Selected work

Representative Papers

OPERA: Operator-residual feedback for reliable autonomous optical experiments with language-model agents

Aug 06, 2026

This work addresses the limitation of existing autonomous agents in optical experimentation, whose scoring metrics often fail to faithfully capture physical outcomes, leading to ineffective or detrimental decisions. To overcome this, the authors propose the OPERA framework, which introduces—for the first time—an operator-residual feedback mechanism. Experimental actions are modeled as optical operators, and physically interpretable residuals quantify the deviation between actual outcomes and desired states, explicitly decoupling executable actions from physical state errors. Integrating language model agents, optical operator representations, and digital twin technology, OPERA enables closed-loop autonomous control grounded in measurable physical evidence. Evaluated across three optical tasks, the framework reduces the proportion of invalid decisions from 23.6–39.0% to 0.9–1.9%, substantially improving task success rates, stability, and the efficiency of protocol transfer and reconstruction on real instruments.

0 citationsRead paper

A Geometric Probe of the Accuracy-Robustness Trade-off: Sharp Boundaries in Symmetry-Breaking Dimensional Expansion

Feb 20, 2026

This work investigates the geometric origins of the trade-off between clean accuracy and adversarial robustness in deep learning. By introducing Symmetry-Breaking Dimension Expansion (SBDE)—a method that appends constant-valued pixels to inputs to boost accuracy—and combining it with test-time mask projection to analyze robustness effects, the study provides a clear geometric interpretation of this trade-off. For the first time, SBDE is employed as a controllable geometric probe, revealing that accuracy gains arise from the formation of sharp decision boundaries and steep loss gradients along auxiliary dimensions. Experiments on CIFAR-10 demonstrate that ResNet-18’s clean accuracy improves from 90.47% to 95.63%, while mask projection nearly fully restores adversarial robustness, confirming that the induced vulnerability indeed stems from the inserted dimensions.

0 citationsRead paper

Hardware-Friendly Input Expansion for Accelerating Function Approximation

Feb 20, 2026

This work addresses the challenges in one-dimensional function approximation—namely, flat loss landscapes, slow convergence, and difficulty fitting high-frequency components—stemming from parameter symmetry in neural networks. To mitigate these issues, the authors propose a hardware-friendly input space expansion strategy that maps the original one-dimensional input into a higher-dimensional space by incorporating constants such as π. This approach effectively breaks parameter symmetry without increasing the number of model parameters. When combined with the L-BFGS optimizer, the method substantially enhances both training efficiency and approximation accuracy across a range of benchmark one-dimensional functions. Experimental results demonstrate that, with an optimal five-dimensional expansion using π as the expansion constant, the mean squared error is reduced by 66.3% and the average number of L-BFGS iterations decreases by 12%.

0 citationsRead paper

Recent publications

Latest Papers

OPERA: Operator-residual feedback for reliable autonomous optical experiments with language-model agents

Aug 06, 2026

This work addresses the limitation of existing autonomous agents in optical experimentation, whose scoring metrics often fail to faithfully capture physical outcomes, leading to ineffective or detrimental decisions. To overcome this, the authors propose the OPERA framework, which introduces—for the first time—an operator-residual feedback mechanism. Experimental actions are modeled as optical operators, and physically interpretable residuals quantify the deviation between actual outcomes and desired states, explicitly decoupling executable actions from physical state errors. Integrating language model agents, optical operator representations, and digital twin technology, OPERA enables closed-loop autonomous control grounded in measurable physical evidence. Evaluated across three optical tasks, the framework reduces the proportion of invalid decisions from 23.6–39.0% to 0.9–1.9%, substantially improving task success rates, stability, and the efficiency of protocol transfer and reconstruction on real instruments.

0 citationsRead paper

A Geometric Probe of the Accuracy-Robustness Trade-off: Sharp Boundaries in Symmetry-Breaking Dimensional Expansion

Feb 20, 2026

This work investigates the geometric origins of the trade-off between clean accuracy and adversarial robustness in deep learning. By introducing Symmetry-Breaking Dimension Expansion (SBDE)—a method that appends constant-valued pixels to inputs to boost accuracy—and combining it with test-time mask projection to analyze robustness effects, the study provides a clear geometric interpretation of this trade-off. For the first time, SBDE is employed as a controllable geometric probe, revealing that accuracy gains arise from the formation of sharp decision boundaries and steep loss gradients along auxiliary dimensions. Experiments on CIFAR-10 demonstrate that ResNet-18’s clean accuracy improves from 90.47% to 95.63%, while mask projection nearly fully restores adversarial robustness, confirming that the induced vulnerability indeed stems from the inserted dimensions.

0 citationsRead paper

Hardware-Friendly Input Expansion for Accelerating Function Approximation

Feb 20, 2026

This work addresses the challenges in one-dimensional function approximation—namely, flat loss landscapes, slow convergence, and difficulty fitting high-frequency components—stemming from parameter symmetry in neural networks. To mitigate these issues, the authors propose a hardware-friendly input space expansion strategy that maps the original one-dimensional input into a higher-dimensional space by incorporating constants such as π. This approach effectively breaks parameter symmetry without increasing the number of model parameters. When combined with the L-BFGS optimizer, the method substantially enhances both training efficiency and approximation accuracy across a range of benchmark one-dimensional functions. Experimental results demonstrate that, with an optimal five-dimensional expansion using π as the expansion constant, the mean squared error is reduced by 66.3% and the average number of L-BFGS iterations decreases by 12%.

0 citationsRead paper