arXiv subjects

Qin Huang

Publications and source records attributed to Qin Huang.

At least 19 recordsLinked to original sources

arXiv2026

MARCO: Click-Intent Decomposition for Calibrated Ads Conversion Prediction

Not all clicks are equal. Industrial ads ranking decouples conversion probability into click-through rate (CTR) and post-click conversion rate (CVR), yet treats every click as the same event. In reality, users provide a free, self-generated signal of intent through their physical UI interactions. Different click types on the same ad exhibit a 4-fold difference in actual conversion rates. By conflating these signals, the standard CVR model under-predicts high-intent clicks and over-predicts low-intent ones, which is a bias masked by near-perfect aggregate calibration. We propose MARCO (Multi-intent Ads Ranking Composition Optimization), a framework that resolves this bias by decomposing each click by intent. Using the logged click type as a free behavioral label, MARCO trains per-intent CVR heads on homogeneous populations, and at serving time composes their per-intent CVR estimates under a predicted distribution over intents. Theoretically, we prove that decomposition never raises population risk, give the exact headroom under squared loss and non-negativity under the deployed loss, and show through a routing-efficiency dial how much of it reaches serving. Because the population-optimal score is unchanged, any gain is a finite-capacity estimation and calibration effect that we validated both offline and online. For deployment at scale, we further cast multi-impression, multi-click attribution as credit assignment with a bias-variance tradeoff analogous to RL return estimation, showing last-impression, first-click attribution is the low-bias, low-variance, deterministic choice under production constraints, and derive three consistency conditions enforced end-to-end at scale. Deployed at binary intent granularity, MARCO corrects per-intent calibration to approximately 100%, lifts conversions per click by +2.80%, and drives +0.98% cumulative improvement in topline metrics.

cs.LG↗

arXiv2026

TALAN: Task-Aligned Latent Adaptation Networks for Targeted Post-Training of Large Language Models

Targeted post-training aims to improve reasoning, math, and code without degrading strengths. Low-rank adapters are efficient but task-global; activation interventions are input-aware but often require separate probes, vectors, or inference-time steering. We introduce TALAN (Task-Aligned Latent Adaptation Networks), a sequence-conditioned latent side path inserted into a transformer's residual stream and co-trained with a low-rank adapter in one SFT loop. TALAN compresses the active sequence into latent memory, remixes it into token-level perturbations, and writes them back through a controlled residual update. It is configured along six axes: insertion location, memory size, mixer, writeback rule, trainability scope, and gradient scale. Across four Qwen3-family backbones and four STEM/code benchmarks, TALAN improves matched LoRA and DoRA baselines. With LoRA, it yields a +1.41 point cross-model mean gain, positive on all four backbones and non-negative on all 16 model-benchmark cells. With DoRA, it yields a +1.85 point mean gain, positive on all backbones and on 13 of 16 cells. Paired seed checks support positive average effects but show nontrivial variance, so we treat them as sensitivity checks. Cost is small: <1% trainable parameters relative to the backbone and 1.01-1.02x inference overhead versus matched LoRA. A Llama-3.2-1B transfer probe is also positive under LoRA and rsLoRA across seven paired seeds, supporting a transfer beyond Qwen. Internal-state analyses suggest TALAN is a small complementary activation intervention. The matched adapter update is 80-1,700x larger than the TALAN perturbation, yet their directions have near-zero cosine; per-layer measurements show this small orthogonal perturbation propagates and amplifies through depth. TALAN offers a practical platform for studying steerable activation-level adaptation within standard adapter-based post-training.

cs.LG↗

arXiv2026

Steering Tropical Cyclones Using Small Perturbations in an AI Weather Model

Tropical cyclone (TC) trajectories are governed by large-scale steering flows and exhibit sensitive dependence on atmospheric initial conditions. Using Hurricane Sandy (2012) in the Aurora AI weather model, we investigate whether targeted thermodynamic perturbations can induce meaningful track deviations. Two distinct perturbation regimes emerge. In the Caribbean, forward finite-time Lyapunov exponent (FTLE) diagnostics identify dynamically sensitive regions within Sandy's steering flow, where perturbations produce substantially larger responses than random placement. In the Pacific, a preferred corridor near 165W influences Sandy through Rossby wave teleconnections, confirmed using Takaya-Nakamura wave activity flux diagnostics. Despite their different physical pathways, both regimes share a common amplification mechanism: small initial perturbations generate modest trajectory offsets that are rapidly amplified when Sandy enters the highly sensitive recurvature region. The largest experiments produce track deviations exceeding 500 km after seven days. These results provide a proof-of-concept demonstration of the Weather Jiu-Jitsu framework, illustrating how targeted perturbations can be amplified through atmospheric dynamics in an AI weather model. Because the required perturbations exceed current operational cloud-seeding capabilities, the experiments should be interpreted as a theoretical sensitivity analysis rather than an operational weather modification strategy.

physics.ao-ph↗

arXiv2026

LoopFM: Learning frOm HistOrical RePresentations of Foundation Model for Recommendation

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical RePresentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5% conversion improvement in the first half after its initial launch, and +1.03% and +1.22% conversion improvement from two individual launches in the subsequent half.

cs.LG↗

arXiv2026

A Simulation Methodology Testbed for Typhoon Sensitivity Analysis: Framework Development and Perturbation-Response Experiments with the Pangu Weather Model

Understanding how typhoons respond to localized perturbations in their environmental fields is fundamental to assessing the limits of predictability and exploring the potential for track or intensity intervention. This study develops a dedicated simulation methodology testbed for typhoon sensitivity analysis by integrating the Pangu weather model, a high-precision AI forecasting system, with Proportional-Integral-Derivative (PID) closed-loop techniques. The testbed is constructed with modular functional blocks including a meteorological prediction module, an artificial perturbation input interface, a typhoon quantitative modeling module, and a PID closed-loop test module, implemented via a cross-platform MATLAB/ONNX technical framework. A Single-Input Single-Output (SISO) test system was built, with velocity and thermal perturbations set as the core inputs and typhoon track and intensity as the key output targets, to perform controlled perturbation-response experiments. The experiments reveal the feasible perturbation-response range, the parameter tuning behavior of the PID module, and the energy-scale response characteristics under different perturbation modes, and quantify the input-output coupling relationships of the test system. By constructing this testbed on an operational AI weather forecasting model, this study provides a framework that goes beyond idealized sensitivity studies typically validated only on low-order dynamical models. The testbed offers an expandable platform for investigating typhoon sensitivity to artificial environmental perturbations and provides a foundation for subsequent expansion toward multi-input multi-output architectures and advanced analysis strategies such as nonlinear PID or model predictive control.

physics.ao-ph↗

arXiv2026

Instability-Aware Steering of an Extreme Atmospheric River in an AI Weather Foundation Model

Advances in deep learning methods for weather forecasting are creating opportunities to computationally explore the potential for steering or control of extreme weather trajectories for societal risk reduction. We present initial investigations into the feasibility of redirecting extreme atmospheric rivers (ARs) through small, instability-aware perturbations. Using the Aurora AI weather foundation model, we identify sensitive upstream locations using finite-time Lyapunov exponents and jet-eddy interaction criteria. We apply an idealized cloud-seeding operator that mimics latent heat release to assess whether these Lyapunov-guided interventions can influence downstream evolution. In a case study of a severe California AR, perturbations induce coherent downstream shifts in moisture transport, reducing intensity at landfall under favorable kinematic conditions. The response is nonlinear and contingent on the local flow geometry. These initial results suggest that the atmosphere's intrinsic chaotic sensitivity could be leveraged for dynamical control, offering a new research direction for extreme event risk mitigation.

physics.ao-ph↗

arXiv2026

SOLARIS: Speculative Offloading of Latent-bAsed Representation for Inference Scaling

Recent advances in recommendation scaling laws have led to foundation models of unprecedented complexity. While these models offer superior performance, their computational demands make real-time serving impractical, often forcing practitioners to rely on knowledge distillation-compromising serving quality for efficiency. To address this challenge, we present SOLARIS (Speculative Offloading of Latent-bAsed Representation for Inference Scaling), a novel framework inspired by speculative decoding. SOLARIS proactively precomputes user-item interaction embeddings by predicting which user-item pairs are likely to appear in future requests, and asynchronously generating their foundation model representations ahead of time. This approach decouples the costly foundation model inference from the latency-critical serving path, enabling real-time knowledge transfer from models previously considered too expensive for online use. Deployed across Meta's advertising system serving billions of daily requests, SOLARIS achieves 0.67% revenue-driving top-line metrics gain, demonstrating its effectiveness at scale.

cs.LG↗

arXiv2026

Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding

We present DeepEarth, a self-supervised multi-modal world model with Earth4D, a novel planetary-scale 4D space-time positional encoder. Earth4D extends 3D multi-resolution hash encoding to include time, efficiently scaling across the planet over centuries with sub-meter, sub-second precision. Multi-modal encoders (e.g. vision-language models) are fused with Earth4D embeddings and trained via masked reconstruction. We demonstrate Earth4D's expressive power by achieving state-of-the-art performance on an ecological forecasting benchmark. Earth4D with learnable hash probing surpasses a multi-modal foundation model pre-trained on substantially more data. Access open source code and download models at: https://github.com/legel/deepearth

cs.AI↗

arXiv2026

Evaluating the Predictability of Selected Weather Extremes with Aurora, an AI Weather Forecast Model

AI weather foundation models now achieve forecast skill comparable to numerical weather prediction at far lower computational cost, yet their predictability for high-impact extremes across dynamical regimes remains uncertain. We evaluate Aurora using an event-based framework spanning tropical cyclones, freezes, heatwaves, atmospheric rivers, and extreme precipitation at lead times from 1 to 21 days. Aurora demonstrates strong short-range (1-7 day) skill across event types, including competitive tropical cyclone track accuracy and high spatial agreement for temperature and moisture extremes. However, a consistent subseasonal failure mode emerges: while large-scale circulation patterns remain moderately skillful at 14-21 day leads, threshold-based extreme intensity collapses as fields regress toward climatology. This divergence indicates that Aurora retains synoptic-scale dynamical structure but loses surface-impact amplitude beyond 7-10 days. The practical predictability horizon for deterministic AI extreme-event forecasting therefore remains constrained by intrinsic atmospheric dynamics.

physics.ao-ph↗

arXiv2025

Meta Lattice: Model Space Redesign for Cost-Effective Industry-Scale Ads Recommendations

The rapidly evolving landscape of products, surfaces, policies, and regulations poses significant challenges for deploying state-of-the-art recommendation models at industry scale, primarily due to data fragmentation across domains and escalating infrastructure costs that hinder sustained quality improvements. To address this challenge, we propose Lattice, a recommendation framework centered around model space redesign that extends Multi-Domain, Multi-Objective (MDMO) learning beyond models and learning objectives. Lattice addresses these challenges through a comprehensive model space redesign that combines cross-domain knowledge sharing, data consolidation, model unification, distillation, and system optimizations to achieve significant improvements in both quality and cost-efficiency. Our deployment of Lattice at Meta has resulted in 10% revenue-driving top-line metrics gain, 11.5% user satisfaction improvement, 6% boost in conversion rate, with 20% capacity saving.

cs.IR↗

arXiv2025

Regime identification and control of extremes in the non-autonomous Lorenz model with chaos and intransitivity

Adaptive chaos control has been studied extensively for autonomous systems. For real world, non-autonomous systems, such as the planetary weather, observations of the system state in response to seasonally and diurnally varying forcing are available only at discrete times and locations, over which system trajectories are likely to have diverged given uncertainties in initial conditions. We consider a stochastic representation of such systems, as a building block for adaptive control, and develop and test control strategies in an idealized setting. We present the first example of finite time adaptive chaos control for a seasonally forced and noise-perturbed Lorenz84 model. We demonstrate two strategies for triggering control: (1) local Lyapunov exponents (LLE), and (2) transition probabilities for the latent states of a non-homogeneous Hidden Markov Model (NHMM). The second approach is motivated by thinking of future applications to a latent embedding space of planetary atmospheric circulation that would get us closer to real world analyses. The NHMM triggers are found to coincide with strongly positive LLE regimes, confirming their dynamical interpretability. These results provide a conceptual bridge towards the use of deep learning based weather and climate foundation models, whose hidden states could be leveraged for adaptive control to mitigate extreme weather events.

nlin.CD↗

arXiv2025

Improving Large-Scale Recommender Systems with Auxiliary Learning

Training large-scale recommendation models under a single global objective implicitly assumes homogeneity across user populations. However, real-world data are composites of heterogeneous cohorts with distinct conditional distributions. As models increase in scale and complexity and as more data is used for training, they become dominated by central distribution patterns, neglecting head and tail regions. This imbalance limits the model's learning ability and can result in inactive attention weights or dead neurons. In this paper, we reveal how the attention mechanism can play a key role in factorization machines for shared embedding selection, and propose to address this challenge by analyzing the substructures in the dataset and exposing those with strong distributional contrast through auxiliary learning. Unlike previous research, which heuristically applies weighted labels or multi-task heads to mitigate such biases, we leverage partially conflicting auxiliary labels to regularize the shared representation. This approach customizes the learning process of attention layers to preserve mutual information with minority cohorts while improving global performance. We evaluated proposed method on massive production datasets with billions of data points each for six SOTA models. Experiments show that the factorization machine is able to capture fine-grained user-ad interactions using the proposed method, achieving up to a 0.16% reduction in normalized entropy overall and delivering gains exceeding 0.30% on targeted minority cohorts.

cs.LG↗

arXiv2025

Affine-Doppler Division Multiplexing for High-Mobility Wireless Communications Systems

Affine Frequency Division Multiplexing (AFDM) has been regarded as a candidate integrated sensing and communications (ISAC) waveform owing to its superior communication performance, outperforming the Orthogonal Time-Frequency Space (OTFS) that has been researched for a longer time. However, since the above two waveforms are incompatible with each other, the state-of-the-art methods well-designed for OTFS may not be directly applicable to AFDM. This paper introduces a new orthogonal multicarrier waveform, namely Affine-Doppler Division Multiplexing (ADDM), which can provide a generic framework and subsume the existing OTFS and AFDM as a particular case. ADDM modulating information symbols in the Affine-Doppler (A-D) domain based on a two-dimensional (2D) transform can enjoy both excellent unambiguous Doppler and Doppler resolution, which is the same as AFDM but outperforms OTFS. Moreover, benefiting from the 2D transform, the symbols block of ADDM in the A-D domain undergoes a 2D cyclic shift produced by the delay and the Doppler of the channel, similar to the 2D cyclic shift in the delay-Doppler domain of cyclic prefix (CP)-OTFS. This offers a potential to directly apply the state-of-the-art methods well-designed for OTFS and AFDM to ADDM. Numerical results show that ADDM achieves comparable BER performance with AFDM but outperforms OTFS in high-mobility scenarios.

eess.SP↗

arXiv2025

Weather Jiu-Jitsu: Climate Adaptation for the 21st Century

Extreme climate events, e.g., droughts, floods, heat waves, and freezes, are becoming more frequent and intense with severe global socio-economic impacts. Growing populations and economic activity leads to increased exposure to these events. Scaling existing physical, financial, and social infrastructure to provide resilience against these extreme events is daunting, especially as one ponders projected climate changes. Climate extremes pose a challenge even if decarbonization and geoengineering are able to regulate Earth's radiation balance. We argue that there is an urgent need to explore a novel adaptive strategy that we call "Weather Jiu-Jitsu," which leverages the intrinsic chaotic dynamics of weather systems to subtly redirect or dissipate their destructive trajectories through precisely timed, small-energy interventions. By leveraging insights from adaptive chaos control, combined with improved observations, prediction and low-energy weather system interventions, humanity could develop a novel nature assisted global infrastructure to limit the impact of climate extremes in the 21st century.

physics.ao-ph↗

arXiv2025

External Large Foundation Model: How to Efficiently Serve Trillions of Parameters for Online Ads Recommendation

Ads recommendation is a prominent service of online advertising systems and has been actively studied. Recent studies indicate that scaling-up and advanced design of the recommendation model can bring significant performance improvement. However, with a larger model scale, such prior studies have a significantly increasing gap from industry as they often neglect two fundamental challenges in industrial-scale applications. First, training and inference budgets are restricted for the model to be served, exceeding which may incur latency and impair user experience. Second, large-volume data arrive in a streaming mode with data distributions dynamically shifting, as new users/ads join and existing users/ads leave the system. We propose the External Large Foundation Model (ExFM) framework to address the overlooked challenges. Specifically, we develop external distillation and a data augmentation system (DAS) to control the computational cost of training/inference while maintaining high performance. We design the teacher in a way like a foundation model (FM) that can serve multiple students as vertical models (VMs) to amortize its building cost. We propose Auxiliary Head and Student Adapter to mitigate the data distribution gap between FM and VMs caused by the streaming data issue. Comprehensive experiments on internal industrial-scale applications and public datasets demonstrate significant performance gain by ExFM.

cs.IR↗

arXiv2025

Unbiased and Error-Detecting Combinatorial Pooling Experiments with Balanced Constant-Weight Gray Codes for Consecutive Positives Detection

Combinatorial pooling schemes have enabled the measurement of thousands of experiments in a small number of reactions. This efficiency is achieved by distributing the items to be measured across multiple reaction units called pools. However, current methods for the design of pooling schemes do not adequately address the need for balanced item distribution across pools, a property particularly important for biological applications. Here, we introduce balanced constant-weight Gray codes for detecting consecutive positives (DCP-CWGCs) for the efficient construction of combinatorial pooling schemes. Balanced DCP-CWGCs ensure uniform item distribution across pools, allow for the identification of consecutive positive items such as overlapping biological sequences, and enable error detection by keeping the number of tests on individual and consecutive positive items constant. For the efficient construction of balanced DCP-CWGCs, we have released an open-source python package codePub, with implementations of the two core algorithms: a branch-and-bound algorithm (BBA) and a recursive combination with BBA (rcBBA). Simulations using codePub show that our algorithms can construct long, balanced DCP-CWGCs that allow for error detection in tractable runtime.

cs.IT↗

arXiv2025

An Integrated Sensing and Communications System Based on Affine Frequency Division Multiplexing

This paper proposes an integrated sensing and communications (ISAC) system based on affine frequency division multiplexing (AFDM) waveform. To this end, a metric set is designed according to not only the maximum tolerable delay/Doppler, but also the weighted spectral efficiency as well as the outage/error probability of sensing and communications. This enables the analytical investigation of the performance trade-offs of AFDM-ISAC system using the derived analytical relation among metrics and AFDM waveform parameters. Moreover, by revealing that delay and the integral/fractional parts of normalized Doppler can be decoupled in the affine Fourier transform-Doppler domain, an efficient estimation method is proposed for our AFDM-ISAC system, whose unambiguous Doppler can break through the limitation of subcarrier spacing. Theoretical analyses and numerical results verify that our proposed AFDM-ISAC system may significantly enlarge unambiguous delay/Doppler while possessing good spectral efficiency and peak-to-sidelobe level ratio in high-mobility scenarios.

eess.SP↗

arXiv2024

An adaptive heavy ball method for ill-posed inverse problems

In this paper we consider ill-posed inverse problems, both linear and nonlinear, by a heavy ball method in which a strongly convex regularization function is incorporated to detect the feature of the sought solution. We develop ideas on how to adaptively choose the step-sizes and the momentum coefficients to achieve acceleration over the Landweber-type method. We then analyze the method and establish its regularization property when it is terminated by the discrepancy principle. Various numerical results are reported which demonstrate the superior performance of our method over the Landweber-type method by reducing substantially the required number of iterations and the computational time.

math.NA↗