Selected work

Representative Papers

CSB: A Counting and Sampling tool for Bit-vectors

Jul 05, 2026

This work presents the first systematic solution to the problems of model counting and sampling in the theory of bit-vectors. By leveraging bit-blasting to translate bit-vector formulas into conjunctive normal form (CNF), the approach integrates modern CNF counters and samplers to uniformly support a wide range of modes, including exact and approximate counting, projected and unprojected counting, as well as near-uniform and uniform sampling. The resulting tool, csb, addresses a critical gap in efficient counting and sampling for bit-vector constraints and demonstrates substantial performance advantages over existing methods in empirical evaluations, highlighting its practicality and effectiveness.

0 citationsRead paper

Virtual Speech Therapist: A Clinician-in-the-Loop AI Speech Therapy Agent for Personalized and Supervised Therapy

May 01, 2026

This study addresses the limitations of traditional speech therapy, which suffers from low assessment efficiency, heavy burdens in developing personalized treatment plans, and a lack of automated support. The authors propose a virtual speech therapist platform that integrates deep learning with a multi-agent large language model (LLM). The system first employs deep learning to automatically analyze speech samples and classify stuttering types, then leverages multiple LLM agents—collaborating through generation, critique, and iterative refinement—to produce individualized therapy plans. Crucially, a dedicated critique agent enforces adherence to clinical evidence-based guidelines, ensuring safety and regulatory compliance. Expert evaluations demonstrate that the system consistently generates high-quality, evidence-supported therapeutic recommendations, significantly reducing clinical workload under physician supervision while enhancing treatment accessibility and efficacy.

0 citationsRead paper

Evidence Packing for Cross-Domain Image Deepfake Detection with LVLMs

Mar 18, 2026

This work addresses the limitations of existing large vision-language model (LVLM)-based deepfake detection methods, which rely on costly fine-tuning and exhibit poor generalization across domains and to novel forgery types. To overcome these challenges, we propose Semantic-Consistent Evidence Packets (SCEP), a training-free LVLM inference framework that identifies high-confidence suspicious image patches and integrates multi-dimensional cues—semantic, frequency-domain, and noise-based—into a compact, non-redundant evidence set. Using the CLS token from the frozen visual encoder as a global reference, SCEP guides the LVLM to produce interpretable forgery judgments without any parameter updates. Our approach achieves state-of-the-art performance across multiple benchmarks, demonstrating for the first time that high-accuracy, cross-domain deepfake detection is feasible without fine-tuning.

0 citationsRead paper

Reframe Your Life Story: Interactive Narrative Therapist and Innovative Moment Assessment with Large Language Models

Jul 27, 2025

Existing large language models (LLMs) lack domain-specific therapeutic simulation capabilities and fail to track longitudinal healing progress in mental health support. Method: We propose the first interactive LLM system designed specifically for narrative therapy, integrating treatment-stage planning, reflective-level scaffolding, context-aware response generation, and narrative transformation detection to form a closed-loop intervention framework. We further introduce a novel “Moment of Innovation” quantification model to dynamically monitor narrative reconstruction and deliver personalized feedback. Contribution/Results: Evaluated on 260 simulated dialogues and 230 human participants, the system significantly enhances dialogue depth and therapeutic quality (p < 0.01), generating more empowering and socially supportive responses. It overcomes key limitations of general-purpose LLMs—namely insufficient therapeutic depth, poor temporal continuity, and low clinical adaptability—thereby advancing AI’s role in evidence-informed, narrative-based mental health interventions.

0 citationsRead paper

Approximate SMT Counting Beyond Discrete Domains

Jul 24, 2025

Projection model counting for hybrid SMT formulas—i.e., counting satisfying assignments over discrete variables while marginalizing out continuous ones—remains computationally intractable for large-scale instances. Method: This paper introduces the first hashing-based approximate counting framework for SMT, leveraging universal hash functions and an iterative query mechanism to efficiently estimate solution space size. The approach supports formulas with both discrete and continuous variables, requires only logarithmically many calls to an off-the-shelf SMT solver, and provides rigorous probabilistic guarantees on estimation error. Contribution/Results: Evaluated on 14,202 benchmark instances, our method successfully solved 603 counting tasks—nearly 50× more than the best baseline (13), and substantially outperforming all prior approaches in scalability and robustness.

0 citationsRead paper

Recent publications

Latest Papers

CSB: A Counting and Sampling tool for Bit-vectors

Jul 05, 2026

This work presents the first systematic solution to the problems of model counting and sampling in the theory of bit-vectors. By leveraging bit-blasting to translate bit-vector formulas into conjunctive normal form (CNF), the approach integrates modern CNF counters and samplers to uniformly support a wide range of modes, including exact and approximate counting, projected and unprojected counting, as well as near-uniform and uniform sampling. The resulting tool, csb, addresses a critical gap in efficient counting and sampling for bit-vector constraints and demonstrates substantial performance advantages over existing methods in empirical evaluations, highlighting its practicality and effectiveness.

0 citationsRead paper

Virtual Speech Therapist: A Clinician-in-the-Loop AI Speech Therapy Agent for Personalized and Supervised Therapy

May 01, 2026

This study addresses the limitations of traditional speech therapy, which suffers from low assessment efficiency, heavy burdens in developing personalized treatment plans, and a lack of automated support. The authors propose a virtual speech therapist platform that integrates deep learning with a multi-agent large language model (LLM). The system first employs deep learning to automatically analyze speech samples and classify stuttering types, then leverages multiple LLM agents—collaborating through generation, critique, and iterative refinement—to produce individualized therapy plans. Crucially, a dedicated critique agent enforces adherence to clinical evidence-based guidelines, ensuring safety and regulatory compliance. Expert evaluations demonstrate that the system consistently generates high-quality, evidence-supported therapeutic recommendations, significantly reducing clinical workload under physician supervision while enhancing treatment accessibility and efficacy.

0 citationsRead paper

Evidence Packing for Cross-Domain Image Deepfake Detection with LVLMs

Mar 18, 2026

This work addresses the limitations of existing large vision-language model (LVLM)-based deepfake detection methods, which rely on costly fine-tuning and exhibit poor generalization across domains and to novel forgery types. To overcome these challenges, we propose Semantic-Consistent Evidence Packets (SCEP), a training-free LVLM inference framework that identifies high-confidence suspicious image patches and integrates multi-dimensional cues—semantic, frequency-domain, and noise-based—into a compact, non-redundant evidence set. Using the CLS token from the frozen visual encoder as a global reference, SCEP guides the LVLM to produce interpretable forgery judgments without any parameter updates. Our approach achieves state-of-the-art performance across multiple benchmarks, demonstrating for the first time that high-accuracy, cross-domain deepfake detection is feasible without fine-tuning.

0 citationsRead paper

Reframe Your Life Story: Interactive Narrative Therapist and Innovative Moment Assessment with Large Language Models

Jul 27, 2025

Existing large language models (LLMs) lack domain-specific therapeutic simulation capabilities and fail to track longitudinal healing progress in mental health support. Method: We propose the first interactive LLM system designed specifically for narrative therapy, integrating treatment-stage planning, reflective-level scaffolding, context-aware response generation, and narrative transformation detection to form a closed-loop intervention framework. We further introduce a novel “Moment of Innovation” quantification model to dynamically monitor narrative reconstruction and deliver personalized feedback. Contribution/Results: Evaluated on 260 simulated dialogues and 230 human participants, the system significantly enhances dialogue depth and therapeutic quality (p < 0.01), generating more empowering and socially supportive responses. It overcomes key limitations of general-purpose LLMs—namely insufficient therapeutic depth, poor temporal continuity, and low clinical adaptability—thereby advancing AI’s role in evidence-informed, narrative-based mental health interventions.

0 citationsRead paper

Approximate SMT Counting Beyond Discrete Domains

Jul 24, 2025

Projection model counting for hybrid SMT formulas—i.e., counting satisfying assignments over discrete variables while marginalizing out continuous ones—remains computationally intractable for large-scale instances. Method: This paper introduces the first hashing-based approximate counting framework for SMT, leveraging universal hash functions and an iterative query mechanism to efficiently estimate solution space size. The approach supports formulas with both discrete and continuous variables, requires only logarithmically many calls to an off-the-shelf SMT solver, and provides rigorous probabilistic guarantees on estimation error. Contribution/Results: Evaluated on 14,202 benchmark instances, our method successfully solved 603 counting tasks—nearly 50× more than the best baseline (13), and substantially outperforming all prior approaches in scalability and robustness.

0 citationsRead paper