Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa Elhoushi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, Carole-Jean Wu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Contrastive Language-Image Pretraining (CLIP) is a popular foundation model, supporting from zero-shot classification, retrieval to encoders for multimodal large language models (MLLMs). Although CLIP is successfully trained on billion-scale image-text pairs from the English world, scaling CLIP's training further to learning from the worldwide web data is still challenging: (1) no curation method is available to handle data points from non-English world; (2) the English performance from existing multilingual CLIP is worse than its English-only counterpart, i.e., "curse of multilinguality" that is common in LLMs. Here, we present Meta CLIP 2, the first recipe training CLIP from scratch on worldwide web-scale image-text pairs. To generalize our findings, we conduct rigorous ablations with minimal changes that are necessary to address the above challenges and present a recipe enabling mutual benefits from English and non-English world data. In zero-shot ImageNet classification, Meta CLIP 2 ViT-H/14 surpasses its English-only counterpart by 0.8% and mSigLIP by 0.7%, and surprisingly sets new state-of-the-art without system-level confounding factors (e.g., translation, bespoke architecture changes) on multilingual benchmarks, such as CVQA with 57.4%, Babel-ImageNet with 50.2% and XM3600 with 64.3% on image-to-text retrieval. Code and model are available at https://github.com/facebookresearch/MetaCLIP.

With increasingly powerful large language models (LLMs) and LLM-based agents tackling an ever-growing list of tasks, we envision a future where numerous LLM agents work seamlessly with other AI agents and humans to solve complex problems and enhance daily life. To achieve these goals, LLM agents must develop collaborative skills such as effective persuasion, assertion and disagreement, which are often overlooked in the prevalent single-turn training and evaluation of LLMs. In this work, we present Collaborative Reasoner (Coral), a framework to evaluate and improve the collaborative reasoning abilities of language models. In particular, tasks and metrics in Coral necessitate agents to disagree with incorrect solutions, convince their partners of a correct solution, and ultimately agree as a team to commit to a final solution, all through a natural multi-turn conversation. Through comprehensive evaluation on six collaborative reasoning tasks covering domains of coding, math, scientific QA and social reasoning, we show that current models cannot effectively collaborate due to undesirable social behaviors, collapsing even on problems that they can solve singlehandedly. To improve the collaborative reasoning capabilities of LLMs, we propose a self-play method to generate synthetic multi-turn preference data and further train the language models to be better collaborators. Experiments with Llama-3.1, Ministral and Qwen-2.5 models show that our proposed self-improvement approach consistently outperforms finetuned chain-of-thought performance of the same base model, yielding gains up to 16.7% absolute. Human evaluations show that the models exhibit more effective disagreement and produce more natural conversations after training on our synthetic interaction data.

As artificial intelligence and machine learning algorithms become increasingly prevalent in society, multiple stakeholders are calling for these algorithms to provide explanations. At the same time, these stakeholders, whether they be affected citizens, government regulators, domain experts, or system developers, have different explanation needs. To address these needs, in 2019, we created AI Explainability 360, an open source software toolkit featuring ten diverse and state-of-the-art explainability methods and two evaluation metrics. This paper examines the impact of the toolkit with several case studies, statistics, and community feedback. The different ways in which users have experienced AI Explainability 360 have resulted in multiple types of impact and improvements in multiple metrics, highlighted by the adoption of the toolkit by the independent LF AI & Data Foundation. The paper also describes the flexible design of the toolkit, examples of its use, and the significant educational material and documentation available to its users.

Contraceptive use improves the health of women and children in several ways, yet data shows high rates of discontinuation which is not well understood. We introduce an AI-based decision platform capable of analyzing event data to identify patterns of contraceptive uptake that are unique to a subpopulation of interest. These discriminatory patterns provide valuable, interpretable insights to policy-makers. The sequences then serve as a hypothesis for downstream causal analysis to estimate the effect of specific variables on discontinuation outcomes. Our platform presents a way to visualize, stratify, compare, and perform a causal analysis on covariates that determine contraceptive uptake behavior, and yet is general enough to be extended to a variety of applications.

As artificial intelligence algorithms make further inroads in high-stakes societal applications, there are increasing calls from multiple stakeholders for these algorithms to explain their outputs. To make matters more challenging, different personas of consumers of explanations have different requirements for explanations. Toward addressing these needs, we introduce AI Explainability 360, an open-source Python toolkit featuring ten diverse and state-of-the-art explainability methods and two evaluation metrics. Equally important, we provide a taxonomy to help entities requiring explanations to navigate the space of interpretation and explanation methods, not only those in the toolkit but also in the broader literature on explainability. For data scientists and other users of the toolkit, we have implemented an extensible software architecture that organizes methods according to their place in the AI modeling pipeline. The toolkit is not only the software, but also guidance material, tutorials, and an interactive web demo to introduce AI explainability to different audiences. Together, our toolkit and taxonomy can help identify gaps where more explainability methods are needed and provide a platform to incorporate them as they are developed.

To improve the performance of data-intensive applications, existing datacenter schedulers optimize either the placement of tasks or the scheduling of network flows. The task scheduler strives to place tasks close to their input data (i.e., maximize data locality) to minimize network traffic, while assuming fair sharing of the network. The network scheduler strives to finish flows as quickly as possible based on their sources and destinations determined by the task scheduler, while the scheduling is based on flow properties (e.g., size, deadline, and correlation) and not bound to fair sharing. Inconsistent assumptions of the two schedulers can compromise the overall application performance. In this paper, we propose NEAT+, a task scheduling framework that leverages information from the underlying network scheduler and available compute resources to make task placement decisions. The core of NEAT+ is a task completion time predictor that estimates the completion time of a task under given network condition and a given network scheduling policy. NEAT+ leverages the predicted task completion times to minimize the average completion time of active tasks. Evaluation using ns2 simulations and real-testbed shows that NEAT+ improves application performance by up to 3.7x for the suboptimal network scheduling policies and up to 33% for the optimal network scheduling policy.

Managing large-scale distributed systems is a difficult task. System administrators are responsible for the upkeep and maintenance of numerous components with complex dependencies. With the shift to microservices-based architectures, these systems can consist of 100s to 1000s of interconnected nodes. To combat this difficulty, administrators rely on analyzing logs and metrics collected from the different services. However, the number of available metrics for large systems presents complexity and scaling issues. To combat these issues, we present Minerva, an unsupervised Machine Learning (ML) framework for performing network diagnosis analysis. Minerva is composed of a multi-stage pipeline, where each component can act individually or cohesively to perform various management tasks. Our system offers a unified and extensible framework for managing the complexity of large networks, and presents administrators with a swiss-army knife for diagnosing the overall health of their systems. To demonstrate the feasibility of Minerva, we evaluate its performance on a production-scale system. We present use cases for the various management tools made available by Minerva, and show how these tools can be used to make strong inferences about the system using unsupervised techniques.

In this paper we provide a critical analysis with metrics that will inform guidelines for designing distributed systems for Collective Situational Understanding (CSU). CSU requires both collective insight-i.e., accurate and deep understanding of a situation derived from uncertain and often sparse data and collective foresight-i.e., the ability to predict what will happen in the future. When it comes to complex scenarios, the need for a distributed CSU naturally emerges, as a single monolithic approach not only is unfeasible: it is also undesirable. We therefore propose a principled, critical analysis of AI techniques that can support specific tasks for CSU to derive guidelines for designing distributed systems for CSU.

To improve the performance of data-intensive applications, existing datacenter schedulers optimize either the placement of tasks or the scheduling of network flows. The task scheduler strives to place tasks close to their input data (i.e., maximize data locality) to minimize network traffic, while assuming fair sharing of the network. The network scheduler strives to finish flows as quickly as possible based on their sources and destinations determined by the task scheduler, while the scheduling is based on flow properties (e.g., size, deadline, and correlation) and not bound to fair sharing. Inconsistent assumptions of the two schedulers can compromise the overall application performance. In this paper, we propose NEAT, a task scheduling framework that leverages information from the underlying network scheduler to make task placement decisions. The core of NEAT is a task completion time predictor that estimates the completion time of a task under given network condition and a given network scheduling policy. NEAT leverages the predicted task completion times to minimize the average completion time of active tasks. Evaluation using ns2 simulations and real-testbed shows that NEAT improves application performance by up to 3.7x for the suboptimal network scheduling policies and up to 30% for the optimal network scheduling policy.

The poor performance of TCP in wireless networks is a well-known problem, and a large amount of research effort has been devoted to it. However, our own experiments show that existing solutions including split TCP and recently developed congestion control algorithms still suffer significant performance degradation in lossy environments, leaving considerable room for improvement. Rather than developing a more sophisticated congestion control algorithm, we explore a radically different approach: we instead propose to completely remove congestion control between the Wi-Fi access points and the wireless receivers. We introduce TCP Fixed, a TCP used in split TCP environments that eliminates congestion control for the last lossy hop, allowing TCP to send data as fast as the channel allows. Extensive evaluations in both lab and real-world settings demonstrate that our approach significantly improves TCP performance in the presence of packet losses. In addition, our results indicate that existing 802.11 rate adaptation schemes which strive to minimize frame loss unnecessarily decrease the data rate in the presence of TCP Fixed. These results offer new opportunities for future link-layer rate adaptation designs.

Fast service placement, finding a set of nodes with enough free capacity of computation, storage, and network connectivity, is a routine task in daily cloud administration. In this work, we formulate this as a subgraph matching problem. Different from the traditional setting, including approximate and probabilistic graphs, subgraph matching on data-center networks has two unique properties. (1) Node/edge labels representing vacant CPU cycles and network bandwidth change rapidly, while the network topology varies little. (2) There is a partial order on node/edge labels. Basically, one needs to place service in nodes with enough free capacity. Existing graph indexing techniques have not considered very frequent label updates, and none of them supports partial order on numeric labels. Therefore, we resort to a new graph index framework, Gradin, to address both challenges. Gradin encodes subgraphs into multi-dimensional vectors and organizes them with indices such that it can efficiently search the matches of a query's subgraphs and combine them to form a full match. In particular, we analyze how the index parameters affect update and search performance with theoretical results. Moreover, a revised pruning algorithm is introduced to reduce unnecessary search during the combination of partial matches. Using both real and synthetic datasets, we demonstrate that Gradin outperforms the baseline approaches up to 10 times.

We consider the problem of endhost-based shortest path routing in a network with unknown, time-varying link qualities. Endhost-based routing is needed when internal nodes of the network do not have the scope or capability to provide globally optimal paths to given source-destination pairs, as can be the case in networks consisting of autonomous subnetworks or those with endhost-based routing restrictions. Assuming the source can probe links along selected paths, we formulate the problem as an online learning problem, where an existing solution achieves a performance loss (called regret) that is logarithmic in time with respect to (wrt) an offline algorithm that knows the link qualities. Current solutions assume coupled probing and routing; in contrast, we give a simple algorithm based on decoupled probing and routing, whose regret is only constant in time. We then extend our solution to support multi-path probing and cooperative learning between multiple sources, where we show an inversely proportional decay in regret wrt the probing rate. We also show that without the decoupling, the regret grows at least logarithmically in time, thus establishing decoupling as critical for obtaining constant regret. Although our analysis assumes certain conditions (i.i.d.) on link qualities, our solution applies with straightforward amendments to much broader scenarios where these conditions are relaxed. The efficacy of the proposed solution is verified by trace-driven simulations.

Rate adaptation is a MAC-layer mechanism in IEEE 802.11 networks that ensures efficient utilization of constantly varying wireless channel. For broadcast communication, however, the current practice is to send data at base rates (i.e., 1 or 2 Mbps in 802.11b/b+g and 6Mbps in 802.11a/g), leaving the channel capacity significantly underutilized. In this paper, we present a practical rate adaptation framework designed for broadcasting of streaming data to improve the channel utilization of 802.11 networks. The proposed framework jointly selects the data rate and the error protection level of FEC in a novel manner, and does not require changes to either existent applications or to the 802.11 standards - thus can be deployed in existing 802.11 networks. Using the performance measurement study in various live networking environments, we show that the proposed scheme provides substantial improvement in wireless channel utilization (more than 10x improvement compared to the current broadcast mode) and can adapt to different operational environments.

Abstract Monitoring and troubleshooting a large wireless mesh network (WMN) presents several challenges. Diagnosis of problems related to wireless access in these networks requires a comprehensive set of metrics and network monitoring data. Collection and offloading of a large amount of data are infeasible in a bandwidth constrained mesh network. Additionally, the processing required to analyze data from the entire network restricts the scalability of the system and impacts the ability to perform real‐time fault diagnosis. To this end, we propose MeshMon, a network monitoring framework that includes a multi‐tiered method of data collection. MeshMon, dynamically controls the granularity of data collection based on observed events in the network, thereby achieving significant bandwidth savings and enabling real‐time automated management. Our evaluation of MeshMon on a real testbed shows that we can diagnose a majority (87%) of network faults with a 66% savings in bandwidth required for network monitoring. Copyright © 2010 John Wiley & Sons, Ltd.

Wireless networks have evolved into an important technology for connecting users to the Internet. As the utility of wireless technology grows, wireless networks are being deployed in more widely varying conditions. The monitoring of wireless networks continues to reveal key implementation deficiencies that need to be corrected in order to improve protocol operation and end-to-end network performance. In wireless networks, where the medium is shared, unwanted traffic can pose significant overhead and lead to suboptimal network performance. Much of the previous analyses of unwanted traffic in wireless networks focus on malicious traffic. However, another major contributor of unwanted traffic is incorrect link layer behavior. Using data we collected from the 67th Internet Engineering Task Force (IETF) meeting held in November 2006, we show that a significant portion of link layer traffic stems from mechanisms that initiate, maintain, and change client-AP associations. We further show that under conditions of high medium utilization and packet loss rate, handoffs are initiated incorrectly. We analyze the traffic to understand when handoffs occur and whether the handoffs were beneficial or should have been avoided.

Power delivery, electricity consumption, and heat management are becoming key challenges in data center environments. Several past solutions have individually evaluated different techniques to address separate aspects of this problem, in hardware and software, and at local and global levels. Unfortunately, there has been no corresponding work on coordinating all these solutions. In the absence of such coordination, these solutions are likely to interfere with one another, in unpredictable (and potentially dangerous) ways. This paper seeks to address this problem. We make two key contributions. First, we propose and validate a power management solution that coordinates different individual approaches. Using simulations based on 180 server traces from nine different real-world enterprises, we demonstrate the correctness, stability, and efficiency advantages of our solution. Second, using our unified architecture as the base, we perform a detailed quantitative sensitivity analysis and draw conclusions about the impact of different architectures, implementations, workloads, and system design choices.

We study a fundamental yet under-explored facet in wireless communication -- the width of the spectrum over which transmitters spread their signals, or the channel width. Through detailed measurements in controlled and live environments, and using only commodity 802.11 hardware, we first quantify the impact of channel width on throughput, range, and power consumption. Taken together, our findings make a strong case for wireless systems that adapt channel width. Such adaptation brings unique benefits. For instance, when the throughput required is low, moving to a narrower channel increases range and reduces power consumption; in fixed-width systems, these two quantities are always in conflict. We then present a channel width adaptation algorithm, called SampleWidth, for the base case of two communicating nodes. This algorithm is based on a simple search process that builds on top of existing techniques for adapting modulation. Per specified policy, it can maximize throughput or minimize power consumption. Evaluation using a prototype implementation shows that SampleWidth correctly identities the optimal width under a range of scenarios. In our experiments with mobility, it increases throughput by more than 60% compared to the best fixed-width configuration.

Power and cooling are emerging to be key challenges in data center environments. A recent IDC report estimated the worldwide spending on enterprise power and cooling to be more than $30 billion and likely to even surpass spending on new server hardware. Server rated power consumptions have increased by nearly 10X over the past ten years. This has led to increased spending on cooling and power delivery equipment. A 30,000 square feet 10MW data center can need up to five million dollars of cooling infrastructure; similarly, power delivery beyond 60 Amps per rack can pose fundamental issues. The increased power also has implications on electricity costs, with many data centers reporting millions of dollars for annual usage. From an environmental point of view, the Department of Energy’s 2007 estimate of 59 billion KWhrs spent in U.S. servers and data centers translates to several million tons of coal consumption and greenhouse gas emission per year. The U.S. Congress recently passed Public Law 109—431, directing the Environmental Protection Agency (EPA) to study enterprise energy use, and several industry consortiums such as the GreenGrid have been formed to address these issues. In addition, power and cooling can also impact compaction and reliability.

As the utility of wireless technology grows, wireless networks are being deployed in more widely varying conditions. The monitoring of these networks continues to reveal key implementation deficiencies that need to be corrected in order to improve protocol operation and end-to-end performance. Using data we collected from the 67th Internet Engineering Task Force (IETF) meeting held in November 2006, we show that under conditions of high medium utilization and packet loss, handoffs can be incorrectly initiated. Using the notion of persistence and prevalence for the association of a client to an Access Point (AP), we show that although the clients were predominantly static, the handoff rate is surprisingly high. Through the analysis of the data set, we show that unnecessary handoff events not only increase the amount of management traffic in the network, but also severely impact client performance.