Selected work

Representative Papers

Ghost Vectors: Soft-Deleted Embeddings Remain Reconstructible in HNSW Vector Databases

Jun 16, 2026

This study addresses a critical compliance gap in HNSW-based vector databases: their soft deletion mechanisms fail to meet stringent data erasure requirements under regulations such as GDPR and HIPAA, as sensitive information—including personally identifiable information (PII)—can be reconstructed with near-perfect accuracy (empirically achieving 100% PII recovery) from residual embeddings. To mitigate this risk, the authors propose Epoch Key Rotation, a novel approach that encrypts embeddings upon deletion and immediately discards the encryption key, thereby rendering the data irrecoverable. The method further incorporates ECDSA signatures to generate cryptographically verifiable proofs of deletion. Experimental results demonstrate that this technique reduces PII recovery rates to 0% while maintaining high efficiency, with per-record deletion latency of approximately 0.005 milliseconds, thus simultaneously ensuring regulatory compliance, performance, and auditability.

0 citationsRead paper

Agentic Design of Compositional Descriptors via Autoresearch for Materials Science Applications

May 14, 2026

This work proposes an automated descriptor design methodology that eliminates the need for manual feature engineering to predict the bandgap and ferromagnetic Curie temperature of inorganic materials solely from their chemical formulas. Leveraging an autoresearch paradigm, an AI agent named Automat—powered by OpenAI Codex as its code-generation engine—iteratively generates, implements, and evaluates task-specific descriptors within a random forest framework. This approach achieves, for the first time, fully automated construction of composition-based descriptors that simultaneously exhibit high predictive performance and chemical interpretability. On both prediction tasks, the method significantly outperforms established baselines, including fractional descriptors, Magpie features, and their combinations, thereby overcoming the limitations inherent in conventional handcrafted or template-driven feature engineering strategies.

0 citationsRead paper

Scalability Matters: Overcoming Challenges in InstructGLM with Similarity-Degree-Based Sampling

May 02, 2025

Large language models (LLMs) struggle to model graph-structured data efficiently due to token-length limitations, lack of native graph awareness, and reliance on graph neural networks (GNNs). To address this, we propose InstructGLM—the first GNN-free, pure-LLM framework for graph learning. Our method introduces: (1) a similarity-degree joint-driven biased random walk mechanism for adaptive and scalable graph substructure sampling; and (2) a graph-structured instruction-tuning paradigm coupled with token-efficient serialization, overcoming the long-graph-input bottleneck. Experiments demonstrate that InstructGLM matches or surpasses state-of-the-art GNN baselines on node classification and link prediction across multiple large-scale graph benchmarks, while reducing redundant token consumption by over 30%. The framework achieves high scalability and inherent interpretability without compromising performance.

0 citationsRead paper

Generating new coordination compounds via multireference simulations, genetic algorithms and machine learning: the case of Co(II) molecular magnets

Apr 18, 2025

The rational design of Co(II)-based mononuclear molecular magnets is traditionally time-consuming and computationally expensive. Method: We propose a closed-loop intelligent design framework integrating multireference ab initio methods (CASPT2/NEVPT2), genetic algorithms, and graph neural networks to autonomously generate novel, out-of-database organic ligands and enable efficient, targeted exploration of chemical space. Machine learning–guided pre-screening accelerates candidate structure evaluation, while genetic algorithm–driven intelligent sampling enhances the probability of discovering high-energy-barrier configurations. Contribution/Results: Within minutes, the framework automatically designed multiple new Co(II) complexes exhibiting record-breaking magnetic relaxation energy barriers (U_eff > 1000 K) and effective operating temperatures (> 4 K)—surpassing both state-of-the-art experimental and purely computational approaches. This work establishes a scalable, rational design paradigm for functional coordination compounds.

0 citationsRead paper

Recent publications

Latest Papers

Ghost Vectors: Soft-Deleted Embeddings Remain Reconstructible in HNSW Vector Databases

Jun 16, 2026

This study addresses a critical compliance gap in HNSW-based vector databases: their soft deletion mechanisms fail to meet stringent data erasure requirements under regulations such as GDPR and HIPAA, as sensitive information—including personally identifiable information (PII)—can be reconstructed with near-perfect accuracy (empirically achieving 100% PII recovery) from residual embeddings. To mitigate this risk, the authors propose Epoch Key Rotation, a novel approach that encrypts embeddings upon deletion and immediately discards the encryption key, thereby rendering the data irrecoverable. The method further incorporates ECDSA signatures to generate cryptographically verifiable proofs of deletion. Experimental results demonstrate that this technique reduces PII recovery rates to 0% while maintaining high efficiency, with per-record deletion latency of approximately 0.005 milliseconds, thus simultaneously ensuring regulatory compliance, performance, and auditability.

0 citationsRead paper

Agentic Design of Compositional Descriptors via Autoresearch for Materials Science Applications

May 14, 2026

This work proposes an automated descriptor design methodology that eliminates the need for manual feature engineering to predict the bandgap and ferromagnetic Curie temperature of inorganic materials solely from their chemical formulas. Leveraging an autoresearch paradigm, an AI agent named Automat—powered by OpenAI Codex as its code-generation engine—iteratively generates, implements, and evaluates task-specific descriptors within a random forest framework. This approach achieves, for the first time, fully automated construction of composition-based descriptors that simultaneously exhibit high predictive performance and chemical interpretability. On both prediction tasks, the method significantly outperforms established baselines, including fractional descriptors, Magpie features, and their combinations, thereby overcoming the limitations inherent in conventional handcrafted or template-driven feature engineering strategies.

0 citationsRead paper

Scalability Matters: Overcoming Challenges in InstructGLM with Similarity-Degree-Based Sampling

May 02, 2025

Large language models (LLMs) struggle to model graph-structured data efficiently due to token-length limitations, lack of native graph awareness, and reliance on graph neural networks (GNNs). To address this, we propose InstructGLM—the first GNN-free, pure-LLM framework for graph learning. Our method introduces: (1) a similarity-degree joint-driven biased random walk mechanism for adaptive and scalable graph substructure sampling; and (2) a graph-structured instruction-tuning paradigm coupled with token-efficient serialization, overcoming the long-graph-input bottleneck. Experiments demonstrate that InstructGLM matches or surpasses state-of-the-art GNN baselines on node classification and link prediction across multiple large-scale graph benchmarks, while reducing redundant token consumption by over 30%. The framework achieves high scalability and inherent interpretability without compromising performance.

0 citationsRead paper

Generating new coordination compounds via multireference simulations, genetic algorithms and machine learning: the case of Co(II) molecular magnets

Apr 18, 2025

The rational design of Co(II)-based mononuclear molecular magnets is traditionally time-consuming and computationally expensive. Method: We propose a closed-loop intelligent design framework integrating multireference ab initio methods (CASPT2/NEVPT2), genetic algorithms, and graph neural networks to autonomously generate novel, out-of-database organic ligands and enable efficient, targeted exploration of chemical space. Machine learning–guided pre-screening accelerates candidate structure evaluation, while genetic algorithm–driven intelligent sampling enhances the probability of discovering high-energy-barrier configurations. Contribution/Results: Within minutes, the framework automatically designed multiple new Co(II) complexes exhibiting record-breaking magnetic relaxation energy barriers (U_eff > 1000 K) and effective operating temperatures (> 4 K)—surpassing both state-of-the-art experimental and purely computational approaches. This work establishes a scalable, rational design paradigm for functional coordination compounds.

0 citationsRead paper