Aman Gupta

Johannes Schmude, Sujit Roy, Will Trojak et al.

Triggered by the realization that AI emulators can rival the performance of traditional numerical weather prediction models running on HPC systems, there is now an increasing number of large AI models that address use cases such as forecasting, downscaling, or nowcasting. While the parallel developments in the AI literature focus on foundation models -- models that can be effectively tuned to address multiple, different use cases -- the developments on the weather and climate side largely focus on single-use cases with particular emphasis on mid-range forecasting. We close this gap by introducing Prithvi WxC, a 2.3 billion parameter foundation model developed using 160 variables from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). Prithvi WxC employs an encoder-decoder-based architecture, incorporating concepts from various recent transformer models to effectively capture both regional and global dependencies in the input data. The model has been designed to accommodate large token counts to model weather phenomena in different topologies at fine resolutions. Furthermore, it is trained with a mixed objective that combines the paradigms of masked reconstruction with forecasting. We test the model on a set of challenging downstream tasks namely: Autoregressive rollout forecasting, Downscaling, Gravity wave flux parameterization, and Extreme events estimation. The pretrained model with 2.3 billion parameters, along with the associated fine-tuning workflows, has been publicly released as an open-source contribution via Hugging Face.

Kayhan Behdin, Qingquan Song, Aman Gupta et al.

Modern deep learning models are over-parameterized, where different optima can result in widely varying generalization performance. The Sharpness-Aware Minimization (SAM) technique modifies the fundamental loss function that steers gradient descent methods toward flatter minima, which are believed to exhibit enhanced generalization prowess. Our study delves into a specific variant of SAM known as micro-batch SAM (mSAM). This variation involves aggregating updates derived from adversarial perturbations across multiple shards (micro-batches) of a mini-batch during training. We extend a recently developed and well-studied general framework for flatness analysis to theoretically show that SAM achieves flatter minima than SGD, and mSAM achieves even flatter minima than SAM. We provide a thorough empirical evaluation of various image classification and natural language processing tasks to substantiate this theoretical advancement. We also show that contrary to previous work, mSAM can be implemented in a flexible and parallelizable manner without significantly increasing computational costs. Our implementation of mSAM yields superior generalization performance across a wide range of tasks compared to SAM, further supporting our theoretical framework.

Kayhan Behdin, Qingquan Song, Aman Gupta et al.

Modern deep learning models are over-parameterized, where the optimization setup strongly affects the generalization performance. A key element of reliable optimization for these systems is the modification of the loss function. Sharpness-Aware Minimization (SAM) modifies the underlying loss function to guide descent methods towards flatter minima, which arguably have better generalization abilities. In this paper, we focus on a variant of SAM known as mSAM, which, during training, averages the updates generated by adversarial perturbations across several disjoint shards of a mini-batch. Recent work suggests that mSAM can outperform SAM in terms of test accuracy. However, a comprehensive empirical study of mSAM is missing from the literature -- previous results have mostly been limited to specific architectures and datasets. To that end, this paper presents a thorough empirical evaluation of mSAM on various tasks and datasets. We provide a flexible implementation of mSAM and compare the generalization performance of mSAM to the performance of SAM and vanilla training on different image classification and natural language processing tasks. We also conduct careful experiments to understand the computational cost of training with mSAM, its sensitivity to hyperparameters and its correlation with the flatness of the loss landscape. Our analysis reveals that mSAM yields superior generalization performance and flatter minima, compared to SAM, across a wide range of tasks without significantly increasing computational costs.

Kayhan Behdin, Ayan Acharya, Aman Gupta et al.

With the rising popularity of Large Language Models (LLMs), there has been an increasing interest in compression techniques that enable their efficient deployment. This study focuses on the Post-Training Quantization (PTQ) of LLMs. Drawing from recent advances, our work introduces QuantEase, a layer-wise quantization framework where individual layers undergo separate quantization. The problem is framed as a discrete-structured non-convex optimization, prompting the development of algorithms rooted in Coordinate Descent (CD) techniques. These CD-based methods provide high-quality solutions to the complex non-convex layer-wise quantization problems. Notably, our CD-based approach features straightforward updates, relying solely on matrix and vector operations, circumventing the need for matrix inversion or decomposition. We also explore an outlier-aware variant of our approach, allowing for retaining significant weights (outliers) with complete precision. Our proposal attains state-of-the-art performance in terms of perplexity and zero-shot accuracy in empirical evaluations across various LLMs and datasets, with relative improvements up to 15% over methods such as GPTQ. Leveraging careful linear algebra optimizations, QuantEase can quantize models like Falcon-180B on a single NVIDIA A100 GPU in $\sim$3 hours. Particularly noteworthy is our outlier-aware algorithm's capability to achieve near or sub-3-bit quantization of LLMs with an acceptable drop in accuracy, obviating the need for non-uniform quantization or grouping techniques, improving upon methods such as SpQR by up to two times in terms of perplexity.

Jiwoo Hong, Noah Lee, Eunki Kim et al.

The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss are prone to over-optimization, losing generalizability to unseen input distributions. In this paper, we study the cause of over-optimization in RM training and its downstream effects on the RLHF procedure, accentuating the importance of distributional robustness of RMs in unseen data. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Then, we propose batch-wise sum-to-zero regularization (BSR) to enforce zero-centered reward sum per batch, constraining the rewards with extreme magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness. Subsequently, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0 reduces generation length by 40% while adding a 7% increase in win rate, further highlighting that robustness in RMs induces robustness in RLHF training. We release the code, data, and models: https://github.com/LinkedIn-XFACT/RM-Robustness.

Fedor Borisyuk, Qingquan Song, Mingzhou Zhou et al.

This paper introduces LiNR, LinkedIn's large-scale, GPU-based retrieval system. LiNR supports a billion-sized index on GPU models. We discuss our experiences and challenges in creating scalable, differentiable search indexes using TensorFlow and PyTorch at production scale. In LiNR, both items and model weights are integrated into the model binary. Viewing index construction as a form of model training, we describe scaling our system for large indexes, incorporating full scans and efficient filtering. A key focus is on enabling attribute-based pre-filtering for exhaustive GPU searches, addressing the common challenge of post-filtering in KNN searches that often reduces system quality. We further provide multi-embedding retrieval algorithms and strategies for tackling cold start issues in retrieval. Our advancements in supporting larger indexes through quantization are also discussed. We believe LiNR represents one of the industry's first Live-updated model-based retrieval indexes. Applied to out-of-network post recommendations on LinkedIn Feed, LiNR has contributed to a 3% relative increase in professional daily active users. We envisage LiNR as a step towards integrating retrieval and ranking into a single GPU model, simplifying complex infrastructures and enabling end-to-end optimization of the entire differentiable infrastructure through gradient descent.

Rajat Shinde, Christopher E. Phillips, Kumar Ankur et al.

High-quality machine learning (ML)-ready datasets play a foundational role in developing new artificial intelligence (AI) models or fine-tuning existing models for scientific applications such as weather and climate analysis. Unfortunately, despite the growing development of new deep learning models for weather and climate, there is a scarcity of curated, pre-processed machine learning (ML)-ready datasets. Curating such high-quality datasets for developing new models is challenging particularly because the modality of the input data varies significantly for different downstream tasks addressing different atmospheric scales (spatial and temporal). Here we introduce WxC-Bench (Weather and Climate Bench), a multi-modal dataset designed to support the development of generalizable AI models for downstream use-cases in weather and climate research. WxC-Bench is designed as a dataset of datasets for developing ML-models for a complex weather and climate system, addressing selected downstream tasks as machine learning phenomenon. WxC-Bench encompasses several atmospheric processes from meso-$β$ (20 - 200 km) scale to synoptic scales (2500 km), such as aviation turbulence, hurricane intensity and track monitoring, weather analog search, gravity wave parameterization, and natural language report generation. We provide a comprehensive description of the dataset and also present a technical validation for baseline analysis. The dataset and code to prepare the ML-ready data have been made publicly available on Hugging Face -- https://huggingface.co/datasets/nasa-impact/WxC-Bench

Vansh Kapoor, Aman Gupta, Hao Chen et al.

Multi-step reasoning tasks like mathematical problem solving are vulnerable to cascading failures, where a single incorrect step leads to complete solution breakdown. Current LLM routing methods assign entire queries to one model, treating all reasoning steps as equal. We propose TRIM (Targeted routing in multi-step reasoning tasks), which routes only critical steps$\unicode{x2013}$those likely to derail the solution$\unicode{x2013}$to larger models while letting smaller models handle routine continuations. Our key insight is that targeted step-level interventions can fundamentally transform inference efficiency by confining expensive calls to precisely those steps where stronger models prevent cascading errors. TRIM operates at the step-level: it uses process reward models to identify erroneous steps and makes routing decisions based on step-level uncertainty and budget constraints. We develop several routing strategies within TRIM, ranging from a simple threshold-based policy to more expressive policies that reason about long-horizon accuracy-cost trade-offs and uncertainty in step-level correctness estimates. On MATH-500, even the simplest thresholding strategy surpasses prior routing methods with 5x higher cost efficiency, while more advanced policies match the strong, expensive model's performance using 80% fewer expensive model tokens. On harder benchmarks such as AIME, TRIM achieves up to 6x higher cost efficiency. All methods generalize effectively across math reasoning tasks, demonstrating that step-level difficulty represents fundamental characteristics of reasoning.

Zhoutong Fu, Yihan Cao, Yi-Lin Chen et al.

Large language models (LLMs) have achieved strong performance across a wide range of natural language processing tasks. However, deploying LLMs at scale for domain specific applications, such as job-person fit and explanation in job seeking platforms, introduces distinct challenges. At LinkedIn, the job person fit task requires analyzing a candidate's public profile against job requirements to produce both a fit assessment and a detailed explanation. Directly applying open source or finetuned LLMs to this task often fails to yield high quality, actionable feedback due to the complexity of the domain and the need for structured outputs. Moreover, the large size of these models leads to high inference latency and limits scalability, making them unsuitable for online use. To address these challenges, we introduce LANTERN, a novel LLM knowledge distillation framework tailored specifically for job person fit tasks. LANTERN involves modeling over multiple objectives, an encoder model for classification purpose, and a decoder model for explanation purpose. To better distill the knowledge from a strong black box teacher model to multiple downstream models, LANTERN incorporates multi level knowledge distillation that integrates both data and logit level insights. In addition to introducing the knowledge distillation framework, we share our insights on post training techniques and prompt engineering, both of which are crucial for successfully adapting LLMs to domain specific downstream tasks. Extensive experimental results demonstrate that LANTERN significantly improves task specific metrics for both job person fit and explanation. Online evaluations further confirm its effectiveness, showing measurable gains in job seeker engagement, including a 0.24\% increase in apply rate and a 0.28\% increase in qualified applications.

David Durfee, Aman Gupta, Kinjal Basu

We introduce the notion of heterogeneous calibration that applies a post-hoc model-agnostic transformation to model outputs for improving AUC performance on binary classification tasks. We consider overconfident models, whose performance is significantly better on training vs test data and give intuition onto why they might under-utilize moderately effective simple patterns in the data. We refer to these simple patterns as heterogeneous partitions of the feature space and show theoretically that perfectly calibrating each partition separately optimizes AUC. This gives a general paradigm of heterogeneous calibration as a post-hoc procedure by which heterogeneous partitions of the feature space are identified through tree-based algorithms and post-hoc calibration techniques are applied to each partition to improve AUC. While the theoretical optimality of this framework holds for any model, we focus on deep neural networks (DNNs) and test the simplest instantiation of this paradigm on a variety of open-source datasets. Experiments demonstrate the effectiveness of this framework and the future potential for applying higher-performing partitioning schemes along with more effective calibration techniques.

Hamed Firooz, Maziar Sanjabi, Adrian Englhardt et al.

Ranking and recommendation systems are the foundation for numerous online experiences, ranging from search results to personalized content delivery. These systems have evolved into complex, multilayered architectures that leverage vast datasets and often incorporate thousands of predictive models. The maintenance and enhancement of these models is a labor intensive process that requires extensive feature engineering. This approach not only exacerbates technical debt but also hampers innovation in extending these systems to emerging problem domains. In this report, we present our research to address these challenges by utilizing a large foundation model with a textual interface for ranking and recommendation tasks. We illustrate several key advantages of our approach: (1) a single model can manage multiple predictive tasks involved in ranking and recommendation, (2) decoder models with textual interface due to their comprehension of reasoning capabilities, can generalize to new recommendation surfaces and out-of-domain problems, and (3) by employing natural language interfaces for task definitions and verbalizing member behaviors and their social connections, we eliminate the need for feature engineering and the maintenance of complex directed acyclic graphs of model dependencies. We introduce our research pre-production model, 360Brew V1.0, a 150B parameter, decoder-only model that has been trained and fine-tuned on LinkedIn's data and tasks. This model is capable of solving over 30 predictive tasks across various segments of the LinkedIn platform, achieving performance levels comparable to or exceeding those of current production systems based on offline metrics, without task-specific fine-tuning. Notably, each of these tasks is conventionally addressed by dedicated models that have been developed and maintained over multiple years by teams of a similar or larger size than our own.

Aman Gupta, Shao Tang, Qingquan Song et al.

Reinforcement Learning with Human Feedback (RLHF) and its variants have made huge strides toward the effective alignment of large language models (LLMs) to follow instructions and reflect human values. More recently, Direct Alignment Algorithms (DAAs) have emerged in which the reward modeling stage of RLHF is skipped by characterizing the reward directly as a function of the policy being learned. Some popular examples of DAAs include Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO). These methods often suffer from likelihood displacement, a phenomenon by which the probabilities of preferred responses are often reduced undesirably. In this paper, we argue that, for DAAs the reward (function) shape matters. We introduce \textbf{AlphaPO}, a new DAA method that leverages an $α$-parameter to help change the shape of the reward function beyond the standard log reward. AlphaPO helps maintain fine-grained control over likelihood displacement and over-optimization. Compared to SimPO, one of the best performing DAAs, AlphaPO leads to about 7\% to 10\% relative improvement in alignment performance for the instruct versions of Mistral-7B and Llama3-8B while achieving 15\% to 50\% relative improvement over DPO on the same models. The analysis and results presented highlight the importance of the reward shape and how one can systematically change it to affect training dynamics, as well as improve alignment performance.

Fedor Borisyuk, Mingzhou Zhou, Qingquan Song et al.

We present LiRank, a large-scale ranking framework at LinkedIn that brings to production state-of-the-art modeling architectures and optimization methods. We unveil several modeling improvements, including Residual DCN, which adds attention and residual connections to the famous DCNv2 architecture. We share insights into combining and tuning SOTA architectures to create a unified model, including Dense Gating, Transformers and Residual DCN. We also propose novel techniques for calibration and describe how we productionalized deep learning based explore/exploit methods. To enable effective, production-grade serving of large ranking models, we detail how to train and compress models using quantization and vocabulary compression. We provide details about the deployment setup for large-scale use cases of Feed ranking, Jobs Recommendations, and Ads click-through rate (CTR) prediction. We summarize our learnings from various A/B tests by elucidating the most effective technical approaches. These ideas have contributed to relative metrics improvements across the board at LinkedIn: +0.5% member sessions in the Feed, +1.76% qualified job applications for Jobs search and recommendations, and +4.3% for Ads CTR. We hope this work can provide practical insights and solutions for practitioners interested in leveraging large-scale deep ranking systems.

Zirui Liu, Qingquan Song, Qiang Charles Xiao et al.

The large number of parameters in Pretrained Language Models enhance their performance, but also make them resource-intensive, making it challenging to deploy them on commodity hardware like a single GPU. Due to the memory and power limitations of these devices, model compression techniques are often used to decrease both the model's size and its inference latency. This usually results in a trade-off between model accuracy and efficiency. Therefore, optimizing this balance is essential for effectively deploying LLMs on commodity hardware. A significant portion of the efficiency challenge is the Feed-forward network (FFN) component, which accounts for roughly $\frac{2}{3}$ total parameters and inference latency. In this paper, we first observe that only a few neurons of FFN module have large output norm for any input tokens, a.k.a. heavy hitters, while the others are sparsely triggered by different tokens. Based on this observation, we explicitly split the FFN into two parts according to the heavy hitters. We improve the efficiency-accuracy trade-off of existing compression methods by allocating more resource to FFN parts with heavy hitters. In practice, our method can reduce model size by 43.1\% and bring $1.25\sim1.56\times$ wall clock time speedup on different hardware with negligible accuracy drop.

Aman Gupta, Anirudh Ravichandran, Ziji Zhang et al.

Task-oriented dialogue systems are essential for applications ranging from customer service to personal assistants and are widely used across various industries. However, developing effective multi-domain systems remains a significant challenge due to the complexity of handling diverse user intents, entity types, and domain-specific knowledge across several domains. In this work, we propose DARD (Domain Assigned Response Delegation), a multi-agent conversational system capable of successfully handling multi-domain dialogs. DARD leverages domain-specific agents, orchestrated by a central dialog manager agent. Our extensive experiments compare and utilize various agent modeling approaches, combining the strengths of smaller fine-tuned models (Flan-T5-large & Mistral-7B) with their larger counterparts, Large Language Models (LLMs) (Claude Sonnet 3.0). We provide insights into the strengths and limitations of each approach, highlighting the benefits of our multi-agent framework in terms of flexibility and composability. We evaluate DARD using the well-established MultiWOZ benchmark, achieving state-of-the-art performance by improving the dialogue inform rate by 6.6% and the success rate by 4.1% over the best-performing existing approaches. Additionally, we discuss various annotator discrepancies and issues within the MultiWOZ dataset and its evaluation system.

Lars Hertel, Neil Daftary, Fedor Borisyuk et al.

We study user history modeling via Transformer encoders in deep learning recommendation models (DLRM). Such architectures can significantly improve recommendation quality, but usually incur high latency cost necessitating infrastructure upgrades or very small Transformer models. An important part of user history modeling is early fusion of the candidate item and various methods have been studied. We revisit early fusion and compare concatenation of the candidate to each history item against appending it to the end of the list as a separate item. Using the latter method, allows us to reformulate the recently proposed amortized history inference algorithm M-FALCON \cite{zhai2024actions} for the case of DLRM models. We show via experimental results that appending with cross-attention performs on par with concatenation and that amortization significantly reduces inference costs. We conclude with results from deploying this model on the LinkedIn Feed and Ads surfaces, where amortization reduces latency by 30\% compared to non-amortized inference.

Changshuai Wei, Benjamin Zelditch, Joyce Chen et al.

Computational marketing has become increasingly important in today's digital world, facing challenges such as massive heterogeneous data, multi-channel customer journeys, and limited marketing budgets. In this paper, we propose a general framework for marketing AI systems, the Neural Optimization with Adaptive Heuristics (NOAH) framework. NOAH is the first general framework for marketing optimization that considers both to-business (2B) and to-consumer (2C) products, as well as both owned and paid channels. We describe key modules of the NOAH framework, including prediction, optimization, and adaptive heuristics, providing examples for bidding and content optimization. We then detail the successful application of NOAH to LinkedIn's email marketing system, showcasing significant wins over the legacy ranking system. Additionally, we share details and insights that are broadly useful, particularly on: (i) addressing delayed feedback with lifetime value, (ii) performing large-scale linear programming with randomization, (iii) improving retrieval with audience expansion, (iv) reducing signal dilution in targeting tests, and (v) handling zero-inflated heavy-tail metrics in statistical testing.

Yingying Zhuang, Aman Gupta, Anurag Beniwal

Multilingual information retrieval has emerged as powerful tools for expanding knowledge sharing across languages. On the other hand, resources on high quality knowledge base are often scarce and in limited languages, therefore an effective embedding model to transform sentences from different languages into a feature vector space same as the knowledge base language becomes the key ingredient for cross language knowledge sharing, especially to transfer knowledge available in high-resource languages to low-resource ones. In this paper we propose a novel strategy to fine-tune multilingual embedding models with weighted sampling for contrastive learning, enabling multilingual information retrieval with a monolingual knowledge base. We demonstrate that the weighted sampling strategy produces performance gains compared to standard ones by up to 31.03\% in MRR and up to 33.98\% in Recall@3. Additionally, our proposed methodology is language agnostic and applicable for both multilingual and code switching use cases.

Ruofan Wang, Prakruthi Prabhakar, Gaurav Srivastava et al.

In the realm of recommender systems, the ubiquitous adoption of deep neural networks has emerged as a dominant paradigm for modeling diverse business objectives. As user bases continue to expand, the necessity of personalization and frequent model updates have assumed paramount significance to ensure the delivery of relevant and refreshed experiences to a diverse array of members. In this work, we introduce an innovative meta-learning solution tailored to the personalization of models for individual members and other entities, coupled with the frequent updates based on the latest user interaction signals. Specifically, we leverage the Model-Agnostic Meta Learning (MAML) algorithm to adapt per-task sub-networks using recent user interaction data. Given the near infeasibility of productionizing original MAML-based models in online recommendation systems, we propose an efficient strategy to operationalize meta-learned sub-networks in production, which involves transforming them into fixed-sized vectors, termed meta embeddings, thereby enabling the seamless deployment of models with hundreds of billions of parameters for online serving. Through extensive experimentation on production data drawn from various applications at LinkedIn, we demonstrate that the proposed solution consistently outperforms the baseline models of those applications, including strong baselines such as using wide-and-deep ID based personalization approach. Our approach has enabled the deployment of a range of highly personalized AI models across diverse LinkedIn applications, leading to substantial improvements in business metrics as well as refreshed experience for our members.

Qiang Charles Xiao, Ajith Muralidharan, Birjodh Tiwana et al.

In this paper, we propose a generic model-based re-ranking framework, MultiSlot ReRanker, which simultaneously optimizes relevance, diversity, and freshness. Specifically, our Sequential Greedy Algorithm (SGA) is efficient enough (linear time complexity) for large-scale production recommendation engines. It achieved a lift of $+6\%$ to $ +10\%$ offline Area Under the receiver operating characteristic Curve (AUC) which is mainly due to explicitly modeling mutual influences among items of a list, and leveraging the second pass ranking scores of multiple objectives. In addition, we have generalized the offline replay theory to multi-slot re-ranking scenarios, with trade-offs among multiple objectives. The offline replay results can be further improved by Pareto Optimality. Moreover, we've built a multi-slot re-ranking simulator based on OpenAI Gym integrated with the Ray framework. It can be easily configured for different assumptions to quickly benchmark both reinforcement learning and supervised learning algorithms.

Aman Gupta, Denny O'Shea, Fazl Barez

Large language models (LLMs) are increasingly used for tasks where outputs shape human decisions, so it is critical to test whether their responses reflect consistent human values. Existing benchmarks mostly track refusals or predefined safety violations, but these only check rule compliance and do not reveal whether a model upholds a coherent value system when facing controversial real-world issues. We introduce the Value ALignment Benchmark (VAL-Bench), which evaluates whether models maintain a stable value stance across paired prompts that frame opposing sides of public debates. VAL-Bench consists of 115K such pairs from Wikipedia's controversial sections. A well-aligned model should express similar underlying views regardless of framing, which we measure using an LLM-as-judge to score agreement or divergence between paired responses. Applied across leading open- and closed-source models, the benchmark reveals large variation in alignment and highlights trade-offs between safety strategies (e.g., refusals) and more expressive value systems. By providing a scalable, reproducible benchmark, VAL-Bench enables systematic comparison of how reliably LLMs embody human values.

Aman Gupta, Rafael Celente, Abhishek Shivanna et al.

The Muon optimizer, based on matrix orthogonalization, has recently shown faster convergence and up to 2x computational efficiency over AdamW in LLM pretraining. Like AdamW, Muon is stateful, requiring storage of both model weights and accumulated gradients. While 8-bit AdamW variants mitigate this overhead using blockwise quantization, they are typically stable only under dynamic quantization - which improves stability on linear quantization for extreme values. In this paper, we introduce the 8-bit Muon optimizer using blockwise quantization, supporting both linear and dynamic schemes. We demonstrate that 8-bit Muon maintains stability under both, while delivering $\sim$74\% reduction in memory footprint compared to full-precision Muon. In extensive experiments, 8-bit Muon closely matches the performance of Muon while outperforming AdamW and 8-bit AdamW in pre-training a 1.6B model on 4B FineWeb tokens. It also shows competitive results when fine-tuning the Llama 3.2 3B model on post-training data. We also provide a theoretical perspective to help explain this robustness under quantization.

Aman Gupta, Aditi Sheshadri, Sujit Roy et al.

Global climate models parameterize a range of atmospheric-oceanic processes like gravity waves, clouds, moist convection, and turbulence that cannot be sufficiently resolved. These subgrid-scale closures for unresolved processes are a leading source of model uncertainty. Here, we present a new approach to developing machine learning parameterizations of small-scale climate processes by fine-tuning a pre-trained AI foundation model (FM). FMs are largely unexplored in climate research. A pre-trained encoder-decoder from a 2.3 billion parameter FM (NASA and IBM Research's Prithvi WxC) -- which contains a latent probabilistic representation of atmospheric evolution -- is fine-tuned (or reused) to create a deep learning parameterization for atmospheric gravity waves (GWs). The parameterization captures GW effects for a coarse-resolution climate model by learning the fluxes from an atmospheric reanalysis with 10 times finer resolution. A comparison of monthly averages and instantaneous evolution with a machine learning model baseline (an Attention U-Net) reveals superior predictive performance of the FM parameterization throughout the atmosphere, even in regions excluded from pre-training. This performance boost is quantified using the Hellinger distance, which is 0.11 for the baseline and 0.06 for the fine-tuned model. Our findings emphasize the versatility and reusability of FMs, which could be used to accomplish a range of atmosphere- and climate-related applications, leading the way for the creation of observations-driven and physically accurate parameterizations for more earth-system processes.

Aman Gupta, Aditi Sheshadri, Dhruv Suri

Accurate weather forecasts are critical for societal planning and disaster preparedness. Yet these forecasts remain challenging to produce and evaluate, especially in regions with sparse observational coverage. Current evaluation of artificial intelligence (AI) weather prediction relies primarily on reanalyses, which can obscure important deficiencies. Here we present MAUSAM (Measuring AI Uncertainty during South Asian Monsoon), an evaluation of seven leading AI-based forecasting systems - FourCastNet, FourCastNet-SFNO, Pangu-Weather, GraphCast, Aurora, AIFS, and GenCast - during the South Asian Monsoon, using ground-based weather stations, rain gauge networks, and geostationary satellite imagery. The AI models demonstrate impressive forecast skill during monsoon across a broad range of variables, ranging from large-scale surface temperature and winds to precipitation, cloud cover, and subseasonal to seasonal eddy statistics, highlighting the strength of data-driven weather prediction. However, the models still exhibit systematic errors at finer scales like the underprediction of extreme precipitation, divergent cyclone tracks, and the mesoscale kinetic energy spectra, highlighting avenues for future improvement. A comparison against observations reveals forecast errors up to 15-45% larger than those relative to reanalysis and traditional forecasts, indicating that reanalysis-centric benchmarks can overstate forecast skill. Of the models assessed, AIFS achieves the most consistent representation of atmospheric variables, with GraphCast and GenCast also showing strong skill. The analysis presents a framework for evaluating AI weather models on regional prediction and highlights both the promise and current limitations of AI weather prediction in data-sparse regions, underscoring the importance of observational evaluation for future operational adoption.

Aman Gupta, Yingying Zhuang, Zhou Yu et al.

Despite advances in the multilingual capabilities of Large Language Models (LLMs), their performance varies substantially across different languages and tasks. In multilingual retrieval-augmented generation (RAG)-based systems, knowledge bases (KB) are often shared from high-resource languages (such as English) to low-resource ones, resulting in retrieved information from the KB being in a different language than the rest of the context. In such scenarios, two common practices are pre-translation to create a mono-lingual prompt and cross-lingual prompting for direct inference. However, the impact of these choices remains unclear. In this paper, we systematically evaluate the impact of different prompt translation strategies for classification tasks with RAG-enhanced LLMs in multilingual systems. Experimental results show that an optimized prompting strategy can significantly improve knowledge sharing across languages, therefore improve the performance on the downstream classification task. The findings advocate for a broader utilization of multilingual resource sharing and cross-lingual prompt optimization for non-English languages, especially the low-resource ones.

Kayhan Behdin, Ata Fatahibaarzi, Qingquan Song et al.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of industrial applications, from search and recommendation systems to generative tasks. Although scaling laws indicate that larger models generally yield better generalization and performance, their substantial computational requirements often render them impractical for many real-world scenarios at scale. In this paper, we present a comprehensive set of insights for training and deploying small language models (SLMs) that deliver high performance for a variety of industry use cases. We focus on two key techniques: (1) knowledge distillation and (2) model compression via structured pruning and quantization. These approaches enable SLMs to retain much of the quality of their larger counterparts while significantly reducing training/serving costs and latency. We detail the impact of these techniques on a variety of use cases in a large professional social network platform and share deployment lessons, including hardware optimization strategies that improve speed and throughput for both predictive and reasoning-based applications in Recommendation Systems.

Fedor Borisyuk, Lars Hertel, Ganesh Parameswaran et al.

We present LiGR, a large-scale ranking framework developed at LinkedIn that brings state-of-the-art transformer-based modeling architectures into production. We introduce a modified transformer architecture that incorporates learned normalization and simultaneous set-wise attention to user history and ranked items. This architecture enables several breakthrough achievements, including: (1) the deprecation of most manually designed feature engineering, outperforming the prior state-of-the-art system using only few features (compared to hundreds in the baseline), (2) validation of the scaling law for ranking systems, showing improved performance with larger models, more training data, and longer context sequences, and (3) simultaneous joint scoring of items in a set-wise manner, leading to automated improvements in diversity. To enable efficient serving of large ranking models, we describe techniques to scale inference effectively using single-pass processing of user history and set-wise attention. We also summarize key insights from various ablation studies and A/B tests, highlighting the most impactful technical approaches.

Aman Gupta, Aditi Sheshadri, Sujit Roy et al.

Global climate models typically operate at a grid resolution of hundreds of kilometers and fail to resolve atmospheric mesoscale processes, e.g., clouds, precipitation, and gravity waves (GWs). Model representation of these processes and their sources is essential to the global circulation and planetary energy budget, but subgrid scale contributions from these processes are often only approximately represented in models using parameterizations. These parameterizations are subject to approximations and idealizations, which limit their capability and accuracy. The most drastic of these approximations is the "single-column approximation" which completely neglects the horizontal evolution of these processes, resulting in key biases in current climate models. With a focus on atmospheric GWs, we present the first-ever global simulation of atmospheric GW fluxes using machine learning (ML) models trained on the WINDSET dataset to emulate global GW emulation in the atmosphere, as an alternative to traditional single-column parameterizations. Using an Attention U-Net-based architecture trained on globally resolved GW momentum fluxes, we illustrate the importance and effectiveness of global nonlocality, when simulating GWs using data-driven schemes.

Gregory Dexter, Borja Ocejo, Sathiya Keerthi et al.

Stochastic Gradient Descent (SGD) stands as a cornerstone optimization algorithm with proven real-world empirical successes but relatively limited theoretical understanding. Recent research has illuminated a key factor contributing to its practical efficacy: the implicit regularization it instigates. Several studies have investigated the linear stability property of SGD in the vicinity of a stationary point as a predictive proxy for sharpness and generalization error in overparameterized neural networks (Wu et al., 2022; Jastrzebski et al., 2019; Cohen et al., 2021). In this paper, we delve deeper into the relationship between linear stability and sharpness. More specifically, we meticulously delineate the necessary and sufficient conditions for linear stability, contingent on hyperparameters of SGD and the sharpness at the optimum. Towards this end, we introduce a novel coherence measure of the loss Hessian that encapsulates pertinent geometric properties of the loss function that are relevant to the linear stability of SGD. It enables us to provide a simplified sufficient condition for identifying linear instability at an optimum. Notably, compared to previous works, our analysis relies on significantly milder assumptions and is applicable for a broader class of loss functions than known before, encompassing not only mean-squared error but also cross-entropy loss.

Aman Gupta, Rohan Ramanath, Jun Shi et al.

Over-parameterized deep networks trained using gradient-based optimizers are a popular choice for solving classification and ranking problems. Without appropriately tuned $\ell_2$ regularization or weight decay, such networks have the tendency to make output scores (logits) and network weights large, causing training loss to become too small and the network to lose its adaptivity (ability to move around) in the parameter space. Although regularization is typically understood from an overfitting perspective, we highlight its role in making the network more adaptive and enabling it to escape more easily from weights that generalize poorly. To provide such a capability, we propose a method called Logit Attenuating Weight Normalization (LAWN), that can be stacked onto any gradient-based optimizer. LAWN controls the logits by constraining the weight norms of layers in the final homogeneous sub-network. Empirically, we show that the resulting LAWN variant of the optimizer makes a deep network more adaptive to finding minimas with superior generalization performance on large-scale image classification and recommender systems. While LAWN is particularly impressive in improving Adam, it greatly improves all optimizers when used with large batch sizes

Aman Gupta, Deepak Bhatt, Anubha Pandey

With the advent of generative modeling techniques, synthetic data and its use has penetrated across various domains from unstructured data such as image, text to structured dataset modeling healthcare outcome, risk decisioning in financial domain, and many more. It overcomes various challenges such as limited training data, class imbalance, restricted access to dataset owing to privacy issues. To ensure the trained model used for automated decisioning purposes makes a fair decision there exist prior work to quantify and mitigate those issues. This study aims to establish a trade-off between bias and fairness in the models trained using synthetic data. Variants of synthetic data generation techniques were studied to understand bias amplification including differentially private generation schemes. Through experiments on a tabular dataset, we demonstrate there exist a varying levels of bias impact on models trained using synthetic data. Techniques generating less correlated feature performs well as evident through fairness metrics with 94\%, 82\%, and 88\% relative drop in DPD (demographic parity difference), EoD (equality of odds) and EoP (equality of opportunity) respectively, and 24\% relative improvement in DRP (demographic parity ratio) with respect to the real dataset. We believe the outcome of our research study will help data science practitioners understand the bias in the use of synthetic data.

Sirjan Kafle, Aman Gupta, Xue Xia et al.

Cluster-and-aggregate techniques such as Vector of Locally Aggregated Descriptors (VLAD), and their end-to-end discriminatively trained equivalents like NetVLAD have recently been popular for video classification and action recognition tasks. These techniques operate by assigning video frames to clusters and then representing the video by aggregating residuals of frames with respect to the mean of each cluster. Since some clusters may see very little video-specific data, these features can be noisy. In this paper, we propose a new cluster-and-aggregate method which we call smoothed Gaussian mixture model (SGMM), and its end-to-end discriminatively trained equivalent, which we call deep smoothed Gaussian mixture model (DSGMM). SGMM represents each video by the parameters of a Gaussian mixture model (GMM) trained for that video. Low-count clusters are addressed by smoothing the video-specific estimates with a universal background model (UBM) trained on a large number of videos. The primary benefit of SGMM over VLAD is smoothing which makes it less sensitive to small number of training samples. We show, through extensive experiments on the YouTube-8M classification task, that SGMM/DSGMM is consistently better than VLAD/NetVLAD by a small but statistically significant margin. We also show results using a dataset created at LinkedIn to predict if a member will watch an uploaded video.