Federated Learning (FL) enables collaborative model training across decentralized clients without sharing raw data, yet faces significant challenges in real-world settings where client data distributions evolve dynamically over time. This paper tackles the critical problem of covariate and label shifts in streaming FL environments, where non-stationary data distributions degrade model performance and necessitate a middleware layer that adapts FL to distributional shifts. We introduce ShiftEx a shift-aware mixture of experts framework that dynamically creates and trains specialized global models in response to detected distribution shifts using Maximum Mean Discrepancy for covariate shifts. The framework employs a latent memory mechanism for expert reuse and implements facility location-based optimization to jointly minimize covariate mismatch, expert creation costs, and label imbalance. Through theoretical analysis and comprehensive experiments on five datasets, we demonstrate 5.5–12.9 percentage point accuracy improvements and 22–95% faster adaptation compared to state-of-the-art FL baselines across diverse shift scenarios. The proposed approach offers a scalable, privacy-preserving middleware solution for FL systems operating in non-stationary, real-world conditions while minimizing communication and computational overhead.

There are increasingly many large language models (LLMs) available to the public. While these LLMs have exhibited impressive abilities on a variety of task, any individual LLM in particular may do well on some tasks and worse on others. Additionally, the performance of these models is heavily dependent on the choice of prompt template used. For instance, they exhibit sensitivity to the few shot examples chosen or brittleness to the wording of instructions. Moreover, a prompt template that makes a model perform well for one input may not be the optimal template for another input. This necessitates an approach for adaptively selecting LLM and prompt template pairs for each input. Recent work has shown that the accuracy of LLM's responses is correlated with the LLM's confidence in the response. Thus, a natural choice for selecting which model and prompt template to use is to select the pair that is most confident in its response. However, existing confidence metrics are expensive to calculate - necessitating multiple calls to each LLm and prompt pair. We thus propose an approach to predict the confidence of each pair using an auxiliary regression model that is inexpensive to run. Using this auxiliary model, we select the LLM and prompt template with the highest predicted confidence for a given input. Results on a range of benchmark datasets show that our confidence-based instance-level prompt search method consistently improves the performance of LLMs.

Existing federated learning models that follow the standard risk minimization paradigm of machine learning often fail to generalize in the presence of spurious correlations in the training data. In many real-world distributed settings, spurious correlations exist due to biases and data sampling issues on distributed devices or clients that can erroneously influence models. Current generalization approaches are designed for centralized training and attempt to identify features that have an invariant causal relationship with the target, thereby reducing the effect of spurious features. However, such invariant risk minimization approaches rely on apriori knowledge of training data distributions which is hard to obtain in many applications. In this work, we present a generalizable federated learning framework called FedGen, which allows clients to identify and distinguish between spurious and invariant features in a collaborative manner without prior knowledge of training distributions. We evaluate our approach on real-world datasets from different domains and show that FedGen results in models that achieve significantly better generalization and can outperform the accuracy of current federated learning approaches by over 24%.

The popularity of conversational digital assistants has resulted in the availability of large amounts of conversational data which can be utilized for improved user experience and personalized response generation.Building these assistants using popular large language models like ChatGPT also require additional emphasis on prompt engineering and evaluation methods.Textual similarity metrics are a key ingredient for such analysis and evaluations.While many similarity metrics have been proposed in the literature, they have not proven effective for taskoriented conversations as they do not take advantage of unique conversational features.To address this gap, we present TaskDiff, a novel conversational similarity metric that utilizes different dialogue components (utterances, intents, and slots) and their distributions to compute similarity.Extensive experimental evaluation of TaskDiff on a benchmark dataset demonstrates its superior performance and improved robustness over other related approaches.

Dialogue State Tracking (DST), a key component of task-oriented conversation systems, represents user intentions by determining the values of pre-defined slots in an ongoing dialogue.Existing approaches use hand-crafted templates and additional slot information to fine-tune and prompt large pre-trained language models and elicit slot values from the dialogue context.Significant manual effort and domain knowledge is required to design effective prompts, limiting the generalizability of these approaches to new domains and tasks.In this work, we propose DiSTRICT, a generalizable in-context tuning approach for DST that retrieves highly relevant training examples for a given dialogue to fine-tune the model without any hand-crafted templates.Experiments with the MultiWOZ benchmark datasets show that DiSTRICT outperforms existing approaches in various zeroshot and few-shot settings using a much smaller model, thereby providing an important advantage for real-world deployments that often have limited resource availability.

The importance of IoT analytics in smart deploy-ments has resulted in an increased use of powerful Deep Neural Network (DNN) models to extract insights from the growing amount of IoT sensor data. Traditional approaches that entirely offload computation and model deployment to cloud servers have been shown to be inefficient due to network congestion and latency concerns. However, with the improved capabilities of IoT devices, it has now become possible to distribute and host DNNs across IoT devices, edge servers and the cloud. In this paper, we propose a multi-user system, called T2C, to dynamically choose, deploy, monitor and control DNN-driven IoT analytics in a thing-to-cloud continuum. T2C leverages strategies such as multi-task learning, hitchhiking, early exit, and dynamic reconfiguration, to maximize the number of served user requests while simultaneously satisfying accuracy and latency requirements. We propose a suite of deployment planning and reconfiguration algorithms to dynamically deploy and migrate DNN layers between IoT devices, edge servers, and the cloud. We implement T2C in a prototype testbed and show that our system: (i) achieves 6.8X throughput boost compared to baseline algorithms in the planning phase, and (ii) improves the satisfied ratio by up to 35% in the operation and reconfiguration phase.

The generalization of predictive models that follow the standard risk minimization paradigm of machine learning can be hindered by the presence of spurious correlations in the data. Identifying invariant predictors while training on data from multiple environments can influence models to focus on features that have an invariant causal relationship with the target, while reducing the effect of spurious features. Such invariant risk minimization approaches heavily rely on clearly defined environments and data being perfectly segmented into these environments for training. However, in real-world settings, perfect segmentation is challenging to achieve and these environment-aware approaches prove to be sensitive to segmentation errors. In this work, we present an environment-agnostic approach to develop generalizable models for classification tasks in sequential datasets without needing prior knowledge of environments. We show that our approach results in models that can generalize to out-of-distribution data and are not influenced by spurious correlations. We evaluate our approach on real-world sequential datasets from various domains.

An increasing number of community spaces are being instrumented with heterogeneous IoT sensors and actuators that enable continuous monitoring of the surrounding environments. Data streams generated from the devices are analyzed using a range of analytics operators and transformed into meaningful information for community monitoring applications. To ensure high quality results, timely monitoring, and application reliability, we argue that these operators must be hosted at edge servers located in close proximity to the community space. In this paper, we present a Resource Efficient Adaptive Monitoring (REAM) framework at the edge that adaptively selects workflows of devices and operators to maintain adequate quality of information for the application at hand while judiciously consuming the limited resources available on edge servers. IoT deployments in community spaces are in a state of continuous flux that are dictated by the nature of activities and events within the space. Since these spaces are complex and change dynamically, and events can take place under different environmental contexts, developing a one-size-fits-all model that works for all types of spaces is infeasible. The REAM framework utilizes deep reinforcement learning agents that learn by interacting with each individual community spaces and take decisions based on the state of the environment in each space and other contextual information. We evaluate our framework on two real-world testbeds in Orange County, USA and NTHU, Taiwan. The evaluation results show that community spaces using REAM can achieve > 90% monitoring accuracy while incurring ~ 50% less resource consumption costs compared to existing static monitoring and Machine Learning driven approaches.