Searcharxiv⌕ Search

arXiv subjects

Zhuo Zhi

Publications and source records attributed to Zhuo Zhi.

16 recordsLinked to original sources

Firepool: Brokering Warm GPU Capacity for Agentic and RL Workloads at LinkedIn

AI agents and reinforcement-learning (RL) trainers run thousands of short, isolated GPU jobs: tool calls, verifiers, and rollouts. A job often finishes in seconds, but starting a GPU pod for it takes 20-35 s even when the container image is already on the node, and minutes when it is not. A warm pod removes the wait only if it can serve the next job, and keeping pods warm must neither waste GPUs nor leak state between callers. Firepool, the sandbox and execution broker behind agentic and RL workloads at LinkedIn, meets both requirements. It makes pods interchangeable by building them from a shared base image per hardware profile and binding each job's code, extra dependencies, and cluster topology when the job starts. It keeps them warm cheaply and safely by reusing a pod only for the caller that used it last, staging zero-VRAM standby pods in fractional GPU shares alongside active pods, and capping the GPUs that warm pools may hold via queue-driven budgets. Over a sampled peak period in production, Firepool serves 94.7% of 3,206 pod requests from warm or standby capacity. In paired benchmarks with fixed-duration jobs on production H200 nodes, agent and RL loops finish 13x and 10.8x faster than with a fresh pod per job; against a fresh reuse tag per job, warm reuse cuts p90 start latency from 59.1 s to 3.5 s. When a job cannot reuse the previous pod, standby promotion shortens the handover wait from 51 s to 0.94 s. In local and simulated runs, a keep-alive matched to the loop's pauses keeps the latency benefit while holding about half the capacity of an always-on pool.

cs.DC↗

Agent Stack for ML Training Lifecycle: Generation, Validation, and Optimization

Production model development requires repeated architecture design, platform integration, debugging, training, and evaluation. We present an agent stack deployed at LinkedIn with two independently usable agent loops: AutoTorch for model generation and migration, and AutoResearch for model optimization. AutoTorch generates PyTorch models and training pipelines for TensorFlow migration, XGBoost replacement, and new-model development. Executable validation hooks check platform API compliance, trainability, and conversion fidelity on sampled data, guiding iterative repair before engineer review and full-scale distributed training. AutoResearch explores architecture and training-configuration changes through expert-steered or autonomous experimentation. An external orchestrator manages parallel trials, experiment history, candidate selection, and wall-clock budgets across agent sessions. A shared runtime preserves execution state independently of agent reasoning and allocates GPUs separately from agent execution. Across ranking, content understanding, and recommendation workloads, AutoTorch completed generation and full-scale training in 92 of 115 workflow runs (80%). Of these, 77 (83.7%) produced models meeting post-training quality requirements. In a separate cohort, AutoResearch improved the primary task metric on held-out data for 28 of 100 production pipelines. Engineering teams estimate that the stack reduced elapsed end-to-end migration time from approximately eight weeks to two weeks, a 4x speedup.

cs.SE↗

Scalable Distributed Learning-Based MPC for Heterogeneous Consumer-Prosumer Building Aggregations: MPC-Aware Feature Selection and Convex Constraint-Coupled Optimization

Buildings provide demand-response flexibility through controllable thermal loads and thermal inertia, while distributed photovoltaic deployment creates opportunities to share surplus generation across heterogeneous consumer-prosumer aggregations. This study develops a scalable distributed learning-based model predictive control (MPC) framework that preserves convexity to enable a fully distributed optimizer with an asymptotic global-optimality guarantee. The Input-Convex Encoder-Only Transformer (IC-EoT) is embedded through recursive substitution to preserve local MPC convexity, while the aggregation-wide energy-trading balance yields a convex constraint-coupled problem solved by Tracking Alternating Direction Method of Multipliers (Tracking-ADMM) with bounded-neighbour communication and fully distributed finite-iteration stopping. An MPC-aware feature-selection procedure additionally accounts for future-value availability and recursive error propagation. In EnergyPlus-Energym co-simulation, centralized MPC becomes unable to meet the real-time solution requirement when the aggregation size reaches six buildings. In contrast, distributed MPC (DMPC) remains real-time feasible throughout the tested range and shows only mild solution-time growth, accompanied primarily by increasing Tracking ADMM iterations as aggregation size grows. Where centralized references are available, finite-iteration DMPC approaches the centralized objective while satisfying the stopping tolerances. Coordination yields bill savings of up to 9.6% when sufficient photovoltaic surplus and consumer demand coincide, whereas low-surplus savings can be masked by closed-loop numerical variation associated with finite solver convergence tolerances. The proposed feature sets reduce horizon-mean weighted standardized primary-target mean squared error by up to 9.7% and 33.8% for the office and apartment testbeds, respectively.

eess.SY↗

End-to-End Differentiable Predictive Control with Probabilistic Constraint-Satisfaction Guarantees for Building Demand Response

Building Demand Response (DR) benefits from Model Predictive Control (MPC) because thermal flexibility can be optimized over a future horizon while enforcing operational constraints, but the repeated online optimization required by MPC motivates Differentiable Predictive Control (DPC) as a computationally efficient explicit-policy alternative. Conventional DPC, however, separates model identification from policy optimization and remains sensitive to plant-model mismatch. This paper proposes an End-to-End DPC (E2E-DPC) framework that jointly optimizes the learned dynamics model and control policy using prediction, constraint, and economic objectives. Stage-wise conformal prediction (CP) calibrates recursive prediction errors to construct horizon-dependent tightened state constraints, while independent Hoeffding certification provides a high-confidence lower bound on plant-level satisfaction of the original state and input constraints under receding-horizon deployment. The framework is evaluated in a high-fidelity residential building simulation using EnergyPlus through the Energym interface and compared with online MPC and conventional DPC. Constraint tightening eliminates thermal violations for both DPC variants in the stress test and reduces violations by 99.4 percent for online MPC. Compared with conventional DPC, E2E-DPC reduces the mean constraint-tightening margin at the first prediction stage by 31.0 percent and increases the median certified plant-level lower bound from 0 to 0.814. The results further reveal a trade-off between stronger certification of plant-level constraint satisfaction and the ability to maintain nonempty tightened state sets as the prediction horizon increases, with stronger multi-step identification mitigating this trade-off.

eess.SY↗

Input Convex Encoder-Only Transformer for Computationally Efficient Model Predictive Control in Building Demand Response

Learning-based Model Predictive Control (MPC) has emerged as a powerful strategy for building demand response (DR). However, its practical deployment is often hindered by the non-convex optimization problems induced by standard neural network models. These problems lead to long solver times and a lack of global optimality guarantees, making long-horizon real-time control challenging. Because forecasting building thermal dynamics and energy consumption relies heavily on historical time-series data, Input Convex Neural Networks (ICNNs) tailored for temporal tasks have been developed to address this issue. Notable examples include Input Convex Recurrent Neural Networks (ICRNNs) and Input Convex Long Short-Term Memory (IC-LSTM). Although IC-LSTM improves the modelling capability of simpler ICRNNs, its recurrent architecture remains computationally expensive and, in the experiments reported here, exhibits numerical training instability at longer historical sequence lengths. To address these limitations, this paper introduces the Input Convex Encoder-Only Transformer (IC-EoT), which combines an encoder-only attention architecture with input convexity guarantees. IC-EoT was evaluated in a co-simulation framework using the Energym Python library and the EnergyPlus building simulator, and compared with IC-LSTM and standard non-convex models in residential and commercial testbeds. IC-EoT maintained predictive accuracy comparable to IC-LSTM, showed no non-finite loss events across 60 multi-seed training runs, and reduced mean per-epoch training time by 26.2% and 29.8% in the two testbeds. At the eight-hour prediction horizon, IC-EoT reduced mean MPC solution time by factors of 5.8 and 5.3 relative to IC-LSTM, while providing comparable closed-loop cost and comfort performance.

eess.SY↗

Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models

Recent studies posit that Reinforcement Learning with Verifiable Rewards (RLVR) primarily amplifies behaviors inherent to the pre-training distribution rather than inducing new capabilities, but these insights are predominantly limited to language-only domains, leaving the dynamics of visual-centric spatial reasoning under-explored. To examine the impact of RLVR on the capability boundaries of Vision-Language Models (VLMs), we introduce \textbf{Ariadne}, a controlled framework based on synthetic maze navigation where the reasoning difficulty is precisely regulated by path length and the number of turns. We demonstrate that applying RLVR extends the spatial reasoning boundary, achieving success on problems where the base policy VLM consistently attains $0\%$ accuracy despite increasing pass@k sampling budgets, indicating that the optimized policy successfully navigates search spaces that were effectively unreachable by the base distribution. Furthermore, despite being trained exclusively on synthetic mazes, we evaluate the model on two real-world navigation benchmarks (MapBench and ReasonMap) in a zero-shot setting. The observed improvements in these out-of-domain tasks suggest genuine spatial reasoning capability expansion rather than mere sampling efficiency.

cs.AI↗

Deconstructing the Failure of Ideal Noise Correction: A Three-Pillar Diagnosis

Statistically consistent methods based on the noise transition matrix ($T$) offer a theoretically grounded solution to Learning with Noisy Labels (LNL), with guarantees of convergence to the optimal clean-data classifier. In practice, however, these methods are often outperformed by empirical approaches such as sample selection, and this gap is usually attributed to the difficulty of accurately estimating $T$. The common assumption is that, given a perfect $T$, noise-correction methods would recover their theoretical advantage. In this work, we put this longstanding hypothesis to a decisive test. We conduct experiments under idealized conditions, providing correction methods with a perfect, oracle transition matrix. Even under these ideal conditions, we observe that these methods still suffer from performance collapse during training. This compellingly demonstrates that the failure is not fundamentally a $T$-estimation problem, but stems from a more deeply rooted flaw. To explain this behaviour, we provide a unified analysis that links three levels: macroscopic convergence states, microscopic optimisation dynamics, and information-theoretic limits on what can be learned from noisy labels. Together, these results give a formal account of why ideal noise correction fails and offer concrete guidance for designing more reliable methods for learning with noisy labels.

cs.LG↗

VideoAgent2: Enhancing the LLM-Based Agent System for Long-Form Video Understanding by Uncertainty-Aware CoT

Long video understanding has emerged as an increasingly important yet challenging task in computer vision. Agent-based approaches are gaining popularity for processing long videos, as they can handle extended sequences and integrate various tools to capture fine-grained information. However, existing methods still face several challenges: (1) they often rely solely on the reasoning ability of large language models (LLMs) without dedicated mechanisms to enhance reasoning in long video scenarios; and (2) they remain vulnerable to errors or noise from external tools. To address these issues, we propose a specialized chain-of-thought (CoT) process tailored for long video analysis. Our proposed CoT with plan-adjust mode enables the LLM to incrementally plan and adapt its information-gathering strategy. We further incorporate heuristic uncertainty estimation of both the LLM and external tools to guide the CoT process. This allows the LLM to assess the reliability of newly collected information, refine its collection strategy, and make more robust decisions when synthesizing final answers. Empirical experiments show that our uncertainty-aware CoT effectively mitigates noise from external tools, leading to more reliable outputs. We implement our approach in a system called VideoAgent2, which also includes additional modules such as general context acquisition and specialized tool design. Evaluation on three dedicated long video benchmarks (and their subsets) demonstrates that VideoAgent2 outperforms the previous state-of-the-art agent-based method, VideoAgent, by an average of 13.1% and achieves leading performance among all zero-shot approaches

cs.CV↗

Seeing and Reasoning with Confidence: Supercharging Multimodal LLMs with an Uncertainty-Aware Agentic Framework

Multimodal large language models (MLLMs) show promise in tasks like visual question answering (VQA) but still face challenges in multimodal reasoning. Recent works adapt agentic frameworks or chain-of-thought (CoT) reasoning to improve performance. However, CoT-based multimodal reasoning often demands costly data annotation and fine-tuning, while agentic approaches relying on external tools risk introducing unreliable output from these tools. In this paper, we propose Seeing and Reasoning with Confidence (SRICE), a training-free multimodal reasoning framework that integrates external vision models with uncertainty quantification (UQ) into an MLLM to address these challenges. Specifically, SRICE guides the inference process by allowing MLLM to autonomously select regions of interest through multi-stage interactions with the help of external tools. We propose to use a conformal prediction-based approach to calibrate the output of external tools and select the optimal tool by estimating the uncertainty of an MLLM's output. Our experiment shows that the average improvement of SRICE over the base MLLM is 4.6% on five datasets and the performance on some datasets even outperforms fine-tuning-based methods, revealing the significance of ensuring reliable tool use in an MLLM agent.

cs.AI↗

PROSAC: Provably Safe Certification for Machine Learning Models under Adversarial Attacks

It is widely known that state-of-the-art machine learning models, including vision and language models, can be seriously compromised by adversarial perturbations. It is therefore increasingly relevant to develop capabilities to certify their performance in the presence of the most effective adversarial attacks. Our paper offers a new approach to certify the performance of machine learning models in the presence of adversarial attacks with population level risk guarantees. In particular, we introduce the notion of $(α,ζ)$-safe machine learning model. We propose a hypothesis testing procedure, based on the availability of a calibration set, to derive statistical guarantees providing that the probability of declaring that the adversarial (population) risk of a machine learning model is less than $α$ (i.e. the model is safe), while the model is in fact unsafe (i.e. the model adversarial population risk is higher than $α$), is less than $ζ$. We also propose Bayesian optimization algorithms to determine efficiently whether a machine learning model is $(α,ζ)$-safe in the presence of an adversarial attack, along with statistical guarantees. We apply our framework to a range of machine learning models - including various sizes of vision Transformer (ViT) and ResNet models - impaired by a variety of adversarial attacks, such as PGDAttack, MomentumAttack, GenAttack and BanditAttack, to illustrate the operation of our approach. Importantly, we show that ViT's are generally more robust to adversarial attacks than ResNets, and large models are generally more robust than smaller models. Our approach goes beyond existing empirical adversarial risk-based certification guarantees. It formulates rigorous (and provable) performance guarantees that can be used to satisfy regulatory requirements mandating the use of state-of-the-art technical tools.

cs.LG↗

Deep Learning-Based Noninvasive Screening of Type 2 Diabetes with Chest X-ray Images and Electronic Health Records

The imperative for early detection of type 2 diabetes mellitus (T2DM) is challenged by its asymptomatic onset and dependence on suboptimal clinical diagnostic tests, contributing to its widespread global prevalence. While research into noninvasive T2DM screening tools has advanced, conventional machine learning approaches remain limited to unimodal inputs due to extensive feature engineering requirements. In contrast, deep learning models can leverage multimodal data for a more holistic understanding of patients' health conditions. However, the potential of chest X-ray (CXR) imaging, one of the most commonly performed medical procedures, remains underexplored. This study evaluates the integration of CXR images with other noninvasive data sources, including electronic health records (EHRs) and electrocardiography signals, for T2DM detection. Utilising datasets meticulously compiled from the MIMIC-IV databases, we investigated two deep fusion paradigms: an early fusion-based multimodal transformer and a modular joint fusion ResNet-LSTM architecture. The end-to-end trained ResNet-LSTM model achieved an AUROC of 0.86, surpassing the CXR-only baseline by 2.3% with just 9863 training samples. These findings demonstrate the diagnostic value of CXRs within multimodal frameworks for identifying at-risk individuals early. Additionally, the dataset preprocessing pipeline has also been released to support further research in this domain.

cs.LG↗

Borrowing Treasures from Neighbors: In-Context Learning for Multimodal Learning with Missing Modalities and Data Scarcity

Multimodal machine learning with missing modalities is an increasingly relevant challenge arising in various applications such as healthcare. This paper extends the current research into missing modalities to the low-data regime, i.e., a downstream task has both missing modalities and limited sample size issues. This problem setting is particularly challenging and also practical as it is often expensive to get full-modality data and sufficient annotated training samples. We propose to use retrieval-augmented in-context learning to address these two crucial issues by unleashing the potential of a transformer's in-context learning ability. Diverging from existing methods, which primarily belong to the parametric paradigm and often require sufficient training samples, our work exploits the value of the available full-modality data, offering a novel perspective on resolving the challenge. The proposed data-dependent framework exhibits a higher degree of sample efficiency and is empirically demonstrated to enhance the classification model's performance on both full- and missing-modality data in the low-data regime across various multimodal learning tasks. When only 1% of the training data are available, our proposed method demonstrates an average improvement of 6.1% over a recent strong baseline across various datasets and missing states. Notably, our method also reduces the performance gap between full-modality and missing-modality data compared with the baseline.

cs.LG↗

HgbNet: predicting hemoglobin level/anemia degree from EHR data

Anemia is a prevalent medical condition that typically requires invasive blood tests for diagnosis and monitoring. Electronic health records (EHRs) have emerged as valuable data sources for numerous medical studies. EHR-based hemoglobin level/anemia degree prediction is non-invasive and rapid but still faces some challenges due to the fact that EHR data is typically an irregular multivariate time series containing a significant number of missing values and irregular time intervals. To address these issues, we introduce HgbNet, a machine learning-based prediction model that emulates clinicians' decision-making processes for hemoglobin level/anemia degree prediction. The model incorporates a NanDense layer with a missing indicator to handle missing values and employs attention mechanisms to account for both local irregularity and global irregularity. We evaluate the proposed method using two real-world datasets across two use cases. In our first use case, we predict hemoglobin level/anemia degree at moment T+1 by utilizing records from moments prior to T+1. In our second use case, we integrate all historical records with additional selected test results at moment T+1 to predict hemoglobin level/anemia degree at the same moment, T+1. HgbNet outperforms the best baseline results across all datasets and use cases. These findings demonstrate the feasibility of estimating hemoglobin levels and anemia degree from EHR data, positioning HgbNet as an effective non-invasive anemia diagnosis solution that could potentially enhance the quality of life for millions of affected individuals worldwide. To our knowledge, HgbNet is the first machine learning model leveraging EHR data for hemoglobin level/anemia degree prediction.

cs.LG↗

Volctrans Parallel Corpus Filtering System for WMT 2020

In this paper, we describe our submissions to the WMT20 shared task on parallel corpus filtering and alignment for low-resource conditions. The task requires the participants to align potential parallel sentence pairs out of the given document pairs, and score them so that low-quality pairs can be filtered. Our system, Volctrans, is made of two modules, i.e., a mining module and a scoring module. Based on the word alignment model, the mining module adopts an iterative mining strategy to extract latent parallel sentences. In the scoring module, an XLM-based scorer provides scores, followed by reranking mechanisms and ensemble. Our submissions outperform the baseline by 3.x/2.x and 2.x/2.x for km-en and ps-en on From Scratch/Fine-Tune conditions, which is the highest among all submissions.

cs.CL↗

Extended Answer and Uncertainty Aware Neural Question Generation

In this paper, we study automatic question generation, the task of creating questions from corresponding text passages where some certain spans of the text can serve as the answers. We propose an Extended Answer-aware Network (EAN) which is trained with Word-based Coverage Mechanism (WCM) and decodes with Uncertainty-aware Beam Search (UBS). The EAN represents the target answer by its surrounding sentence with an encoder, and incorporates the information of the extended answer into paragraph representation with gated paragraph-to-answer attention to tackle the problem of the inadequate representation of the target answer. To reduce undesirable repetition, the WCM penalizes repeatedly attending to the same words at different time-steps in the training stage. The UBS aims to seek a better balance between the model confidence in copying words from an input text paragraph and the confidence in generating words from a vocabulary. We conduct experiments on the SQuAD dataset, and the results show our approach achieves significant performance improvement.

cs.CL↗

Generative Hierarchical Models for Parts, Objects, and Scenes

Compositional structures between parts and objects are inherent in natural scenes. Modeling such compositional hierarchies via unsupervised learning can bring various benefits such as interpretability and transferability, which are important in many downstream tasks. In this paper, we propose the first deep latent variable model, called RICH, for learning Representation of Interpretable Compositional Hierarchies. At the core of RICH is a latent scene graph representation that organizes the entities of a scene into a tree structure according to their compositional relationships. During inference, taking top-down approach, RICH is able to use higher-level representation to guide lower-level decomposition. This avoids the difficult problem of routing between parts and objects that is faced by bottom-up approaches. In experiments on images containing multiple objects with different part compositions, we demonstrate that RICH is able to learn the latent compositional hierarchy and generate imaginary scenes.

cs.LG↗