George Karypis

Zhuosheng Zhang, Aston Zhang, Mu Li et al.

Large language models (LLMs) have shown impressive performance on complex reasoning by leveraging chain-of-thought (CoT) prompting to generate intermediate reasoning chains as the rationale to infer the answer. However, existing CoT studies have primarily focused on the language modality. We propose Multimodal-CoT that incorporates language (text) and vision (images) modalities into a two-stage framework that separates rationale generation and answer inference. In this way, answer inference can leverage better generated rationales that are based on multimodal information. Experimental results on ScienceQA and A-OKVQA benchmark datasets show the effectiveness of our proposed approach. With Multimodal-CoT, our model under 1 billion parameters achieves state-of-the-art performance on the ScienceQA benchmark. Our analysis indicates that Multimodal-CoT offers the advantages of mitigating hallucination and enhancing convergence speed. Code is publicly available at https://github.com/amazon-science/mm-cot.

Hongkuan Zhou, Da Zheng, Israt Nisa et al.

Many real world graphs contain time domain information. Temporal Graph Neural Networks capture temporal information as well as structural and contextual information in the generated dynamic node embeddings. Researchers have shown that these embeddings achieve state-of-the-art performance in many different tasks. In this work, we propose TGL, a unified framework for large-scale offline Temporal Graph Neural Network training where users can compose various Temporal Graph Neural Networks with simple configuration files. TGL comprises five main components, a temporal sampler, a mailbox, a node memory module, a memory updater, and a message passing engine. We design a Temporal-CSR data structure and a parallel sampler to efficiently sample temporal neighbors to formtraining mini-batches. We propose a novel random chunk scheduling technique that mitigates the problem of obsolete node memory when training with a large batch size. To address the limitations of current TGNNs only being evaluated on small-scale datasets, we introduce two large-scale real-world datasets with 0.2 and 1.3 billion temporal edges. We evaluate the performance of TGL on four small-scale datasets with a single GPU and the two large datasets with multiple GPUs for both link prediction and node classification tasks. We compare TGL with the open-sourced code of five methods and show that TGL achieves similar or better accuracy with an average of 13x speedup. Our temporal parallel sampler achieves an average of 173x speedup on a multi-core CPU compared with the baselines. On a 4-GPU machine, TGL can train one epoch of more than one billion temporal edges within 1-10 hours. To the best of our knowledge, this is the first work that proposes a general framework for large-scale Temporal Graph Neural Networks training on multiple GPUs.

Zhiqi Bu, Yu-Xiang Wang, Sheng Zha et al.

Differentially private (DP) optimization is the standard paradigm to learn large neural networks that are accurate and privacy-preserving. The computational cost for DP deep learning, however, is notoriously heavy due to the per-sample gradient clipping. Existing DP implementations are 2-1000X more costly in time and space complexity than the standard (non-private) training. In this work, we develop a novel Book-Keeping (BK) technique that implements existing DP optimizers (thus achieving the same accuracy), with a substantial improvement on the computational cost. Specifically, BK enables DP training on large models and high dimensional data to be roughly as fast and memory-saving as the standard training, whereas previous DP algorithms can be inefficient or incapable of training due to memory error. The computational advantage of BK is supported by the complexity analysis as well as extensive experiments on vision and language tasks. Our implementation achieves state-of-the-art (SOTA) accuracy with very small extra cost: on GPT2 and at almost the same memory cost (<1% overhead), BK has 1.03X the time complexity of the standard training (0.83X training speed in practice), and 0.61X the time complexity of the most efficient DP implementation (1.36X training speed in practice). We open-source the codebase for the BK algorithm at the FastDP library (https://github.com/awslabs/fast-differential-privacy).

Zhiqi Bu, Yu-Xiang Wang, Sheng Zha et al.

We study the problem of differentially private (DP) fine-tuning of large pre-trained models -- a recent privacy-preserving approach suitable for solving downstream tasks with sensitive data. Existing work has demonstrated that high accuracy is possible under strong privacy constraint, yet requires significant computational overhead or modifications to the network architecture. We propose differentially private bias-term fine-tuning (DP-BiTFiT), which matches the state-of-the-art accuracy for DP algorithms and the efficiency of the standard BiTFiT. DP-BiTFiT is model agnostic (not modifying the network architecture), parameter efficient (only training about 0.1% of the parameters), and computation efficient (almost removing the overhead caused by DP, in both the time and space complexity). On a wide range of tasks, DP-BiTFiT is 2~30X faster and uses 2~8X less memory than DP full fine-tuning, even faster than the standard full fine-tuning. This amazing efficiency enables us to conduct DP fine-tuning on language and vision tasks with long-sequence texts and high-resolution images, which were computationally difficult using existing methods. We open-source our code at FastDP (https://github.com/awslabs/fast-differential-privacy).

Costas Mavromatis, George Karypis

Knowledge Graph Question Answering (KGQA) involves retrieving entities as answers from a Knowledge Graph (KG) using natural language queries. The challenge is to learn to reason over question-relevant KG facts that traverse KG entities and lead to the question answers. To facilitate reasoning, the question is decoded into instructions, which are dense question representations used to guide the KG traversals. However, if the derived instructions do not exactly match the underlying KG information, they may lead to reasoning under irrelevant context. Our method, termed ReaRev, introduces a new way to KGQA reasoning with respect to both instruction decoding and execution. To improve instruction decoding, we perform reasoning in an adaptive manner, where KG-aware information is used to iteratively update the initial instructions. To improve instruction execution, we emulate breadth-first search (BFS) with graph neural networks (GNNs). The BFS strategy treats the instructions as a set and allows our method to decide on their execution order on the fly. Experimental results on three KGQA benchmarks demonstrate the ReaRev's effectiveness compared with previous state-of-the-art, especially when the KG is incomplete or when we tackle complex questions. Our code is publicly available at https://github.com/cmavro/ReaRev_KGQA.

Jiani Zhang, Zhengyuan Shen, Balasubramaniam Srinivasan et al.

Recent advances in large language models have revolutionized many sectors, including the database industry. One common challenge when dealing with large volumes of tabular data is the pervasive use of abbreviated column names, which can negatively impact performance on various data search, access, and understanding tasks. To address this issue, we introduce a new task, called NameGuess, to expand column names (used in database schema) as a natural language generation problem. We create a training dataset of 384K abbreviated-expanded column pairs using a new data fabrication method and a human-annotated evaluation benchmark that includes 9.2K examples from real-world tables. To tackle the complexities associated with polysemy and ambiguity in NameGuess, we enhance auto-regressive language models by conditioning on table content and column header names -- yielding a fine-tuned model (with 2.7B parameters) that matches human performance. Furthermore, we conduct a comprehensive analysis (on multiple LLMs) to validate the effectiveness of table content in NameGuess and identify promising future opportunities. Code has been made available at https://github.com/amazon-science/nameguess.

Tuan Dinh, Jinman Zhao, Samson Tan et al. · amazon-science

Large language models of code (Code-LLMs) have recently brought tremendous advances to code completion, a fundamental feature of programming assistance and code intelligence. However, most existing works ignore the possible presence of bugs in the code context for generation, which are inevitable in software development. Therefore, we introduce and study the buggy-code completion problem, inspired by the realistic scenario of real-time code suggestion where the code context contains potential bugs -- anti-patterns that can become bugs in the completed program. To systematically study the task, we introduce two datasets: one with synthetic bugs derived from semantics-altering operator changes (buggy-HumanEval) and one with realistic bugs derived from user submissions to coding problems (buggy-FixEval). We find that the presence of potential bugs significantly degrades the generation performance of the high-performing Code-LLMs. For instance, the passing rates of CODEGEN-2B-MONO on test cases of buggy-HumanEval drop more than 50% given a single potential bug in the context. Finally, we investigate several post-hoc methods for mitigating the adverse effect of potential bugs and find that there remains a significant gap in post-mitigation performance.

Costas Mavromatis, Vassilis N. Ioannidis, Shen Wang et al. · amazon-science

How can we learn effective node representations on textual graphs? Graph Neural Networks (GNNs) that use Language Models (LMs) to encode textual information of graphs achieve state-of-the-art performance in many node classification tasks. Yet, combining GNNs with LMs has not been widely explored for practical deployments due to its scalability issues. In this work, we tackle this challenge by developing a Graph-Aware Distillation framework (GRAD) to encode graph structures into an LM for graph-free, fast inference. Different from conventional knowledge distillation, GRAD jointly optimizes a GNN teacher and a graph-free student over the graph's nodes via a shared LM. This encourages the graph-free student to exploit graph information encoded by the GNN teacher while at the same time, enables the GNN teacher to better leverage textual information from unlabeled nodes. As a result, the teacher and the student models learn from each other to improve their overall performance. Experiments in eight node classification benchmarks in both transductive and inductive settings showcase GRAD's superiority over existing distillation approaches for textual graphs.

Hengzhi Pei, Jinman Zhao, Leonard Lausen et al. · amazon-science

Pretrained code language models have enabled great progress towards program synthesis. However, common approaches only consider in-file local context and thus miss information and constraints imposed by other parts of the codebase and its external dependencies. Existing code completion benchmarks also lack such context. To resolve these restrictions we curate a new dataset of permissively licensed Python packages that includes full projects and their dependencies and provide tools to extract non-local information with the help of program analyzers. We then focus on the task of function call argument completion which requires predicting the arguments to function calls. We show that existing code completion models do not yield good results on our completion task. To better solve this task, we query a program analyzer for information relevant to a given function call, and consider ways to provide the analyzer results to different code completion models during inference and training. Our experiments show that providing access to the function implementation and function usages greatly improves the argument completion performance. Our ablation study provides further insights on how different types of information available from the program analyzer and different ways of incorporating the information affect the model performance.

Danilo Ribeiro, Shen Wang, Xiaofei Ma et al. · amazon-science

We introduce STREET, a unified multi-task and multi-domain natural language reasoning and explanation benchmark. Unlike most existing question-answering (QA) datasets, we expect models to not only answer questions, but also produce step-by-step structured explanations describing how premises in the question are used to produce intermediate conclusions that can prove the correctness of a certain answer. We perform extensive evaluation with popular language models such as few-shot prompting GPT-3 and fine-tuned T5. We find that these models still lag behind human performance when producing such structured reasoning steps. We believe this work will provide a way for the community to better train and test systems on multi-step reasoning and explanations in natural language.

Yulun Wu, Robert A. Barton, Zichen Wang et al. · amazon-science, harvard

Predicting the responses of a cell under perturbations may bring important benefits to drug discovery and personalized therapeutics. In this work, we propose a novel graph variational Bayesian causal inference framework to predict a cell's gene expressions under counterfactual perturbations (perturbations that this cell did not factually receive), leveraging information representing biological knowledge in the form of gene regulatory networks (GRNs) to aid individualized cellular response predictions. Aiming at a data-adaptive GRN, we also developed an adjacency matrix updating technique for graph convolutional networks and used it to refine GRNs during pre-training, which generated more insights on gene relations and enhanced model performance. Additionally, we propose a robust estimator within our framework for the asymptotically efficient estimation of marginal perturbation effect, which is yet to be carried out in previous works. With extensive experiments, we exhibited the advantage of our approach over state-of-the-art deep learning models for individual response prediction.

Vassilis N. Ioannidis, Xiang Song, Da Zheng et al. · amazon-science

Can we combine heterogenous graph structure with text to learn high-quality semantic and behavioural representations? Graph neural networks (GNN)s encode numerical node attributes and graph structure to achieve impressive performance in a variety of supervised learning tasks. Current GNN approaches are challenged by textual features, which typically need to be encoded to a numerical vector before provided to the GNN that may incur some information loss. In this paper, we put forth an efficient and effective framework termed language model GNN (LM-GNN) to jointly train large-scale language models and graph neural networks. The effectiveness in our framework is achieved by applying stage-wise fine-tuning of the BERT model first with heterogenous graph information and then with a GNN model. Several system and design optimizations are proposed to enable scalable and efficient training. LM-GNN accommodates node and edge classification as well as link prediction tasks. We evaluate the LM-GNN framework in different datasets performance and showcase the effectiveness of the proposed approach. LM-GNN provides competitive results in an Amazon query-purchase-product application.

Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan et al.

Recent advances in tabular data generation have greatly enhanced synthetic data quality. However, extending diffusion models to tabular data is challenging due to the intricately varied distributions and a blend of data types of tabular data. This paper introduces Tabsyn, a methodology that synthesizes tabular data by leveraging a diffusion model within a variational autoencoder (VAE) crafted latent space. The key advantages of the proposed Tabsyn include (1) Generality: the ability to handle a broad spectrum of data types by converting them into a single unified space and explicitly capture inter-column relations; (2) Quality: optimizing the distribution of latent embeddings to enhance the subsequent training of diffusion models, which helps generate high-quality synthetic data, (3) Speed: much fewer number of reverse steps and faster synthesis speed than existing diffusion-based methods. Extensive experiments on six datasets with five metrics demonstrate that Tabsyn outperforms existing methods. Specifically, it reduces the error rates by 86% and 67% for column-wise distribution and pair-wise column correlation estimations compared with the most competitive baselines.

Zhiqi Bu, Yu-Xiang Wang, Sheng Zha et al.

Per-example gradient clipping is a key algorithmic step that enables practical differential private (DP) training for deep learning models. The choice of clipping threshold R, however, is vital for achieving high accuracy under DP. We propose an easy-to-use replacement, called automatic clipping, that eliminates the need to tune R for any DP optimizers, including DP-SGD, DP-Adam, DP-LAMB and many others. The automatic variants are as private and computationally efficient as existing DP optimizers, but require no DP-specific hyperparameters and thus make DP training as amenable as the standard non-private training. We give a rigorous convergence analysis of automatic DP-SGD in the non-convex setting, showing that it can enjoy an asymptotic convergence rate that matches the standard SGD, under a symmetric gradient noise assumption of the per-sample gradients (commonly used in the non-DP literature). We demonstrate on various language and vision tasks that automatic clipping outperforms or matches the state-of-the-art, and can be easily employed with minimal changes to existing codebases.

Yongjun He, Shuai Zhang, Jiading Gai et al. · amazon-science

As large language models (LLMs) continue to scale and new GPUs are released even more frequently, there is an increasing demand for LLM post-training in heterogeneous environments to fully leverage underutilized mid-range or previous-generation GPUs and alleviate the shortage of homogeneous high-end GPUs within a single availability zone. However, achieving high-performance reinforcement learning (RL) training for LLMs on such computing resources remains challenging because the workflow involves multiple models and tasks with complex computation and data dependencies. In this paper, we present HetRL, a distributed system for efficient RL training in infrastructures with heterogeneous GPUs and networks. HetRL formulates the scheduling of RL training in heterogeneous environments as a constrained joint optimization problem and provides two complementary approaches for addressing this problem: (1) a hybrid scheduling algorithm that efficiently identifies near-optimal solutions, and (2) an integer linear programming (ILP)-based scheduling algorithm that obtains optimal solutions, enabling flexible trade-offs between solution optimality and efficiency. Our extensive evaluation, consuming 20,000 GPU-hours, shows that HetRL achieves up to 9.17x the throughput of state-of-the-art systems, and 3.17x on average, across a wide range of workloads and settings.

Jonas Kübler, Kailash Budhathoki, Matthäus Kleindessner et al.

Minimizing the inference cost and latency of foundation models has become a crucial area of research. Optimization approaches include theoretically lossless methods and others without accuracy guarantees like quantization. In all of these cases it is crucial to ensure that the model quality has not degraded. However, even at temperature zero, model generations are not necessarily robust even to theoretically lossless model optimizations due to numerical errors. We thus require statistical tools to decide whether a finite-sample accuracy deviation is an evidence of a model's degradation or whether it can be attributed to (harmless) noise in the evaluation. We propose a statistically sound hypothesis testing framework based on McNemar's test allowing to efficiently detect model degradations, while guaranteeing a controlled rate of false positives. The crucial insight is that we have to confront the model scores on each sample, rather than aggregated on the task level. Furthermore, we propose three approaches to aggregate accuracy estimates across multiple benchmarks into a single decision. We provide an implementation on top of the largely adopted open source LM Evaluation Harness and provide a case study illustrating that the method correctly flags degraded models, while not flagging model optimizations that are provably lossless. We find that with our tests even empirical accuracy degradations of 0.3% can be confidently attributed to actual degradations rather than noise.

Pei Chen, Soumajyoti Sarkar, Leonard Lausen et al.

Language models pretrained on large collections of tabular data have demonstrated their effectiveness in several downstream tasks. However, many of these models do not take into account the row/column permutation invariances, hierarchical structure, etc. that exist in tabular data. To alleviate these limitations, we propose HYTREL, a tabular language model, that captures the permutation invariances and three more structural properties of tabular data by using hypergraphs - where the table cells make up the nodes and the cells occurring jointly together in each row, column, and the entire table are used to form three different types of hyperedges. We show that HYTREL is maximally invariant under certain conditions for tabular data, i.e., two tables obtain the same representations via HYTREL iff the two tables are identical up to permutations. Our empirical results demonstrate that HYTREL consistently outperforms other competitive baselines on four downstream tasks with minimal pretraining, illustrating the advantages of incorporating the inductive biases associated with tabular data into the representations. Finally, our qualitative analyses showcase that HYTREL can assimilate the table structures to generate robust representations for the cells, rows, columns, and the entire table.

Jiacheng Li, Tong Zhao, Jin Li et al.

Sequential recommendation aims to model dynamic user behavior from historical interactions. Self-attentive methods have proven effective at capturing short-term dynamics and long-term preferences. Despite their success, these approaches still struggle to model sparse data, on which they struggle to learn high-quality item representations. We propose to model user dynamics from shopping intents and interacted items simultaneously. The learned intents are coarse-grained and work as prior knowledge for item recommendation. To this end, we present a coarse-to-fine self-attention framework, namely CaFe, which explicitly learns coarse-grained and fine-grained sequential dynamics. Specifically, CaFe first learns intents from coarse-grained sequences which are dense and hence provide high-quality user intent representations. Then, CaFe fuses intent representations into item encoder outputs to obtain improved item representations. Finally, we infer recommended items based on representations of items and corresponding intents. Experiments on sparse datasets show that CaFe outperforms state-of-the-art self-attentive recommenders by 44.03% NDCG@5 on average.

Costas Mavromatis, Balasubramaniam Srinivasan, Zhengyuan Shen et al.

Large Language Models (LLMs) can adapt to new tasks via in-context learning (ICL). ICL is efficient as it does not require any parameter updates to the trained LLM, but only few annotated examples as input for the LLM. In this work, we investigate an active learning approach for ICL, where there is a limited budget for annotating examples. We propose a model-adaptive optimization-free algorithm, termed AdaICL, which identifies examples that the model is uncertain about, and performs semantic diversity-based example selection. Diversity-based sampling improves overall effectiveness, while uncertainty sampling improves budget efficiency and helps the LLM learn new information. Moreover, AdaICL poses its sampling strategy as a Maximum Coverage problem, that dynamically adapts based on the model's feedback and can be approximately solved via greedy algorithms. Extensive experiments on nine datasets and seven LLMs show that AdaICL improves performance by 4.4% accuracy points over SOTA (7.7% relative improvement), is up to 3x more budget-efficient than performing annotations uniformly at random, while it outperforms SOTA with 2x fewer ICL examples.

Hongkuan Zhou, Da Zheng, Xiang Song et al.

Memory-based Temporal Graph Neural Networks are powerful tools in dynamic graph representation learning and have demonstrated superior performance in many real-world applications. However, their node memory favors smaller batch sizes to capture more dependencies in graph events and needs to be maintained synchronously across all trainers. As a result, existing frameworks suffer from accuracy loss when scaling to multiple GPUs. Evenworse, the tremendous overhead to synchronize the node memory make it impractical to be deployed to distributed GPU clusters. In this work, we propose DistTGL -- an efficient and scalable solution to train memory-based TGNNs on distributed GPU clusters. DistTGL has three improvements over existing solutions: an enhanced TGNN model, a novel training algorithm, and an optimized system. In experiments, DistTGL achieves near-linear convergence speedup, outperforming state-of-the-art single-machine method by 14.5% in accuracy and 10.17x in training throughput.

Vishakh Padmakumar, Leonard Lausen, Miguel Ballesteros et al.

Recent work has found that multi-task training with a large number of diverse tasks can uniformly improve downstream performance on unseen target tasks. In contrast, literature on task transferability has established that the choice of intermediate tasks can heavily affect downstream task performance. In this work, we aim to disentangle the effect of scale and relatedness of tasks in multi-task representation learning. We find that, on average, increasing the scale of multi-task learning, in terms of the number of tasks, indeed results in better learned representations than smaller multi-task setups. However, if the target tasks are known ahead of time, then training on a smaller set of related tasks is competitive to the large-scale multi-task training at a reduced computational cost.

Chunxing Yin, Da Zheng, Israt Nisa et al.

This paper describes a new method for representing embedding tables of graph neural networks (GNNs) more compactly via tensor-train (TT) decomposition. We consider the scenario where (a) the graph data that lack node features, thereby requiring the learning of embeddings during training; and (b) we wish to exploit GPU platforms, where smaller tables are needed to reduce host-to-GPU communication even for large-memory GPUs. The use of TT enables a compact parameterization of the embedding, rendering it small enough to fit entirely on modern GPUs even for massive graphs. When combined with judicious schemes for initialization and hierarchical graph partitioning, this approach can reduce the size of node embedding vectors by 1,659 times to 81,362 times on large publicly available benchmark datasets, achieving comparable or better accuracy and significant speedups on multi-GPU systems. In some cases, our model without explicit node features on input can even match the accuracy of models that use node features.

Zeren Shui, Daniel S. Karls, Mingjian Wen et al.

For decades, atomistic modeling has played a crucial role in predicting the behavior of materials in numerous fields ranging from nanotechnology to drug discovery. The most accurate methods in this domain are rooted in first-principles quantum mechanical calculations such as density functional theory (DFT). Because these methods have remained computationally prohibitive, practitioners have traditionally focused on defining physically motivated closed-form expressions known as empirical interatomic potentials (EIPs) that approximately model the interactions between atoms in materials. In recent years, neural network (NN)-based potentials trained on quantum mechanical (DFT-labeled) data have emerged as a more accurate alternative to conventional EIPs. However, the generalizability of these models relies heavily on the amount of labeled training data, which is often still insufficient to generate models suitable for general-purpose applications. In this paper, we propose two generic strategies that take advantage of unlabeled training instances to inject domain knowledge from conventional EIPs to NNs in order to increase their generalizability. The first strategy, based on weakly supervised learning, trains an auxiliary classifier on EIPs and selects the best-performing EIP to generate energies to supplement the ground-truth DFT energies in training the NN. The second strategy, based on transfer learning, first pretrains the NN on a large set of easily obtainable EIP energies, and then fine-tunes it on ground-truth DFT energies. Experimental results on three benchmark datasets demonstrate that the first strategy improves baseline NN performance by 5% to 51% while the second improves baseline performance by up to 55%. Combining them further boosts performance.

Zhiqi Bu, Justin Chiu, Ruixuan Liu et al.

Deep learning using large models have achieved great success in a wide range of domains. However, training these models on billions of parameters is very challenging in terms of the training speed, memory cost, and communication efficiency, especially under the privacy-preserving regime with differential privacy (DP). On the one hand, DP optimization has comparable efficiency to the standard non-private optimization on a single GPU, but on multiple GPUs, existing DP distributed learning (such as pipeline parallel) has suffered from significantly worse efficiency. On the other hand, the Zero Redundancy Optimizer (ZeRO) is a state-of-the-art solution to the standard distributed learning, exhibiting excellent training efficiency on large models, but to work compatibly with DP is technically complicated. In this work, we develop a new systematic solution, DP-ZeRO, (I) to scale up the trainable DP model size, e.g. to GPT-100B, (II) to obtain the same computation and communication efficiency as the standard ZeRO, and (III) to enable mixed-precision DP training. Our DP-ZeRO, like the standard ZeRO, has the potential to train models with arbitrary size and is evaluated on the world's largest DP models in terms of the number of trainable parameters.

Yulun Wu, Layne C. Price, Zichen Wang et al.

Estimating an individual's potential outcomes under counterfactual treatments is a challenging task for traditional causal inference and supervised learning approaches when the outcome is high-dimensional (e.g. gene expressions, impulse responses, human faces) and covariates are relatively limited. In this case, to construct one's outcome under a counterfactual treatment, it is crucial to leverage individual information contained in its observed factual outcome on top of the covariates. We propose a deep variational Bayesian framework that rigorously integrates two main sources of information for outcome construction under a counterfactual treatment: one source is the individual features embedded in the high-dimensional factual outcome; the other source is the response distribution of similar subjects (subjects with the same covariates) that factually received this treatment of interest.

Zhiqi Bu, Ruixuan Liu, Yu-Xiang Wang et al.

Recent advances have substantially improved the accuracy, memory cost, and training speed of differentially private (DP) deep learning, especially on large vision and language models with millions to billions of parameters. In this work, we thoroughly study the per-sample gradient clipping style, a key component in DP optimization. We show that different clipping styles have the same time complexity but instantiate an accuracy-memory trade-off: while the all-layer clipping (of coarse granularity) is the most prevalent and usually gives the best accuracy, it incurs heavier memory cost compared to other group-wise clipping, such as the layer-wise clipping (of finer granularity). We formalize this trade-off through our convergence theory and complexity analysis. Importantly, we demonstrate that the accuracy gap between group-wise clipping and all-layer clipping becomes smaller for larger models, while the memory advantage of the group-wise clipping remains. Consequently, the group-wise clipping allows DP optimization of large models to achieve high accuracy and low peak memory simultaneously.

Trong Nghia Hoang, Anoop Deoras, Tong Zhao et al.

This paper studies the item-to-item recommendation problem in recommender systems from a new perspective of metric learning via implicit feedback. We develop and investigate a personalizable deep metric model that captures both the internal contents of items and how they were interacted with by users. There are two key challenges in learning such model. First, there is no explicit similarity annotation, which deviates from the assumption of most metric learning methods. Second, these approaches ignore the fact that items are often represented by multiple sources of meta data and different users use different combinations of these sources to form their own notion of similarity. To address these challenges, we develop a new metric representation embedded as kernel parameters of a probabilistic model. This helps express the correlation between items that a user has interacted with, which can be used to predict user interaction with new items. Our approach hinges on the intuition that similar items induce similar interactions from the same user, thus fitting a metric-parameterized model to predict an implicit feedback signal could indirectly guide it towards finding the most suitable metric for each user. To this end, we also analyze how and when the proposed method is effective from a theoretical lens. Its empirical effectiveness is also demonstrated on several real-world datasets.

Hengrui Zhang, Shen Wang, Vassilis N. Ioannidis et al.

Graph Neural Networks (GNNs) are currently dominating in modeling graph-structure data, while their high reliance on graph structure for inference significantly impedes them from widespread applications. By contrast, Graph-regularized MLPs (GR-MLPs) implicitly inject the graph structure information into model weights, while their performance can hardly match that of GNNs in most tasks. This motivates us to study the causes of the limited performance of GR-MLPs. In this paper, we first demonstrate that node embeddings learned from conventional GR-MLPs suffer from dimensional collapse, a phenomenon in which the largest a few eigenvalues dominate the embedding space, through empirical observations and theoretical analysis. As a result, the expressive power of the learned node representations is constrained. We further propose OrthoReg, a novel GR-MLP model to mitigate the dimensional collapse issue. Through a soft regularization loss on the correlation matrix of node embeddings, OrthoReg explicitly encourages orthogonal node representations and thus can naturally avoid dimensionally collapsed representations. Experiments on traditional transductive semi-supervised classification tasks and inductive node classification for cold-start scenarios demonstrate its effectiveness and superiority.

Chaoyang He, Shuai Zheng, Aston Zhang et al.

The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.

Youngsuk Park, Kailash Budhathoki, Liangfu Chen et al.

Powerful foundation models, including large language models (LLMs), with Transformer architectures have ushered in a new era of Generative AI across various industries. Industry and research community have witnessed a large number of new applications, based on those foundation models. Such applications include question and answer, customer services, image and video generation, and code completions, among others. However, as the number of model parameters reaches to hundreds of billions, their deployment incurs prohibitive inference costs and high latency in real-world scenarios. As a result, the demand for cost-effective and fast inference using AI accelerators is ever more higher. To this end, our tutorial offers a comprehensive discussion on complementary inference optimization techniques using AI accelerators. Beginning with an overview of basic Transformer architectures and deep learning system frameworks, we deep dive into system optimization techniques for fast and memory-efficient attention computations and discuss how they can be implemented efficiently on AI accelerators. Next, we describe architectural elements that are key for fast transformer inference. Finally, we examine various model compression and fast decoding strategies in the same context.

Ruixuan Liu, Zhiqi Bu, Yu-xiang Wang et al.

The success of large neural networks is crucially determined by the availability of data. It has been observed that training only on a small amount of public data, or privately on the abundant private data can lead to undesirable degradation of accuracy. In this work, we leverage both private and public data to improve the optimization, by coupling their gradients via a weighted linear combination. We formulate an optimal solution for the optimal weight in the convex setting to indicate that the weighting coefficient should be hyperparameter-dependent. Then, we prove the acceleration in the convergence of non-convex loss and the effects of hyper-parameters such as privacy budget, number of iterations, batch size, and model size on the choice of the weighting coefficient. We support our analysis with empirical experiments across language and vision benchmarks, and provide a guideline for choosing the optimal weight of the gradient coupling.

Jiefu Ou, Sapana Chaudhary, Kaj Bostrom et al. · amazon-science

Large Language Models (LLMs) demonstrate strong capabilities in general coding tasks but encounter two key challenges when optimizing code: (i) the complexity of writing optimized code (such as performant CUDA kernels and competition-level CPU code) requires expertise in systems, algorithms and specific languages and (ii) requires interpretation of performance metrics like timing and device utilization beyond binary correctness. In this work, we explore inference-time search algorithms that guide the LLM to discover better solutions through iterative refinement based on execution feedback. Our approach, called MaxCode unifies existing search methods under a max-reward reinforcement learning framework, making the observation and action-value functions modular for modification. To enhance the observation space, we integrate a natural language critique model that converts raw execution feedback into diagnostic insights about errors and performance bottlenecks, and the best-discounted reward seen so far. Together, these provide richer input to the code proposal function. To improve exploration during search, we train a generative reward-to-go model using action values from rollouts to rerank potential solutions. Testing on the KernelBench (CUDA) and PIE (C++) optimization benchmarks shows that MaxCode improves optimized code performance compared to baselines, achieving 20.3% and 10.1% relative improvements in absolute speedup value and relative speedup ranking, respectively.

Soumajyoti Sarkar, Leonard Lausen, Volkan Cevher et al.

Sparse Mixture of Expert (SMoE) models have emerged as a scalable alternative to dense models in language modeling. These models use conditionally activated feedforward subnetworks in transformer blocks, allowing for a separation between total model parameters and per-example computation. However, large token-routed SMoE models face a significant challenge: during inference, the entire model must be used for a sequence or a batch, resulting in high latencies in a distributed setting that offsets the advantages of per-token sparse activation. Our research explores task-specific model pruning to inform decisions about designing SMoE architectures, mainly modulating the choice of expert counts in pretraining. We investigate whether such pruned models offer advantages over smaller SMoE models trained from scratch, when evaluating and comparing them individually on tasks. To that end, we introduce an adaptive task-aware pruning technique UNCURL to reduce the number of experts per MoE layer in an offline manner post-training. Our findings reveal a threshold pruning factor for the reduction that depends on the number of experts used in pretraining, above which, the reduction starts to degrade model performance. These insights contribute to our understanding of model design choices when pretraining with SMoE architectures, particularly useful when considering task-specific inference optimization for later stages.

Zhenwei Dai, Vasileios Ioannidis, Soji Adeshina et al.

What target labels are most effective for graph neural network (GNN) training? In some applications where GNNs excel-like drug design or fraud detection, labeling new instances is expensive. We develop a data-efficient active sampling framework, ScatterSample, to train GNNs under an active learning setting. ScatterSample employs a sampling module termed DiverseUncertainty to collect instances with large uncertainty from different regions of the sample space for labeling. To ensure diversification of the selected nodes, DiverseUncertainty clusters the high uncertainty nodes and selects the representative nodes from each cluster. Our ScatterSample algorithm is further supported by rigorous theoretical analysis demonstrating its advantage compared to standard active sampling methods that aim to simply maximize the uncertainty and not diversify the samples. In particular, we show that ScatterSample is able to efficiently reduce the model uncertainty over the whole sample space. Our experiments on five datasets show that ScatterSample significantly outperforms the other GNN active learning baselines, specifically it reduces the sampling cost by up to 50% while achieving the same test accuracy.

Petros Karypis, Julian McAuley, George Karypis

Effectively training language models on long inputs poses many technical challenges. As a cost consideration, languages models are pretrained on a fixed sequence length before being adapted to longer sequences. We explore various methods for adapting models to longer inputs by training on segmented sequences and an interpolation-based method for extending absolute positional embeddings. We develop a training procedure to extend the input context size of pretrained models with no architectural changes and no additional memory costs than training on the original input lengths. By sub-sampling segments from long inputs while maintaining their original position the model is able to learn new positional interactions. Our method benefits both models trained with absolute positional embeddings, by extending their input contexts, as well as popular relative positional embedding methods showing a reduced perplexity on sequences longer than they were trained on. We demonstrate our method can extend input contexts by a factor of 4x while improving perplexity.

Yunyi Zhang, Soji Adeshina, Sheng Guan et al.

Prompt routing dynamically selects the most appropriate large language model from a pool of candidates for each query, optimizing performance while managing costs. As model pools scale to include dozens of frontier models with narrow performance gaps, existing approaches face significant challenges: manually defined task taxonomies cannot capture fine-grained capability distinctions, while monolithic routers struggle to differentiate subtle differences across diverse tasks. We propose a two-stage routing architecture that addresses these limitations through automated fine-grained task discovery and task-aware quality estimation. Our first stage employs graph-based clustering to discover latent task types and trains a classifier to assign prompts to discovered tasks. The second stage uses a mixture-of-experts architecture with task-specific prediction heads for specialized quality estimates. At inference, we aggregate predictions from both stages to balance task-level stability with prompt-specific adaptability. Evaluated on 10 benchmarks with 11 frontier models, our method consistently outperforms existing baselines and surpasses the strongest individual model while incurring less than half its cost.

Zhiqiang Tang, Haoyang Fang, Su Zhou et al.

AutoGluon-Multimodal (AutoMM) is introduced as an open-source AutoML library designed specifically for multimodal learning. Distinguished by its exceptional ease of use, AutoMM enables fine-tuning of foundation models with just three lines of code. Supporting various modalities including image, text, and tabular data, both independently and in combination, the library offers a comprehensive suite of functionalities spanning classification, regression, object detection, semantic matching, and image segmentation. Experiments across diverse datasets and tasks showcases AutoMM's superior performance in basic classification and regression tasks compared to existing AutoML tools, while also demonstrating competitive results in advanced tasks, aligning with specialized toolboxes designed for such purposes.

Philipp Hoellmer, Thomas Egg, Maya M. Martirossyan et al.

The discovery of new materials is essential for enabling technological advancements. Computational approaches for predicting novel materials must effectively learn the manifold of stable crystal structures within an infinite design space. We introduce Open Materials Generation (OMatG), a unifying framework for the generative design and discovery of inorganic crystalline materials. OMatG employs stochastic interpolants (SI) to bridge an arbitrary base distribution to the target distribution of inorganic crystals via a broad class of tunable stochastic processes, encompassing both diffusion models and flow matching as special cases. In this work, we adapt the SI framework by integrating an equivariant graph representation of crystal structures and extending it to account for periodic boundary conditions in unit cell representations. Additionally, we couple the SI flow over spatial coordinates and lattice vectors with discrete flow matching for atomic species. We benchmark OMatG's performance on two tasks: Crystal Structure Prediction (CSP) for specified compositions, and 'de novo' generation (DNG) aimed at discovering stable, novel, and unique structures. In our ground-up implementation of OMatG, we refine and extend both CSP and DNG metrics compared to previous works. OMatG establishes a new state of the art in generative modeling for materials discovery, outperforming purely flow-based and diffusion-based implementations. These results underscore the importance of designing flexible deep learning frameworks to accelerate progress in materials science. The OMatG code is available at https://github.com/FERMat-ML/OMatG.

Haoyang Fang, Shuai Zhang, Yifei Ma et al.

Domain-specific finetuning is essential for dense retrievers, yet not all training pairs contribute equally to the learning process. We introduce OPERA, a data pruning framework that exploits this heterogeneity to improve both the effectiveness and efficiency of retrieval model adaptation. We first investigate static pruning (SP), which retains only high-similarity query-document pairs, revealing an intrinsic quality-coverage tradeoff: ranking (NDCG) improves while retrieval (Recall) can degrade due to reduced query diversity. To resolve this tradeoff, we propose a two-stage dynamic pruning (DP) strategy that adaptively modulates sampling probabilities at both query and document levels throughout training, prioritizing high-quality examples while maintaining access to the full training set. Evaluations across eight datasets spanning six domains demonstrate the effectiveness of both approaches: SP improves ranking over standard finetuning (NDCG@10 +0.5\%), while DP achieves the strongest performance on both ranking (NDCG@10 +1.9\%) and retrieval (Recall@20 +0.7\%), with an average rank of 1.38 across all methods. These findings scale to Qwen3-Embedding, an LLM-based dense retriever, confirming architecture-agnostic benefits. Notably, DP reaches comparable performance in less than 50\% of the training time required by standard finetuning.

Daniil Vankov, Nikita Ivkin, Kyle Ulrich et al.

Mixture-of-Experts (MoE) architectures are increasingly used to efficiently scale large language models. However, in production inference, request batching and speculative decoding significantly amplify expert activation, eroding these efficiency benefits. We address this issue by modeling batch-aware expert selection as a modular optimization problem and designing efficient greedy algorithms for different deployment settings. The proposed method, namely XShare, requires no retraining and dynamically adapts to each batch by maximizing the total gating score of selected experts. It reduces expert activation by up to 30% under standard batching, cuts peak GPU load by up to 3x in expert-parallel deployments, and achieves up to 14% throughput gains in speculative decoding via hierarchical, correlation-aware expert selection even if requests in a batch drawn from heterogeneous datasets.

Jiading Gai, Shuai Zhang, Xiang Song et al.

Modern RL post-training methods such as GRPO and DAPO train on $N$ response sequences of $R$ tokens sampled from a shared prompt of $P$ tokens, but standard FlashAttention replicates all $P$ prompt tokens $N$ times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training ($N{\geq}16$, $P{\geq}8\text{K}$), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose \textbf{DualKV}, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1)~fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2)~a data-pipeline redesign in veRL that repacks $N(P{+}R)$ tokens into $P{+}NR$ tokens per micro-batch, extending the token reduction from attention to the entire model by a factor $ρ= N(P{+}R)/(P{+}NR)$. DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8$\times$H100 GPUs ($N{=}32$, 8K-context), DualKV achieves $1.63$--$2.09\times$ policy-update speedup, enables $2\times$ larger micro-batches, and raises MFU from $36\%$ to $76\%$. Similar gains hold for DAPO ($2.47\times$ speedup, $77\%$ MFU). At 30B MoE scale on 16$\times$H100, DualKV achieves $3.82\times$ policy-update and $3.38\times$ end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM).

Zhiqi Bu, Xinwei Zhang, Mingyi Hong et al.

The superior performance of large foundation models relies on the use of massive amounts of high-quality data, which often contain sensitive, private and copyrighted material that requires formal protection. While differential privacy (DP) is a prominent method to gauge the degree of security provided to the models, its application is commonly limited to the model fine-tuning stage, due to the performance degradation when applying DP during the pre-training stage. Consequently, DP is yet not capable of protecting a substantial portion of the data used during the initial pre-training process. In this work, we first provide a theoretical understanding of the efficacy of DP training by analyzing the per-iteration loss improvement. We make a key observation that DP optimizers' performance degradation can be significantly mitigated by the use of limited public data, which leads to a novel DP continual pre-training strategy. Empirically, using only 10\% of public data, our strategy can achieve DP accuracy of 41.5\% on ImageNet-21k (with $ε=8$), as well as non-DP accuracy of 55.7\% and and 60.0\% on downstream tasks Places365 and iNaturalist-2021, respectively, on par with state-of-the-art standard pre-training and substantially outperforming existing DP pre-trained models. Our DP pre-trained models are released in fastDP library (https://github.com/awslabs/fast-differential-privacy/releases/tag/v2.1)

Costas Mavromatis, Soji Adeshina, Vassilis N. Ioannidis et al.

Knowledge graph question answering (KGQA) presents significant challenges due to the structural and semantic variations across input graphs. Existing works rely on Large Language Model (LLM) agents for graph traversal and retrieval; an approach that is sensitive to traversal initialization, as it is prone to entity linking errors and may not generalize well to custom ("bring-your-own") KGs. We introduce BYOKG-RAG, a framework that enhances KGQA by synergistically combining LLMs with specialized graph retrieval tools. In BYOKG-RAG, LLMs generate critical graph artifacts (question entities, candidate answers, reasoning paths, and OpenCypher queries), and graph tools link these artifacts to the KG and retrieve relevant graph context. The retrieved context enables the LLM to iteratively refine its graph linking and retrieval, before final answer generation. By retrieving context from different graph tools, BYOKG-RAG offers a more general and robust solution for QA over custom KGs. Through experiments on five benchmarks spanning diverse KG types, we demonstrate that BYOKG-RAG outperforms the second-best graph retrieval method by 4.5% points while showing better generalization to custom KGs. BYOKG-RAG framework is open-sourced at https://github.com/awslabs/graphrag-toolkit.

Zhikai Chen, Han Xie, Jian Zhang et al.

Recent years have witnessed significant advancements in graph machine learning (GML), with its applications spanning numerous domains. However, the focus of GML has predominantly been on developing powerful models, often overlooking a crucial initial step: constructing suitable graphs from common data formats, such as tabular data. This construction process is fundamental to applying graph-based models, yet it remains largely understudied and lacks formalization. Our research aims to address this gap by formalizing the graph construction problem and proposing an effective solution. We identify two critical challenges to achieve this goal: 1. The absence of dedicated datasets to formalize and evaluate the effectiveness of graph construction methods, and 2. Existing automatic construction methods can only be applied to some specific cases, while tedious human engineering is required to generate high-quality graphs. To tackle these challenges, we present a two-fold contribution. First, we introduce a set of datasets to formalize and evaluate graph construction methods. Second, we propose an LLM-based solution, AutoG, automatically generating high-quality graph schemas without human intervention. The experimental results demonstrate that the quality of constructed graphs is critical to downstream task performance, and AutoG can generate high-quality graphs that rival those produced by human experts. Our code can be accessible from https://github.com/amazon-science/Automatic-Table-to-Graph-Generation.

Da Zheng, Xiang Song, Qi Zhu et al.

Graph machine learning (GML) is effective in many business applications. However, making GML easy to use and applicable to industry applications with massive datasets remain challenging. We developed GraphStorm, which provides an end-to-end solution for scalable graph construction, graph model training and inference. GraphStorm has the following desirable properties: (a) Easy to use: it can perform graph construction and model training and inference with just a single command; (b) Expert-friendly: GraphStorm contains many advanced GML modeling techniques to handle complex graph data and improve model performance; (c) Scalable: every component in GraphStorm can operate on graphs with billions of nodes and can scale model training and inference to different hardware without changing any code. GraphStorm has been used and deployed for over a dozen billion-scale industry applications after its release in May 2023. It is open-sourced in Github: https://github.com/awslabs/graphstorm.

Zonghan Wu, Da Zheng, Shirui Pan et al.

This paper aims to unify spatial dependency and temporal dependency in a non-Euclidean space while capturing the inner spatial-temporal dependencies for traffic data. For spatial-temporal attribute entities with topological structure, the space-time is consecutive and unified while each node's current status is influenced by its neighbors' past states over variant periods of each neighbor. Most spatial-temporal neural networks for traffic forecasting study spatial dependency and temporal correlation separately in processing, gravely impaired the spatial-temporal integrity, and ignore the fact that the neighbors' temporal dependency period for a node can be delayed and dynamic. To model this actual condition, we propose TraverseNet, a novel spatial-temporal graph neural network, viewing space and time as an inseparable whole, to mine spatial-temporal graphs while exploiting the evolving spatial-temporal dependencies for each node via message traverse mechanisms. Experiments with ablation and parameter studies have validated the effectiveness of the proposed TraverseNet, and the detailed implementation can be found from https://github.com/nnzhan/TraverseNet.

Mufei Li, Jinjing Zhou, Jiajing Hu et al.

Graph neural networks (GNNs) constitute a class of deep learning methods for graph data. They have wide applications in chemistry and biology, such as molecular property prediction, reaction prediction and drug-target interaction prediction. Despite the interest, GNN-based modeling is challenging as it requires graph data pre-processing and modeling in addition to programming and deep learning. Here we present DGL-LifeSci, an open-source package for deep learning on graphs in life science. DGL-LifeSci is a python toolkit based on RDKit, PyTorch and Deep Graph Library (DGL). DGL-LifeSci allows GNN-based modeling on custom datasets for molecular property prediction, reaction prediction and molecule generation. With its command-line interfaces, users can perform modeling without any background in programming and deep learning. We test the command-line interfaces using standard benchmarks MoleculeNet, USPTO, and ZINC. Compared with previous implementations, DGL-LifeSci achieves a speed up by up to 6x. For modeling flexibility, DGL-LifeSci provides well-optimized modules for various stages of the modeling pipeline. In addition, DGL-LifeSci provides pre-trained models for reproducing the test experiment results and applying models without training. The code is distributed under an Apache-2.0 License and is freely accessible at https://github.com/awslabs/dgl-lifesci.

Da Zheng, Chao Ma, Minjie Wang et al.

Graph neural networks (GNN) have shown great success in learning from graph-structured data. They are widely used in various applications, such as recommendation, fraud detection, and search. In these domains, the graphs are typically large, containing hundreds of millions of nodes and several billions of edges. To tackle this challenge, we develop DistDGL, a system for training GNNs in a mini-batch fashion on a cluster of machines. DistDGL is based on the Deep Graph Library (DGL), a popular GNN development framework. DistDGL distributes the graph and its associated data (initial features and embeddings) across the machines and uses this distribution to derive a computational decomposition by following an owner-compute rule. DistDGL follows a synchronous training approach and allows ego-networks forming the mini-batches to include non-local nodes. To minimize the overheads associated with distributed computations, DistDGL uses a high-quality and light-weight min-cut graph partitioning algorithm along with multiple balancing constraints. This allows it to reduce communication overheads and statically balance the computations. It further reduces the communication by replicating halo nodes and by using sparse embedding updates. The combination of these design choices allows DistDGL to train high-quality models while achieving high parallel efficiency and memory scalability. We demonstrate our optimizations on both inductive and transductive GNN models. Our results show that DistDGL achieves linear speedup without compromising model accuracy and requires only 13 seconds to complete a training epoch for a graph with 100 million nodes and 3 billion edges on a cluster with 16 machines. DistDGL is now publicly available as part of DGL:https://github.com/dmlc/dgl/tree/master/python/dgl/distributed.

Ke Yang, Yao Liu, Sapana Chaudhary et al.

Autonomy via agents using large language models (LLMs) for personalized, standardized tasks boosts human efficiency. Automating web tasks (like booking hotels within a budget) is increasingly sought after. Fulfilling practical needs, the web agent also serves as an important proof-of-concept example for various agent grounding scenarios, with its success promising advancements in many future applications. Prior research often handcrafts web agent strategies (e.g., prompting templates, multi-agent systems, search methods, etc.) and the corresponding in-context examples, which may not generalize well across all real-world scenarios. On the other hand, there has been limited study on the misalignment between a web agent's observation/action representation and the pre-training data of the LLM it's based on. This discrepancy is especially notable when LLMs are primarily trained for language completion rather than tasks involving embodied navigation actions and symbolic web elements. Our study enhances an LLM-based web agent by simply refining its observation and action space to better align with the LLM's capabilities. This approach enables our base agent to significantly outperform previous methods on a wide variety of web tasks. Specifically, on WebArena, a benchmark featuring general-purpose web interaction tasks, our agent AgentOccam surpasses the previous state-of-the-art and concurrent work by 9.8 (+29.4%) and 5.9 (+15.8%) absolute points respectively, and boosts the success rate by 26.6 points (+161%) over similar plain web agents with its observation and action space alignment. We achieve this without using in-context examples, new agent roles, online feedback or search strategies. AgentOccam's simple design highlights LLMs' impressive zero-shot performance on web tasks, and underlines the critical role of carefully tuning observation and action spaces for LLM-based agents.

Costas Mavromatis, Petros Karypis, George Karypis

Fusing knowledge from multiple Large Language Models (LLMs) can combine their diverse strengths to achieve improved performance on a given task. However, current fusion approaches either rely on learning-based fusers that do not generalize to new LLMs, or do not take into account how well each LLM understands the input. In this work, we study LLM fusion at test-time, which enables leveraging knowledge from arbitrary user-specified LLMs during inference. We introduce Pack of LLMs (PackLLM), an effective method for test-time fusion that leverages each LLM's expertise, given an input prompt. PackLLM performs model fusion by solving an optimization problem for determining each LLM's importance, so that perplexity over the input prompt is minimized. First, our simple PackLLM-sim variant validates that perplexity is a good indicator for measuring each LLM's expertise. Second, our PackLLM-opt variant approximately solves the perplexity minimization problem via a greedy algorithm. The derived importance weights are used to combine the LLMs during inference. We conduct experiments with over 100 total LLMs on a diverse set of tasks. Experimental results show that (i) perplexity is a reliable measure for LLM fusion, (ii) PackLLM outperforms test-time fusion baselines by 1.89% accuracy points, and (iii) PackLLM can leverage new LLMs to improve performance over learning-based fusion approaches by 3.92-11.94% accuracy points.