. Method: We propose a lightweight, robust OOD detection framework that jointly models retrieval and generation representations via two complementary dimensionality reduction and feature separation strategies—Principal Component Analysis (PCA) and Neural Collapse (NC)—and integrates GPT-4o with regression models for efficient detection. Response quality is rigorously validated through dual-track evaluation using large language models (LLMs) and human annotators. Contribution/Results: Experiments on standard benchmarks and a real-world COVID-19 vaccine chatbot demonstrate significant improvements in OOD detection accuracy and response relevance. Our results underscore the critical role of external OOD detectors in enhancing RAG safety. Notably, the simple PCA-based strategy outperforms complex baselines under realistic adversarial conditions, highlighting its deployment efficiency and practical utility.

0 citationsRead paper

Conventional Contrastive Learning Often Falls Short: Improving Dense Retrieval with Cross-Encoder Listwise Distillation and Synthetic Data

May 25, 2025

Traditional contrastive learning (e.g., InfoNCE) often degrades the performance of state-of-the-art dense retrieval models during corpus-level fine-tuning. This paper identifies and addresses this issue by proposing a robust cross-encoder listwise knowledge distillation framework. Instead of relying on biased human-authored queries, it leverages diverse synthetic queries—declarative, keyword-based, and question-form—generated by large language models. By operating at the ranking-list level rather than sample-wise contrastive matching, the method mitigates instability inherent in pairwise or pointwise contrastive objectives and preserves holistic ranking structure from the teacher model. Evaluated on multiple standard benchmarks, the distilled BERT-based embedding models achieve new state-of-the-art results, outperforming all previously published models in retrieval effectiveness. The code and pretrained models are publicly released.

0 citationsRead paper

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

May 07, 2025

This work addresses the inaccurate hallucination detection by large language models (LLMs) in summarization tasks—particularly the weak faithfulness evaluation in retrieval-augmented generation (RAG) settings. We propose FaithJudge, an LLM-as-a-judge evaluation framework guided by few-shot human annotations. FaithJudge integrates human-annotated priors with self-supervised hallucination modeling to significantly improve automated hallucination detection accuracy. Furthermore, it introduces a dual-track evolutionary hallucination leaderboard that enables dynamic, fine-grained, and task-adaptive quantification and ranking of faithfulness. Extensive experiments demonstrate that FaithJudge outperforms state-of-the-art methods—including HHEM and the Vectara Leaderboard—across multiple benchmarks. By providing a more reliable, scalable, and principled benchmarking infrastructure, FaithJudge advances rigorous, interpretable, and application-aware faithfulness assessment for LLM-generated summaries.

0 citationsRead paper

Teaching Dense Retrieval Models to Specialize with Listwise Distillation and LLM Data Augmentation

Feb 27, 2025

Fine-tuning domain-specialized dense retrieval models often degrades performance—even with hard negative mining and denoising techniques—because the standard pointwise InfoNCE loss impairs generalization. Method: We identify this failure mode and propose listwise knowledge distillation to replace pointwise learning, augmented with high-quality synthetic queries generated by large language models (LLMs); we further employ a cross-encoder teacher to enhance supervision quality. Contribution/Results: Our approach achieves consistent improvements across multiple domain-specific benchmarks. Empirical analysis confirms that LLM-generated queries match human-authored queries in effectiveness. We also find that the cross-encoder teacher’s capacity constitutes the current performance bottleneck. All code and training scripts are publicly released.

0 citationsRead paper