Beijing Advanced Sciences and Innovation Center

Quantum dot light-emitting diodes (QLEDs) are widely recognized as a promising next-generation display technology. Yet, they face critical challenges, including the toxicity of cadmium/lead (Cd/Pb)-based heavy metals and the lack of mature, simple, and cost-effective arraying techniques. Herein, we employed heavy-metal-free red/green/blue QDs (R-/G-/B-QDs) as tricolor emissive centers to fabricate single-emissive-layer white QLEDs (SEL-WQLEDs). By regulating charge distribution and Förster resonant energy transfer (FRET) in the R-/G-/B-QD mixed emissive layer (EML), the electroluminescence (EL) performance and spectrum of SEL-WQLEDs were simultaneously improved. The optimal device exhibited a peak external quantum efficiency (EQE) of 4.9% and emitted relatively balanced R/G/B light. Furthermore, we proposed a full-color display technical route in which interference color filters (ICFs) were transferred onto SEL-WQLED pixels. Benefitting from the narrow bandpass and high transmittance characteristics of the ICFs, the resulting QLED display achieved an exceptional color gamut of 113% NTSC (National Television System Committee) and excellent viewing-angle spectral stability over a range of 0°-30°, meeting the application requirements of near-eye displays, i.e., virtual reality/augmented reality (AR/VR). Based on the technical route, we designed a Google Glass-inspired prototype and successfully demonstrated its display functionality. This work offers a promising eco-friendly technical route for developing full-color QLED displays.

Vision-and-Language Navigation (VLN) plays a critical role in tasks of embodied AI, particularly in unseen environments following natural language instructions. Recent advancements leverage large language models (LLMs) to improve the accuracy and generalizability of VLN systems by encoding image sequences as dense token representations. However, this tokenization approach incurs substantial computational overhead due to two key inefficiencies: 1) ego-centric camera views often include navigation-irrelevant re- gions (e.g., sky or distant backgrounds), and 2) high-frame-rate image sequences introduce temporal redundancy. To address these challenges, we propose Spatial-Temporal Efficient Visual Token Pruning (STEP-Nav), a unified frame- work that simultaneously prunes redundant visual tokens and fine-tunes VLN models to preserve navigation performance. In particular, STEP-Nav incorporates a distance- and content-aware token evaluation mechanism to remove irrelevant tokens at the spatial level, along with temporal level similarity-based filtering to reduce redundancy across sequential frames. To ensure pruning does not harm task performance, we introduce a distortion-aware fine-tuning strategy that aligns pruned-token representations with their full-token counterparts while maintaining navigation accuracy. Experiments on the R2R and RxR benchmarks using Navid-CE and NavGPT-2 as base models demonstrate that STEP-Nav preserves over 95% of the performance while reducing 66.7% of tokens, outperforming existing token pruning baselines.

First-principles electronic structure calculations for large-scale material systems with defects or dopants remain a major computational bottleneck in atomistic simulations. Here, we propose a target-driven, non-learning-based method termed HaMLR (Hash-Matching-based Local density-matrix Reuse) to efficiently construct the density matrix of periodic atomic structures containing local defects or dopants. Leveraging the nearsightedness principle of electronic matter, the method systematically scans all atoms in the target system to extract local substructures within a defined nearsightedness radius, thereby covering the full sample space of local environments. Each substructure is encoded based on its geometric and chemical features, hashed, and deduplicated. Distinct substructures are then evaluated using self-consistent density functional theory (DFT) calculations to obtain the density-matrix blocks between the central atom and its neighbors within the cutoff radius. During reconstruction, the full-system density matrix is assembled by matching local environments via hash values and reusing the precomputed local density blocks—thereby avoiding full-scale DFT calculations. Unlike machine learning-based approaches, HaMLR does not require model training, offering improved physical consistency and computational efficiency. Validation on defective graphene, MoS2, and doped silicon demonstrates that HaMLR achieves high accuracy while significantly accelerating density-matrix construction, providing an efficient and robust alternative for large-scale electronic structure modeling.