arXiv subjects

Haotian Zhong

Publications and source records attributed to Haotian Zhong.

4 recordsLinked to original sources

arXiv2026

Understanding Is Done Early: A Depth Division of Labor in Large Language Models and Its Use for Unbounded-Context Memory

Transformer depth is not used uniformly: lower and middle layers build semantic representations, while upper layers increasingly specialize them for prediction. We turn this division of labor into CoMem (Comprehension Memory), which writes each context chunk only through an intermediate layer, retrieves a fixed number of cached residual states, and recomputes the query-conditioned upper layers over the resulting pack. For a fixed retrieval budget, model-side read compute and memory are independent of stored-context length. We evaluate a continued-trained Qwen3-8B base LM under a unified chat-template-free protocol. The backbone is frozen; the flagship trains only a rank-32 self-distillation LoRA on plain PG19, and we report an adapter-free arm separately. CoMem reaches 97.05 on RULER and 38.27 on LoCoMo versus 34.59 for full-context KV-Direct; the dialogue-memory advantage survives conversation-cluster resampling and an independent judge. Results on additional long-context and long-document tasks expose both the benefits of bounded retrieval and its in-window compression tax. Controlled depth sweeps show that deeper caching lowers per-query recomputation but incurs a fidelity loss that self-distillation substantially repairs. In a separate adapter-free efficiency control on an NVIDIA H20 at 128k, CoMem uses 18.26 GB rather than 89.36 GB and achieves a 7.83x prefill speedup. These results show that long-context memory can be organized along the layer axis, not only the token axis.

cs.CL↗

arXiv2025

Putting a Price on Immobility: Food Deliveries and Pricing Approaches

Urban food delivery services have become an integral part of daily life, yet their mobility and environmental externalities remain poorly addressed by planners. Most studies neglect whether consumers pay enough to internalize the broader social costs of these services. This study quantifies the value of access to and use of food delivery services in Beijing, China, through two discrete choice experiments. The first measures willingness to accept compensation for giving up access, with a median value of CNY588 (approximately USD80). The second captures willingness to pay for reduced waiting time and improved reliability, showing valuations far exceeding typical delivery fees (e.g., CNY96.6/hour and CNY4.83/min at work). These results suggest a substantial consumer surplus and a clear underpricing problem. These findings highlight the need for urban planning to integrate digital service economies into pricing and mobility frameworks. We propose a quantity-based pricing model that targets delivery speed rather than order volume, addressing the primary source of externalities while maintaining net welfare gains. This approach offers a pragmatic, equity-conscious strategy to curb delivery-related congestion, emissions, and safety risks, especially in dense urban cores.

econ.GN↗

arXiv2023

Deep Learning Enabled Design of Terahertz High-Q Metamaterials

Metamaterials open up a new way to manipulate electromagnetic waves and realize various functional devices. Metamaterials with high-quality (Q) resonance responses are widely employed in sensing, detection, and other applications. Traditional design of metamaterials involves laborious simulation-optimization and limits the efficiency. The high-Q metamaterials with abrupt spectral change are even harder to reverse design on-demand. In this paper, we propose novel solutions for designing terahertz high-Q metamaterials based on deep learning, including the forward prediction of spectral responses and the inverse design of structural parameters. For the forward prediction, we develop the Electromagnetic Response Transformer (ERT) model to establish the complex mapping relations between the highly sensitive structural parameters and the abrupt spectra, and realize precise prediction of the high-Q resonance in terahertz spectra from given structural parameters. For the inverse design, we introduce the Visual Attention Network (VAN) model with a large model capability to attentively learn the abrupt shifts in spectral resonances, which can efficiently reduce errors and achieve highly accurate inverse design of structural parameters according to the expected high-Q resonance responses. Both models exhibit outstanding performance, and the accuracy is improved one or two orders higher compared to the traditional machine learning methods. Besides, our ERT model can be 4000 times faster than the conventional full wave simulations in computation time. Our work provides new avenues for the deep learning enabled design of terahertz high-Q metamaterials, which holds potential applications in various fields, such as terahertz communication, sensing, imaging, and functional devices.

physics.optics↗

arXiv2019

A two-dimensional propensity score matching method for longitudinal quasi-experimental studies: A focus on travel behavior and the built environment

The lack of longitudinal studies of the relationship between the built environment and travel behavior has been widely discussed in the literature. This paper discusses how standard propensity score matching estimators can be extended to enable such studies by pairing observations across two dimensions: longitudinal and cross-sectional. Researchers mimic randomized controlled trials (RCTs) and match observations in both dimensions, to find synthetic control groups that are similar to the treatment group and to match subjects synthetically across before-treatment and after-treatment time periods. We call this a two-dimensional propensity score matching (2DPSM). This method demonstrates superior performance for estimating treatment effects based on Monte Carlo evidence. A near-term opportunity for such matching is identifying the impact of transportation infrastructure on travel behavior.

econ.EM↗