Selected work

Representative Papers

Your Embedding Model is SMARTer Than You Think

May 24, 2026

This work addresses the limitations of unimodal single-vector retrieval models, which struggle to preserve fine-grained local information in multimodal tasks, and existing multi-vector approaches that typically require retraining and lack effective global representations. To overcome these challenges, the authors propose SMART, a plug-and-play framework that activates the latent multi-vector capabilities embedded in frozen single-vector models during inference—without any additional training. By fusing global and local information through late interaction of intermediate hidden states, SMART achieves both efficient inference and lightweight adaptation. Experimental results demonstrate that SMART significantly outperforms state-of-the-art multi-vector models on the MMEB-V2 benchmark and visual document retrieval tasks, delivering superior retrieval performance at lower computational cost.

0 citationsRead paper

HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation

Feb 08, 2026

This work addresses the limitations of conventional Euclidean dense retrievers in modeling the hierarchical structure of natural language, which often leads to the retrieval of semantically irrelevant documents and exacerbates hallucination in generation. To overcome this, the authors introduce hyperbolic geometry into RAG retrievers for the first time, proposing a fully hyperbolic Transformer (HyTE-FH) and a hybrid architecture (HyTE-H). They design an Outward Einstein Midpoint pooling operator to preserve hierarchical relationships, employ Lorentz model embeddings with radial norms to explicitly encode document specificity, and incorporate geometric-aware aggregation to prevent representation collapse. Evaluated on MTEB, their approach outperforms Euclidean baselines, and on RAGBench, it achieves up to a 29% improvement in context-answer relevance with a smaller model size, while exhibiting over 20% radial separation in document specificity.

0 citationsRead paper

Reasoning-Augmented Representations for Multimodal Retrieval

Feb 06, 2026

This work addresses the challenge of spurious matches in universal multimodal retrieval when handling queries requiring implicit reasoning—such as coreference resolution or compositional constraints—where reasoning and feature compression are often entangled. The authors propose a data-driven framework that leverages powerful vision-language models to explicitly surface latent semantics prior to retrieval: densely annotating images, resolving ambiguous references, and rewriting complex queries into concise, structured constraints to construct an enhanced training corpus. By externalizing the reasoning process and integrating it into data curation, this approach decouples reasoning from embedding learning, effectively mitigating distributional shifts. Evaluated on the M-BEIR benchmark, the method significantly outperforms strong baselines; ablation studies further demonstrate that corpus augmentation boosts performance on knowledge-intensive queries, while query rewriting proves critical for compositional requests.

0 citationsRead paper

The Challenge of Identifying the Origin of Black-Box Large Language Models

Mar 06, 2025

To address copyright infringement and regulatory compliance risks arising from unauthorized fine-tuning and misuse of black-box large language models (LLMs), this paper proposes PlugAE—the first proactive provenance method operating in the continuous embedding space. Unlike conventional fingerprinting techniques that rely on static output features, PlugAE employs black-box API probing to optimize adversarial tokens in the embedding space and inject them into the target model, thereby eliciting distinguishable response patterns that constitute a robust model fingerprint. Our work systematically exposes fundamental limitations of existing provenance methods in resisting fine-tuning attacks. PlugAE is rigorously validated across 30 open-source LLMs and two real-world commercial black-box APIs, significantly improving detection accuracy for fine-tuned derivative models. The results provide a deployable technical foundation for LLM copyright auditing and inform emerging legal and regulatory frameworks governing model provenance.

0 citationsRead paper

Synthetic Artifact Auditing: Tracing LLM-Generated Synthetic Data Usage in Downstream Applications

Feb 02, 2025

Large language models (LLMs) generating synthetic data introduce bias amplification and hallucination risks when integrated into downstream tools—e.g., classifiers, generators, and visualization systems. Method: This paper proposes the first black-box auditing paradigm specifically designed for synthetic artifacts, requiring no access to model architecture, training data, or parameters. Leveraging only ~200 queries on average, it integrates three disclosure-free, cross-modal, and cross-task auditing strategies: statistical bias measurement, lightweight fine-tuning probes, and supervised classification. Contribution/Results: Evaluated across three text classification, two summarization, and two visualization tasks, the method achieves mean audit accuracy of 0.868–0.880, demonstrating strong effectiveness, practicality, and generalizability in identifying downstream dependence on synthetic data.

0 citationsRead paper

Recent publications

Latest Papers

Your Embedding Model is SMARTer Than You Think

May 24, 2026

This work addresses the limitations of unimodal single-vector retrieval models, which struggle to preserve fine-grained local information in multimodal tasks, and existing multi-vector approaches that typically require retraining and lack effective global representations. To overcome these challenges, the authors propose SMART, a plug-and-play framework that activates the latent multi-vector capabilities embedded in frozen single-vector models during inference—without any additional training. By fusing global and local information through late interaction of intermediate hidden states, SMART achieves both efficient inference and lightweight adaptation. Experimental results demonstrate that SMART significantly outperforms state-of-the-art multi-vector models on the MMEB-V2 benchmark and visual document retrieval tasks, delivering superior retrieval performance at lower computational cost.

0 citationsRead paper

HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation

Feb 08, 2026

This work addresses the limitations of conventional Euclidean dense retrievers in modeling the hierarchical structure of natural language, which often leads to the retrieval of semantically irrelevant documents and exacerbates hallucination in generation. To overcome this, the authors introduce hyperbolic geometry into RAG retrievers for the first time, proposing a fully hyperbolic Transformer (HyTE-FH) and a hybrid architecture (HyTE-H). They design an Outward Einstein Midpoint pooling operator to preserve hierarchical relationships, employ Lorentz model embeddings with radial norms to explicitly encode document specificity, and incorporate geometric-aware aggregation to prevent representation collapse. Evaluated on MTEB, their approach outperforms Euclidean baselines, and on RAGBench, it achieves up to a 29% improvement in context-answer relevance with a smaller model size, while exhibiting over 20% radial separation in document specificity.

0 citationsRead paper

Reasoning-Augmented Representations for Multimodal Retrieval

Feb 06, 2026

This work addresses the challenge of spurious matches in universal multimodal retrieval when handling queries requiring implicit reasoning—such as coreference resolution or compositional constraints—where reasoning and feature compression are often entangled. The authors propose a data-driven framework that leverages powerful vision-language models to explicitly surface latent semantics prior to retrieval: densely annotating images, resolving ambiguous references, and rewriting complex queries into concise, structured constraints to construct an enhanced training corpus. By externalizing the reasoning process and integrating it into data curation, this approach decouples reasoning from embedding learning, effectively mitigating distributional shifts. Evaluated on the M-BEIR benchmark, the method significantly outperforms strong baselines; ablation studies further demonstrate that corpus augmentation boosts performance on knowledge-intensive queries, while query rewriting proves critical for compositional requests.

0 citationsRead paper

The Challenge of Identifying the Origin of Black-Box Large Language Models

Mar 06, 2025

To address copyright infringement and regulatory compliance risks arising from unauthorized fine-tuning and misuse of black-box large language models (LLMs), this paper proposes PlugAE—the first proactive provenance method operating in the continuous embedding space. Unlike conventional fingerprinting techniques that rely on static output features, PlugAE employs black-box API probing to optimize adversarial tokens in the embedding space and inject them into the target model, thereby eliciting distinguishable response patterns that constitute a robust model fingerprint. Our work systematically exposes fundamental limitations of existing provenance methods in resisting fine-tuning attacks. PlugAE is rigorously validated across 30 open-source LLMs and two real-world commercial black-box APIs, significantly improving detection accuracy for fine-tuned derivative models. The results provide a deployable technical foundation for LLM copyright auditing and inform emerging legal and regulatory frameworks governing model provenance.

0 citationsRead paper

Synthetic Artifact Auditing: Tracing LLM-Generated Synthetic Data Usage in Downstream Applications

Feb 02, 2025

Large language models (LLMs) generating synthetic data introduce bias amplification and hallucination risks when integrated into downstream tools—e.g., classifiers, generators, and visualization systems. Method: This paper proposes the first black-box auditing paradigm specifically designed for synthetic artifacts, requiring no access to model architecture, training data, or parameters. Leveraging only ~200 queries on average, it integrates three disclosure-free, cross-modal, and cross-task auditing strategies: statistical bias measurement, lightweight fine-tuning probes, and supervised classification. Contribution/Results: Evaluated across three text classification, two summarization, and two visualization tasks, the method achieves mean audit accuracy of 0.868–0.880, demonstrating strong effectiveness, practicality, and generalizability in identifying downstream dependence on synthetic data.

0 citationsRead paper