Selected work

Representative Papers

A minimum witness for the 3/2 configuration-linear-program gap in two-weight graph balancing, unique at its size

Aug 13, 2026

This study addresses the 3/2 integrality gap between the configuration linear programming (Config-LP) relaxation and integer optimal solutions in restricted assignment scheduling, specifically for instances where each job can be processed on at most two machines and only two distinct processing times exist. By employing graph-theoretic modeling, symmetry reduction, exhaustive search verified with both exact rational and floating-point arithmetic, and complexity analysis, the authors establish the existence of a unique minimal witness instance \( I^* \) with six jobs—improving upon the previously known seven-job example. They further prove that no instance with five or fewer jobs can achieve this gap and that at least four machines are necessary. Additionally, they provide a complete classification of all witness instances with seven jobs (13 in total) and eight jobs (154 in total).

0 citationsRead paper

Two base rates, two weights: base-rate neglect has a second axis

Aug 06, 2026

This study addresses a critical limitation in traditional research on base-rate neglect, which has focused exclusively on the underweighting of outcome priors while overlooking the influence of cue frequencies on judgment, thereby yielding an incomplete understanding of reasoning biases. The authors propose that base-rate neglect comprises two distinct dimensions—outcome priors and cue frequencies—and develop a dual-weight Bayesian model to jointly account for both. Using a novel rating-task experimental paradigm, they identify the cue-density effect as a second form of base-rate neglect and achieve a double dissociation between the two types, overcoming the constraints of single-parameter models. This integrative framework not only explains why six standard models appear convergent in conventional experiments despite underlying divergences but also provides a unified interpretation of classic measures such as causal strength and signal detection metrics.

0 citationsRead paper

Thermalizing Stochastic Programs

Aug 02, 2026

This work proposes the first compiler framework enabling end-to-end mapping of general stochastic programs to thermodynamic sampling hardware. Addressing the challenge of efficiently compiling stochastic programs—expressed as directed factor graphs or parameterized random circuits—onto native energy-based model (EBM) hardware, the approach integrates context-aware pattern matching with a trajectory-level REINFORCE post-training strategy. This combination substantially reduces compilation error and enhances approximation fidelity. Empirical evaluation demonstrates the framework’s effectiveness and generality across diverse applications, including financial market simulation, ecological probabilistic modeling, Gibbs sampling for non-native EBMs, and Bayesian design of Gaussian random circuits, thereby establishing a viable pathway toward energy-efficient stochastic computing.

0 citationsRead paper

Three Failures of Pain Location: Why the Diagnostic Utility of Symptom Localization Is Not One Thing

Jul 28, 2026

This study challenges the conventional view that variations in pain localization accuracy stem solely from a unidimensional gradient of anatomical complexity, arguing instead for multiple underlying mechanisms. It proposes that failures in pain localization arise from three distinct sources: anatomical multiplexing, decentralized amplification, and referred or atypical displacement. For the first time, these are formally unified within a Bayesian generative framework as independent failure modes affecting the likelihood, prior, and loss function—thereby distinguishing between perceived and reported pain location. Integrating inverse problem theory, information theory, and neural field models, the work establishes a formal spatiotemporal dynamic framework. The findings reveal that clinical practice overestimates the diagnostic specificity of pain localization, with its actual utility exhibiting a far more gradual gradient, and offer mechanism-specific optimization strategies for each failure type.

0 citationsRead paper

Text-Preserving Lossy Text Compression: A Study of Strategic Deletion and LLM Reconstruction

May 27, 2026

Traditional lossless text compression struggles to achieve both high fidelity and high compression ratios on natural language. This work proposes a semantically lossy compression framework that strategically removes portions of the input text and leverages large language models to reconstruct the original content from the retained “skeleton.” Multiple deletion strategies—including word frequency, semantic surprisal, linear programming optimization, and hybrid approaches—are designed and evaluated on the BBC News dataset. Results show that WordFreq serves as an efficient baseline, while semantic and hybrid strategies achieve superior reconstruction quality at moderate compression rates. Furthermore, a locally deployed decoder fine-tuned with QLoRA matches the performance of Gemini 2.0 Flash, and the framework demonstrates effective cross-lingual transfer between English and Chinese.

0 citationsRead paper

Recent publications

Latest Papers

A minimum witness for the 3/2 configuration-linear-program gap in two-weight graph balancing, unique at its size

Aug 13, 2026

This study addresses the 3/2 integrality gap between the configuration linear programming (Config-LP) relaxation and integer optimal solutions in restricted assignment scheduling, specifically for instances where each job can be processed on at most two machines and only two distinct processing times exist. By employing graph-theoretic modeling, symmetry reduction, exhaustive search verified with both exact rational and floating-point arithmetic, and complexity analysis, the authors establish the existence of a unique minimal witness instance \( I^* \) with six jobs—improving upon the previously known seven-job example. They further prove that no instance with five or fewer jobs can achieve this gap and that at least four machines are necessary. Additionally, they provide a complete classification of all witness instances with seven jobs (13 in total) and eight jobs (154 in total).

0 citationsRead paper

Two base rates, two weights: base-rate neglect has a second axis

Aug 06, 2026

This study addresses a critical limitation in traditional research on base-rate neglect, which has focused exclusively on the underweighting of outcome priors while overlooking the influence of cue frequencies on judgment, thereby yielding an incomplete understanding of reasoning biases. The authors propose that base-rate neglect comprises two distinct dimensions—outcome priors and cue frequencies—and develop a dual-weight Bayesian model to jointly account for both. Using a novel rating-task experimental paradigm, they identify the cue-density effect as a second form of base-rate neglect and achieve a double dissociation between the two types, overcoming the constraints of single-parameter models. This integrative framework not only explains why six standard models appear convergent in conventional experiments despite underlying divergences but also provides a unified interpretation of classic measures such as causal strength and signal detection metrics.

0 citationsRead paper

Thermalizing Stochastic Programs

Aug 02, 2026

This work proposes the first compiler framework enabling end-to-end mapping of general stochastic programs to thermodynamic sampling hardware. Addressing the challenge of efficiently compiling stochastic programs—expressed as directed factor graphs or parameterized random circuits—onto native energy-based model (EBM) hardware, the approach integrates context-aware pattern matching with a trajectory-level REINFORCE post-training strategy. This combination substantially reduces compilation error and enhances approximation fidelity. Empirical evaluation demonstrates the framework’s effectiveness and generality across diverse applications, including financial market simulation, ecological probabilistic modeling, Gibbs sampling for non-native EBMs, and Bayesian design of Gaussian random circuits, thereby establishing a viable pathway toward energy-efficient stochastic computing.

0 citationsRead paper

Three Failures of Pain Location: Why the Diagnostic Utility of Symptom Localization Is Not One Thing

Jul 28, 2026

This study challenges the conventional view that variations in pain localization accuracy stem solely from a unidimensional gradient of anatomical complexity, arguing instead for multiple underlying mechanisms. It proposes that failures in pain localization arise from three distinct sources: anatomical multiplexing, decentralized amplification, and referred or atypical displacement. For the first time, these are formally unified within a Bayesian generative framework as independent failure modes affecting the likelihood, prior, and loss function—thereby distinguishing between perceived and reported pain location. Integrating inverse problem theory, information theory, and neural field models, the work establishes a formal spatiotemporal dynamic framework. The findings reveal that clinical practice overestimates the diagnostic specificity of pain localization, with its actual utility exhibiting a far more gradual gradient, and offer mechanism-specific optimization strategies for each failure type.

0 citationsRead paper

Text-Preserving Lossy Text Compression: A Study of Strategic Deletion and LLM Reconstruction

May 27, 2026

Traditional lossless text compression struggles to achieve both high fidelity and high compression ratios on natural language. This work proposes a semantically lossy compression framework that strategically removes portions of the input text and leverages large language models to reconstruct the original content from the retained “skeleton.” Multiple deletion strategies—including word frequency, semantic surprisal, linear programming optimization, and hybrid approaches—are designed and evaluated on the BBC News dataset. Results show that WordFreq serves as an efficient baseline, while semantic and hybrid strategies achieve superior reconstruction quality at moderate compression rates. Furthermore, a locally deployed decoder fine-tuned with QLoRA matches the performance of Gemini 2.0 Flash, and the framework demonstrates effective cross-lingual transfer between English and Chinese.

0 citationsRead paper