Search for: All records

Reinforcement Learning from Human Feedback (RLHF) is critical for aligning Large Language Models (LLMs) with human preferences. However, its efficacy is often compromised by the inherent inconsistency and subjectivity of human annotations. Existing preference optimization frameworks, such as Direct Preference Optimization (DPO), typically treat ambiguous pairs with high annotator disagreement identically to those with unanimous consensus, forcing models to overfit to inconsistent supervision signals and leading to suboptimal alignment. In this work, we propose Reliability-Guided Preference Optimization (RGPO), a robust framework designed to mitigate the impact of inconsistent human feedback. RGPO estimates annotator reliability and infers latent ground truth labels from noisy human feedback to identify robust preferences. Furthermore, we introduce a reliability-aware consistency optimization that dynamically modulates the training objective based on the consensus level of annotations, ensuring the model prioritizes high-consensus supervision signals. Extensive experiments on LLM alignment benchmarks demonstrate that RGPO effectively reduces inconsistency and noise in training data and achieves superior performance compared to widely adopted RLHF baselines. 

On-chip optical filters with ultra-high rejection are essential for quantum photonics, nonlinear optics, and communication systems, where strong pump fields must be suppressed to isolate weak signals. Bragg gratings are attractive for these applications due to their compactness, large free spectral range, and CMOS compatibility, but their rejection levels have often been limited by various leakage noises. Here we present a z-shaped Bragg grating filter that achieves an extinction ratio of 83 dB. The design mitigates unwanted modes and scattering while maintaining a simple and compact geometry. Experimental results confirm stable high-rejection performance, making this filter a strong candidate for integration into quantum photonic circuits and other advanced photonic platforms. 

Chromatin is a hierarchically organized soft material whose nanoscale structure and heterogeneity regulate essential genomic functions. Resolving this organization requires molecular imaging materials that combine selective DNA binding with photophysical properties compatible with nanoscale localization and energy transfer under biologically relevant conditions. Conventional bisbenzimide (Hoechst) DNA stains provide robust targeting of nuclear DNA but limited applicability for super-resolution imaging. Here, we investigate the modular molecular engineering strategy in which cyanine chromophores (Cyanine3, Cyanine5, or Cyanine7) are covalently integrated with a bisbenzimide DNA-binding motif through an aliphatic spacer to form hybrid fluorescent materials. This design decouples DNA recognition from optical functionality, allowing independent optimization of binding affinity and photophysical performance. In the resulting conjugates, the cyanine units retain their intrinsic brightness and spectral properties, while the bisbenzimide ligand preserves high-affinity minor-groove binding to nuclear DNA. The Cyanine3–Hoechst and Cyanine5–Hoechst hybrids enable high-contrast imaging of nuclear DNA in fixed and permeabilized cells using long-wavelength excitation (>500 nm), with strongly suppressed extranuclear background. The substantial spectral overlap between Cyanine3 emission and Cyanine5 absorption further enables efficient Förster resonance energy transfer within the nuclear environment, providing a route to probe nanoscale proximity and organization in chromatin. Owing to their favorable photophysical stability and brightness, these hybrid materials also support single-molecule localization microscopy, revealing nanostructured features within the nucleus that are not resolved in diffraction-limited images. This approach provides a versatile platform for developing next-generation imaging materials tailored to the study of chromatin as a dynamic soft matter system.