Selected work

Representative Papers

Asymptotic Risk Calibration for Selective Question Answering

Aug 12, 2026

This work addresses the challenge that large language models often produce fluent yet incorrect answers in question answering, while existing uncertainty scoring methods struggle to reliably distinguish correct from incorrect responses and fail to statistically control error rates with fixed thresholds. To overcome these limitations, the authors propose A-CRC-QA, a novel framework that introduces asymptotic risk calibration into selective question answering for the first time. Inspired by conformal risk control, their approach employs a monotonic empirical risk calibration procedure that reformulates error control as a linear expectation constraint. The method is training-free, model-agnostic, compatible with diverse uncertainty estimators, and applicable to both open-ended and closed-form QA tasks. Experiments on CoQA and MedMCQA demonstrate that A-CRC-QA achieves a significantly better trade-off between answer retention rate and reliability compared to uncalibrated baselines and confidence-bound approaches.

0 citationsRead paper

Uncertainty-Aware Abstention in Large Language Models with Provable Alignment Guarantees

Jul 05, 2026

This work addresses the unreliability of large language models (LLMs) in question answering and their lack of statistically guaranteed uncertainty quantification. To this end, the authors propose the CIC framework, which provides the first finite-sample, provably aligned selective answering mechanism for LLMs. By calibrating arbitrary uncertainty scores using Hoeffding or Clopper–Pearson confidence intervals, CIC constructs response policies that strictly control the error rate under a user-specified risk level. Empirical evaluation across seven prominent LLMs and multiple uncertainty estimators demonstrates that CIC consistently maintains high answer rates while rigorously adhering to the prescribed risk constraints.

0 citationsRead paper

Micrometer-scale displacement and thickness sensing using a single terahertz resonant-tunneling diode

Feb 27, 2026

This work proposes a compact self-mixing radar system based on a single terahertz resonant tunneling diode (RTD) to address the demand for high-precision sensing of micrometer-scale displacements and thin-film thicknesses. For the first time, a single RTD is employed simultaneously as a 280 GHz tunable oscillator and a self-mixing detector, eliminating the need for an external heterodyne receiver chain. By sweeping the oscillation frequency to induce self-mixing interferometry, low-frequency signals are extracted to demodulate displacement and thickness information. Experimental results demonstrate a minimum detectable displacement of approximately 5 μm and successful discrimination of polymer films with thicknesses of 12.5, 25, and 50 μm, confirming the effectiveness and innovation of this approach for miniaturized, high-precision sensing applications.

0 citationsRead paper

RF-MatID: Dataset and Benchmark for Radio Frequency Material Identification

Jan 28, 2026

This work addresses the longstanding limitation in radio-frequency (RF) material identification caused by the absence of large-scale public datasets and systematic benchmarks. We present RF-MatID, the first open-source, large-scale, and wideband (4–43.5 GHz) dataset featuring geometrically diverse samples across 16 fine-grained classes and 5 superclasses, augmented with angular and distance perturbations to emulate real-world conditions. Built upon this dataset, we establish a comprehensive deep learning benchmark comprising 142,000 time–frequency domain samples under multiple protocols and experimental settings, enabling both band-level analysis and practical deployment. Extensive experiments evaluate state-of-the-art models in terms of in-distribution performance and out-of-distribution robustness, providing the community with a reproducible foundation for algorithm development and a standardized evaluation framework.

0 citationsRead paper

Bouncy particle sampler with infinite exchanging parallel tempering

Sep 02, 2025

Sampling from multimodal posterior distributions in Bayesian inference remains challenging due to poor mixing and slow convergence. To address this, we propose a novel framework that integrates the Bouncy Particle Sampler (BPS) with parallel tempering (PT) under the infinite exchange rate regime. Our key innovation lies in reformulating PT in the limit of infinitely frequent temperature swaps, enabling instantaneous state exchanges between discrete temperature levels while preserving continuous-state dynamics—thereby substantially enhancing global exploration. Crucially, the method requires no manual tuning of swap rates or temperature schedules. Numerical experiments demonstrate that our approach achieves faster convergence and higher effective sample size than standard BPS and Hamiltonian Monte Carlo (HMC), particularly for strongly multimodal posteriors. Moreover, it exhibits superior stability and robustness across diverse benchmark problems.

0 citationsRead paper

Recent publications

Latest Papers

Asymptotic Risk Calibration for Selective Question Answering

Aug 12, 2026

This work addresses the challenge that large language models often produce fluent yet incorrect answers in question answering, while existing uncertainty scoring methods struggle to reliably distinguish correct from incorrect responses and fail to statistically control error rates with fixed thresholds. To overcome these limitations, the authors propose A-CRC-QA, a novel framework that introduces asymptotic risk calibration into selective question answering for the first time. Inspired by conformal risk control, their approach employs a monotonic empirical risk calibration procedure that reformulates error control as a linear expectation constraint. The method is training-free, model-agnostic, compatible with diverse uncertainty estimators, and applicable to both open-ended and closed-form QA tasks. Experiments on CoQA and MedMCQA demonstrate that A-CRC-QA achieves a significantly better trade-off between answer retention rate and reliability compared to uncalibrated baselines and confidence-bound approaches.

0 citationsRead paper

Uncertainty-Aware Abstention in Large Language Models with Provable Alignment Guarantees

Jul 05, 2026

This work addresses the unreliability of large language models (LLMs) in question answering and their lack of statistically guaranteed uncertainty quantification. To this end, the authors propose the CIC framework, which provides the first finite-sample, provably aligned selective answering mechanism for LLMs. By calibrating arbitrary uncertainty scores using Hoeffding or Clopper–Pearson confidence intervals, CIC constructs response policies that strictly control the error rate under a user-specified risk level. Empirical evaluation across seven prominent LLMs and multiple uncertainty estimators demonstrates that CIC consistently maintains high answer rates while rigorously adhering to the prescribed risk constraints.

0 citationsRead paper

Micrometer-scale displacement and thickness sensing using a single terahertz resonant-tunneling diode

Feb 27, 2026

This work proposes a compact self-mixing radar system based on a single terahertz resonant tunneling diode (RTD) to address the demand for high-precision sensing of micrometer-scale displacements and thin-film thicknesses. For the first time, a single RTD is employed simultaneously as a 280 GHz tunable oscillator and a self-mixing detector, eliminating the need for an external heterodyne receiver chain. By sweeping the oscillation frequency to induce self-mixing interferometry, low-frequency signals are extracted to demodulate displacement and thickness information. Experimental results demonstrate a minimum detectable displacement of approximately 5 μm and successful discrimination of polymer films with thicknesses of 12.5, 25, and 50 μm, confirming the effectiveness and innovation of this approach for miniaturized, high-precision sensing applications.

0 citationsRead paper

RF-MatID: Dataset and Benchmark for Radio Frequency Material Identification

Jan 28, 2026

This work addresses the longstanding limitation in radio-frequency (RF) material identification caused by the absence of large-scale public datasets and systematic benchmarks. We present RF-MatID, the first open-source, large-scale, and wideband (4–43.5 GHz) dataset featuring geometrically diverse samples across 16 fine-grained classes and 5 superclasses, augmented with angular and distance perturbations to emulate real-world conditions. Built upon this dataset, we establish a comprehensive deep learning benchmark comprising 142,000 time–frequency domain samples under multiple protocols and experimental settings, enabling both band-level analysis and practical deployment. Extensive experiments evaluate state-of-the-art models in terms of in-distribution performance and out-of-distribution robustness, providing the community with a reproducible foundation for algorithm development and a standardized evaluation framework.

0 citationsRead paper

Bouncy particle sampler with infinite exchanging parallel tempering

Sep 02, 2025

Sampling from multimodal posterior distributions in Bayesian inference remains challenging due to poor mixing and slow convergence. To address this, we propose a novel framework that integrates the Bouncy Particle Sampler (BPS) with parallel tempering (PT) under the infinite exchange rate regime. Our key innovation lies in reformulating PT in the limit of infinitely frequent temperature swaps, enabling instantaneous state exchanges between discrete temperature levels while preserving continuous-state dynamics—thereby substantially enhancing global exploration. Crucially, the method requires no manual tuning of swap rates or temperature schedules. Numerical experiments demonstrate that our approach achieves faster convergence and higher effective sample size than standard BPS and Hamiltonian Monte Carlo (HMC), particularly for strongly multimodal posteriors. Moreover, it exhibits superior stability and robustness across diverse benchmark problems.

0 citationsRead paper