Large Language Models (LLMs) have demonstrated remarkable prowess in generating contextually coherent responses, yet their fixed context windows pose fundamental challenges for maintaining consistency over prolonged multi-session dialogues. We introduce Mem0, a scalable memory-centric architecture that addresses this issue by dynamically extracting, consolidating, and retrieving salient information from ongoing conversations. Building on this foundation, we further propose an enhanced variant that leverages graph-based memory representations to capture complex relational structures among conversational elements. Through comprehensive evaluations on the LOCOMO benchmark, we systematically compare our approaches against six baseline categories. Empirical results demonstrate that our methods consistently outperform all existing memory systems across four question categories: single-hop, temporal, multi-hop, and open-domain. Notably, Mem0 achieves 26% relative improvements in the LLM-as-a-Judge metric over OpenAI, while Mem0 with graph memory achieves around 2% higher overall score than the base Mem0 configuration. Beyond accuracy gains, we also markedly reduce computational overhead compared to the full-context approach. In particular, Mem0 attains a 91% lower p95 latency and saves more than 90% token cost, thereby offering a compelling balance between advanced reasoning capabilities and practical deployment constraints. Our findings highlight the critical role of structured, persistent memory mechanisms for long-term conversational coherence, paving the way for more reliable and efficient LLM-driven AI agents. Code: https://mem0.ai/research.

Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk. Our code is available at: https://github.com/saccharomycetes/mllms_know.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but they exhibit problems with logical consistency in the output they generate. How can we harness LLMs’ broad-coverage parametric knowledge in formal reasoning despite their inconsistency? We present a method for directly integrating an LLM into the interpretation function of the formal semantics for a paraconsistent logic. We provide experimental evidence for the feasibility of the method by evaluating the function using datasets created from several short-form factuality benchmarks. Unlike prior work, our method offers a theoretical framework for neurosymbolic reasoning that leverages an LLM’s knowledge while preserving the underlying logic’s soundness and completeness properties.

Food computing has emerged as a prominent multidisciplinary field of research in recent years. An ambitious goal of food computing is to develop end-to-end intelligent systems capable of autonomously producing recipe information for a food image. Current image-to-recipe methods are retrieval-based and their success depends heavily on the dataset size and diversity, as well as the quality of learned embeddings. Meanwhile, the emergence of powerful attention-based vision and language models presents a promising avenue for accurate and generalizable recipe generation, which has yet to be extensively explored. This paper proposes FIRE, a novel multimodal methodology tailored to recipe generation in the food computing domain, which generates the food title, ingredients, and cooking instructions based on input food images. FIRE leverages the BLIP model to generate titles, utilizes a Vision Transformer with a decoder for ingredient extraction, and employs the T5 model to generate recipes incorporating titles and ingredients as inputs. We showcase two practical applications that can benefit from integrating FIRE with large language model prompting: recipe customization to fit recipes to user preferences and recipe-to-code transformation to enable automated cooking processes. Our experimental findings validate the efficacy of our proposed approach, underscoring its potential for future advancements and widespread adoption in food computing.

The rapid development of Artificial Intelligence (AI) has had a significant impact on various industries, including healthcare. The Internet of Medical Things (IoMT) has played a vital role in this evolution. However, while AI has contributed to many benefits in healthcare, concerns about data privacy and security persist. To address these concerns, we propose a framework that combines Federated Learning (FL) and Differential Privacy (DP) to enhance data protection within IoMT. By integrating FL’s decentralized approach with DP’s mechanism to prevent data reconstruction from model outputs, we can improve data confidentiality. This integrated approach is used to develop and analyze high-performing Convolutional Neural Networks (CNNs) for detecting Tuberculosis using chest X-ray datasets. The framework undergo thorough performance evaluation, utilizing various metrics to establish its superiority over baseline models. The results demonstrate the effectiveness of our framework as a robust solution for secure and private AI applications in healthcare.

Communication via natural language is a key aspect of machine intelligence, and it requires computational models to learn and reason about world concepts, with varying levels of supervision. Significant progress has been made on fully-supervised non-interactive tasks, such as question-answering and procedural text understanding. Yet, various sequential interactive tasks, as in text-based games, have revealed limitations of existing approaches in terms of coherence, contextual awareness, and their ability to learn effectively from the environment. In this paper, we propose a knowledge-injection framework for improved functional grounding of agents in text-based games. Specifically, we consider two forms of domain knowledge that we inject into learning-based agents: memory of previous correct actions and affordances of relevant objects in the environment. Our framework supports two representative model classes: reinforcement learning agents and language model agents. Furthermore, we devise multiple injection strategies for the above domain knowledge types and agent architectures, including injection via knowledge graphs and augmentation of the existing input encoding strategies. We experiment with four models on the 10 tasks in the ScienceWorld text-based game environment, to illustrate the impact of knowledge injection on various model configurations and challenging task settings. Our findings provide crucial insights into the interplay between task properties, model architectures, and domain knowledge for interactive contexts.

Summary From the last few years, we have witnessed the fourth generation industrial revolution (Industry 4.0), impact of which will be seen in the years to come in various disciplines such as healthcare, transportation, IoT, smart grid, autonomous vehicles, and image processing. These applications in Industry 4.0 may have data in the form of images, speech signals, videos having high dimensions containing multiple dimensions to represent data along different axis. So, the complexity of data processing increases with an increase in the dimensions of the dataset. Complexity can be viewed in terms of detecting and exploiting the relationships among different features of the dataset. These complexities among different attributes can be reduced with the help of dimensionality reduction techniques. These techniques reduce the dimensions from the original input dataset to a lower dimensional dataset. Dimensionality reduction methods are broadly categorized into two types asfeature extraction and feature selection. In feature selection method, out of the original set, a subset of features are identified to get a smaller subset which can be used to build the model whereas, the feature extraction method reduces the dataset of high dimensions to a lower dimension space, that is, a space with a less number of features having different values in comparison to the original dataset. Keeping focus on these points, in this article, we have compared and analyzed different data dimensionality reduction techniques which reduce the dimensions of a large and complex dataset during data processing. In addition, we have discussed various data dimensionality reduction techniques and compared these techniques with respect to various parameters. The comparison among various techniques provides insights to the readers about the applicability of a specific technique to the stand‐alone or a group of applications.

People need to breathe, and so do other living beings, including plants and animals. It is impossible to overlook the impact of air pollution on nature, human well-being, and concerned countries' economies. Monitoring of air pollution and future predictions of air quality have lately displayed a vital concern. There is a need to predict the air quality index with high accuracy; on a real-time basis to prevent people from health issues caused by air pollution. With the help of Unmanned Aerial Vehicle's onboard sensors, we can collect air quality data easily. The paper proposes a distributed and decentralized Federated Learning approach within a UAV swarm. The accumulated data by the sensors are used as an input to the Long Short Term Memory (LSTM) model. Each UAV used its locally gathered data to train a model before transmitting the local model to the central base station. The central base station creates a master model by combining all the UAV's local model weights of the participating UAVs in the FL process and transmits it to all UAV s in the subsequent cycles. The effectiveness of the proposed model is evaluated with other machine learning models using various evaluation metrics using test data from the capital city of India, i.e., Delhi.

As stated by Spock, “change is the essential process of all existence,” which is reflected in everyday applications in our daily lives. We, as humans, just need to find a way to make the best use of the current technological advances. The pandemic has managed to exploit our deepest vulnerabilities and insecurities. We need to cope with a lot of things, just to be comfortable in the new normal. Hence, we can rely on technology, the greatest asset developed by humans. In this article, we discuss how we can enhance the work environment in offices post-pandemic. We combine federated learning with emotion analysis to create a state-of-the-art, simple, secure, and efficient emotion monitoring system. We combine facial expression and speech signals to find out macroexpressions and create an emotion index that is monitored to find the mental health of the user. Federated learning enables users to locally train the model without compromising his/her privacy. In place of sending data to the centralized server, the proposed scheme sends only model weights that are combined at the server to make a better global model, which is further pushed back to the users. This model is then trained interorganizational as it does not violate the privacy or data sharing to achieve optimal results. The data collected from users are monitored to analyze the mental health and presented with counseling solutions during low times. Technology is a panacea that has enabled us to survive in this pandemic, and by using our solution to improve work culture and the environment in post-pandemic times.

The applications of unmanned aerial vehicles (UAVs) in military, intelligent transportation, agriculture, rescue operations, natural environment mapping, and many other allied domains has increased exponentially during the past few years. Some of the use cases of their applications range from aerial surveillance, data retrieval to their use in real-time communicative networks. Though UAVs were traditionally used only outdoors, many of its indoor applications like for rescue operations, inventory tracking in warehouses, etc., have recently emerged and these use cases are being actively explored. One of the major challenges for indoor drone applications is navigation and obstacle avoidance. Due to indoor operations, the global positioning system fails in accurate localization and navigation. To address this issue, we introduce a scheme that facilitates the autonomous navigation of UAVs (which have an onboard camera) in the indoor corridors of a building using deep-neural-networks-based processing of images. For a deep neural network, the selection of a good combination of hyperparameters for a better prediction is a complicated task. In this article, the hyperparameters tuning of a convolutional neural network is achieved by using genetic algorithms. The proposed architecture (DCNN-GA) is compared with state-of-the-art ImageNet models. The experimental results show the minimum loss and high performance of the proposed algorithm.

Air pollution monitoring, finding the hazardous zone, and future air quality predictions have recently become a significant issue for many researchers. With the adverse effect of low air quality on human health, it has become necessary for predicting the air quality index (AQI) accurately and on time. The unmanned aerial vehicle (UAV) can collect air quality data with high spatial and temporal resolutions. Using a fleet of UAVs could be considered a good option. In the proposed work, we implement a distributed federated learning (FL) algorithm within a UAV swarm that collects air quality data using built-in sensors. A scheme for finding the area with the highest AQI value is proposed using swarm intelligence. The collected data are then fed to a CNN-LSTM model to predict the AQI. The trained local model is sent to the central server, and the server aggregates the received models from UAVs in the swarm. A global model is created and is transmitted to the UAV swarm again in the next iteration. The proposed architecture is compared with other time-series models. The results show that the proposed model predicts AQI daily with a minimal error rate on a real-time data set from Delhi.

Virtualization is an essential feature in the IoT-resource-constrained environment due to which the service providers are facing challenges to minimize the energy consumption by IoT devices. Energy consumption models are pivotal in designing and optimizing energy-efficient operations to curb excessive energy consumption of IoT devices, which are an integral part of the modern data centers. A lot of research work has focused on efficient management of energy consumption by virtue of virtual machine consolidation. The existing virtualization techniques may not be suitable for this problem due to high computational overhead. As containers have been recently getting much popularity to encapsulate fog services, so they are the best candidate to handle this problem, especially for intelligent IoT devices. Keeping the focus on all these issues, in this article, we propose an energy-efficient container migration scheme by migrating the container from the source host server to the destination host server to meet the container's resource requirement. We used a novel approach to find the best destination host for container placement to solve host overload or underload problems using the best-fit container placement technique. The results obtained on the benchmark data set with respect to various performance evaluation metrics prove the efficacy of the designed scheme in comparison to the other existing state-of-the-art schemes.