arXiv subjects

Chunhui Zhu

Publications and source records attributed to Chunhui Zhu.

6 recordsLinked to original sources

arXiv2026

Content Exploration Beyond the Feed: Creator Supply and the Shared Corpus

Industrial recommenders give new content initial views through budgeted exploration, then use early performance to decide further delivery. On many short-video platforms, exploration is the primary way new videos reach viewers. Viewer-side tests measure consumption; the published budget objectives we review omit creator response. We analyze four experiments on a major short-video platform. An eight-month creator ablation finds production exploration raises videos posted per creator by 8.55% and creators posting at least once by 7.10% relative to a minimal floor. A budget-matched reallocation raises creator participation with no detectable short-run viewer-side change. A year-long viewer ablation finds 1.74% more video views but 2.13% less view time. A delivered view creates immediate feed value, can trigger organic take-up, and can induce creator supply. Take-up and supply replenish a shared corpus, creating two measurement limits. Viewer-side A/B tests cancel the corpus effect when both arms consume the same corpus. Giving each arm its own corpus avoids cancellation, but turnover still controls the horizon. If the corpus turns over at rate w per posting cycle, a t-cycle experiment expresses at most wt of the eventual corpus effect. More users reduce noise but do not speed turnover. Before the corpus path visibly bends, data cannot distinguish a modest fast effect from an arbitrarily large slow one, so a valid confidence interval may lack a finite upper endpoint. As predicted, the three-week co-diverted experiment cannot determine the sign of the eventual corpus effect. Within the window, it identifies the direct feed effect, and an exploratory cohort analysis detects organic lift after exploration ends. The experiments establish a positive creator response, measure the gross corpus flow visible within three weeks, and show the design and duration needed to identify total value.

cs.IR↗

arXiv2026

The Price of Isolation: Estimating the Ecosystem Cost of Symmetric Two-Sided A/B Testing

On two-sided content platforms, symmetric two-sided isolation (assigning matched fractions of creators and viewers to isolated treatment and control submarkets) is widely used for creator-side and cold-start experiments because it removes cross-arm marketplace interference. Isolation, however, thins each viewer's candidate catalog, and intuition suggests the resulting engagement cost should fade as the platform grows: a small fraction of a vast catalog is still vast. We show that, in an order-statistics model of engagement, whether this intuition holds depends on the upper tail of match quality. Extreme-value theory yields tail-class loss laws with a sharp dichotomy: for light or bounded tails the loss vanishes as the candidate pool grows, whereas under heavy tails it converges to a size-independent constant, so expanding the candidate pool, even by orders of magnitude, does not asymptotically eliminate the cost. Evidence from two production experiments on a platform with millions of active creators is consistent with this picture: a pure A/A traffic sweep reveals a measurable, depth-graded engagement cost; a one-sided catalog ablation independently shows that per-viewer thinning contributes to the loss; and a tail index calibrated on the small exploration pool predicts an effect consistent with the one observed in the far larger full-catalog ablation. Isolation thus carries a price that experimenters should budget for, like any other cost. We give practitioners a preflight procedure that estimates it before launch, sizes traffic accordingly, and recommends a fallback design when the predicted cost exceeds a chosen tolerance.

cs.IR↗

arXiv2026

LLM-Based Generative Retrieval for Snapchat Content Recommendation

Pretrained large language models (LLMs) are promising retrieval engines because they combine rich semantic priors, strong sequence modeling capabilities, and favorable scaling behavior. However, turning a pretrained LLM into a generative retriever in production deployment raises several challenges: the model must learn an internal item vocabulary that was absent from pretraining, and generate valid item identifiers under strict latency and cost constraints. We address these challenges through the design and launch of SnapLGR, an LLM-based generative retrieval system for short-video recommendation at Snapchat. The system is built around three main designs. First, we construct semantic identifiers (SIDs) from multimodal item embeddings and enhance them with Personalized PageRank (PPR)-based co-engagement contrastive learning, resulting in improved codebook utilization, reduced collisions, and infused collaborative signal. Second, we use continued pretraining (CPT) to ground the introduced SID tokens before supervised fine-tuning (SFT) on user interaction sequences. Third, we make SnapLGR serving practical through TensorRT-LLM CUDA-backed beam search and a decentralized worker-loop architecture. In a live A/B test, the launched system increased View Time by 0.37%, Time Spent by 0.09%, Deep Sessions by 0.18%, and Deep Sessions Unique User by 0.11% relative to the existing TIGER-style generative retrieval baseline. We then decompose this offline gap under a fixed tokenizer and quantify the gains due to model architecture, scaling, and pretraining. Overall, our deployment shows that successful production SnapLGR requires joint design across representation learning, vocabulary grounding, and efficient training and serving.

cs.IR↗

arXiv2026

Electrically steered conduction topologies and period-doubling phase dynamics in VO2

The insulator-to-metal transition (IMT) in strongly correlated materials, such as vanadium dioxide (VO2), offers a transformative platform for next-generation adaptive electronics and neuromorphic computing. However, harnessing this non-equilibrium phase transition for deterministic device operation is fundamentally hindered by the inability to disentangle electric-field effects from Joule heating, owing to a lack of operando techniques capable of resolving phase dynamics at nanoscale spatial and sub-nanosecond temporal scales. Here, using a newly developed electrical-pulse-pump ultrafast transmission electron microscope (E-UTEM), we directly visualize the multi-scale electro-thermo-mechanical dynamics of the IMT in suspended VO2 devices. Our results reveal that electric-field-induced Poole-Frenkel (PF) emission, localized by patterned oxygen vacancies, plays a decisive role in redistributing the internal electric field to trigger a deterministic Mott transition. The extreme non-linearity of this PF effect enables the formation of dynamically reconfigurable connectivity topologies that bypass conventional thermal limits. Furthermore, we observe that the coupling of thermal and elastic energies governs a discrete domain evolution, characterized by step-wise and period-doubling configurational resets, which is a hallmark of non-equilibrium phase dynamics in constrained geometries. By integrating experimental imaging with phase-field simulations, we establish a comprehensive framework for the electrically-driven IMT and predict sub-100-ps switching kinetics. These findings provide a fundamental basis for the rational design of ultrafast, low-energy functional devices through nanoscale defect and strain engineering in correlated systems.

cond-mat.mtrl-sci↗

arXiv2023

The Spectral Integral Method (SIM) for the Scattering from an Arbitrary Number of Circular PEC Cylinders

We present an accurate spectral integral method (SIM) for the analyses of scattering from multiple circular perfect electric conductor (PEC) cylinders. It solves the coupled surface integral equations by using the Fourier series and addition theorem to decouple the system. The SIM has exponential convergence so that the error decreases exponentially with the sample density on the surfaces, and requires only about 2-3 points per wavelength (PPW) to reach engineering accuracy with less than 1% error. Numerical results demonstrate that the SIM is much more accurate and efficient than the method of moments (MoM), and thus can be potentially used as the exact radiation boundary condition in the finite element and spectral element methods.

physics.comp-ph↗

arXiv2016

A High-Performance Mid-infrared Optical Switch Enabled by Bulk Dirac Fermions in Cd3As2

Pulsed lasers operating in the 2-5 {\mu}m band are important for a wide range of applications in sensing, spectroscopy, imaging and communications. Despite recent advances with mid-infrared gain media, the lack of a capable pulse generation mechanism, i.e. a passive optical switch, remains a significant technological challenge. Here we show that mid-infrared optical response of Dirac states in crystalline Cd3As2, a three-dimensional topological Dirac semimetal (TDS), constitutes an ideal ultrafast optical switching mechanism for the 2-5 {\mu}m range. Significantly, fundamental aspects of the photocarrier processes, such as relaxation time scales, are found to be flexibly controlled through element doping, a feature crucial for the development of convenient mid-infrared ultrafast sources. Although various exotic physical phenomena have been uncovered in three-dimensional TDS systems, our findings show for the first time that this emerging class of quantum materials can be harnessed to fill a long known gap in the field of photonics.

cond-mat.mtrl-sci↗