Selected work

Representative Papers

Ontology Concept Overlap as a Training Signal: Knowledge-Grounded Reinforcement Learning for Clinical Question Answering

Oct 05, 2026

This study addresses the challenge of defining reinforcement learning rewards in clinical question answering, where executable verifiers are unavailable and subtle entity-level errors are prevalent. To this end, it proposes a soft verification method grounded in concept overlap within the UMLS ontology. This approach pioneers the use of controlled-vocabulary concept overlap as a graded external reward signal, integrating an entropy-normalized LLM judge with consistency penalties and optimizing model training via the GRPO algorithm to effectively capture entity substitution errors and enhance safety evaluation. Experimental results demonstrate that the proposed model achieves up to a 39% improvement in Token-F1 on MedQA and PubMedQA, significantly outperforming supervised fine-tuning baselines while exhibiting strong transferability.

0 citationsRead paper

Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders

Oct 01, 2026

This study addresses the issue that personalized encoders lose critical preference evidence when compressing user histories, thereby constraining downstream task performance. To overcome this limitation, we propose REPAIR, a method that introduces a novel "encoder-host" repair mechanism. With the encoder frozen, REPAIR recovers lost preference information by comparing cached representations with current states, enabling state correction without re-encoding. Furthermore, it integrates a compact learned coordinate space, timestep-level pattern selection, and aggregated correction injection techniques. Experimental results demonstrate that REPAIR significantly improves MRR and nDCG metrics across multiple recommendation datasets and enhances personalized generation responsiveness by over 25%, comprehensively outperforming conventional fine-tuning paradigms.

0 citationsRead paper

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

Oct 01, 2026

This study addresses the challenge of unifying and reusing multi-source heterogeneous user histories by proposing the PerTIDE architecture. By designing an action-item shared event schema, it constructs a reusable user memory update mechanism. Integrating action gating, multi-timescale state space trajectories, and command-conditioned readout techniques, the framework enables frozen transfer of core modules across data sources and achieves unified encoding for both predictive and generative modalities, supported by theoretical invariance guarantees. Experimental results demonstrate that PerTIDE improves MRR by 15.23 points on the PENS dataset and outperforms baselines by 4.12 points on MIND, effectively validating the advantages of cross-source training.

0 citationsRead paper

Terminal-Register Certification for Finite-Measurement Learning of Multiscale Quantum States

Sep 27, 2026

This study addresses the lack of reliable certification and noise robustness in multi-scale quantum state learning under finite measurements by proposing a terminal-register-certified MERA learning method. Leveraging an inverse binary MERA architecture, coherent coarse-graining, and terminal joint measurements combined with causal-cone-complete local circuit comparison techniques, this work constructs complete bit-distribution certificates. Furthermore, it introduces a noise-resistance theorem and a calibrated total variation budget to establish rigorous performance guarantees in the finite-sample regime. Experimental results demonstrate that for an 8-qubit system, the proposed approach achieves an average fidelity of 0.9969, significantly outperforming parameter-matched local circuits across metrics including long-range errors, energy, and entropy.

0 citationsRead paper

Endpoint Covering of Axis-Parallel Segments:Bichromatic and Monochromatic One-Center

Sep 24, 2026

This study addresses the minimum enclosing square single-center optimization problem for axis-parallel line segments based on endpoint 1-covering, encompassing both monochromatic and bichromatic settings. Methodologically, it proposes novel algorithms for the monochromatic case under segment constraints alongside a deterministic algorithm for the bichromatic scenario, while establishing matching theoretical lower bounds. Technically, the approach integrates computational geometry, the algebraic decision tree model, and combinatorial optimization. The primary contributions lie in achieving time complexities of O(n log n) for the monochromatic case and O(m + n log²n) for the bichromatic case. Furthermore, the authors prove that these bounds tightly match the established theoretical lower limits, thereby providing optimal solutions for this covering optimization problem.

0 citationsRead paper

Recent publications

Latest Papers

Ontology Concept Overlap as a Training Signal: Knowledge-Grounded Reinforcement Learning for Clinical Question Answering

Oct 05, 2026

This study addresses the challenge of defining reinforcement learning rewards in clinical question answering, where executable verifiers are unavailable and subtle entity-level errors are prevalent. To this end, it proposes a soft verification method grounded in concept overlap within the UMLS ontology. This approach pioneers the use of controlled-vocabulary concept overlap as a graded external reward signal, integrating an entropy-normalized LLM judge with consistency penalties and optimizing model training via the GRPO algorithm to effectively capture entity substitution errors and enhance safety evaluation. Experimental results demonstrate that the proposed model achieves up to a 39% improvement in Token-F1 on MedQA and PubMedQA, significantly outperforming supervised fine-tuning baselines while exhibiting strong transferability.

0 citationsRead paper

Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders

Oct 01, 2026

This study addresses the issue that personalized encoders lose critical preference evidence when compressing user histories, thereby constraining downstream task performance. To overcome this limitation, we propose REPAIR, a method that introduces a novel "encoder-host" repair mechanism. With the encoder frozen, REPAIR recovers lost preference information by comparing cached representations with current states, enabling state correction without re-encoding. Furthermore, it integrates a compact learned coordinate space, timestep-level pattern selection, and aggregated correction injection techniques. Experimental results demonstrate that REPAIR significantly improves MRR and nDCG metrics across multiple recommendation datasets and enhances personalized generation responsiveness by over 25%, comprehensively outperforming conventional fine-tuning paradigms.

0 citationsRead paper

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

Oct 01, 2026

This study addresses the challenge of unifying and reusing multi-source heterogeneous user histories by proposing the PerTIDE architecture. By designing an action-item shared event schema, it constructs a reusable user memory update mechanism. Integrating action gating, multi-timescale state space trajectories, and command-conditioned readout techniques, the framework enables frozen transfer of core modules across data sources and achieves unified encoding for both predictive and generative modalities, supported by theoretical invariance guarantees. Experimental results demonstrate that PerTIDE improves MRR by 15.23 points on the PENS dataset and outperforms baselines by 4.12 points on MIND, effectively validating the advantages of cross-source training.

0 citationsRead paper

Terminal-Register Certification for Finite-Measurement Learning of Multiscale Quantum States

Sep 27, 2026

This study addresses the lack of reliable certification and noise robustness in multi-scale quantum state learning under finite measurements by proposing a terminal-register-certified MERA learning method. Leveraging an inverse binary MERA architecture, coherent coarse-graining, and terminal joint measurements combined with causal-cone-complete local circuit comparison techniques, this work constructs complete bit-distribution certificates. Furthermore, it introduces a noise-resistance theorem and a calibrated total variation budget to establish rigorous performance guarantees in the finite-sample regime. Experimental results demonstrate that for an 8-qubit system, the proposed approach achieves an average fidelity of 0.9969, significantly outperforming parameter-matched local circuits across metrics including long-range errors, energy, and entropy.

0 citationsRead paper

Endpoint Covering of Axis-Parallel Segments:Bichromatic and Monochromatic One-Center

Sep 24, 2026

This study addresses the minimum enclosing square single-center optimization problem for axis-parallel line segments based on endpoint 1-covering, encompassing both monochromatic and bichromatic settings. Methodologically, it proposes novel algorithms for the monochromatic case under segment constraints alongside a deterministic algorithm for the bichromatic scenario, while establishing matching theoretical lower bounds. Technically, the approach integrates computational geometry, the algebraic decision tree model, and combinatorial optimization. The primary contributions lie in achieving time complexities of O(n log n) for the monochromatic case and O(m + n log²n) for the bichromatic case. Furthermore, the authors prove that these bounds tightly match the established theoretical lower limits, thereby providing optimal solutions for this covering optimization problem.

0 citationsRead paper