Test-time aggregation strategies, such as generating and combining multiple answers, can enhance LLM performance but eventually hit diminishing returns. Refinement, where model feedback is used to improve answers iteratively, presents an alternative. However, it faces three challenges: (1) excessive refinement, which can lead to over-correction and reduced accuracy; (2) difficulty in identifying and addressing specific… →
Collaborative Multi-Agent Reinforcement Learning (MARL) has emerged as a powerful approach in various domains, including traffic signal control, swarm robotics, and sensor networks. However, MARL faces significant challenges due to the complex interactions between agents, which introduce non-stationarity in the environment. This non-stationarity complicates the learning process and makes it difficult for agents to adapt… →
One of the main challenges in cellular automata (CA) systems, particularly in Conway’s Game of Life (Life), lies in predicting their emergent behavior without explicitly knowing the underlying grid topology. Life and other CA algorithms are computationally simple, yet they generate complex and unpredictable dynamics highly sensitive to initial conditions. This unpredictability complicates the development… →
Federated learning (FL) is a powerful ML paradigm that enables multiple data owners to train models without centralizing their data collaboratively. This approach is particularly valuable in domains where data privacy is critical, such as healthcare, finance, and the energy sector. The core of federated learning lies in training models on decentralized data stored on… →
Governments and humanitarian organizations need reliable data on building and infrastructure changes over time to manage urbanization, allocate resources, and respond to crises. However, many regions across the Global South need more access to timely and accurate data on buildings, making it difficult to track urban growth and infrastructure development. The absence of this data… →
Integer Linear Programming (ILP) is the foundation of combinatorial optimization, which is extensively applied across numerous industries to resolve challenging decision-making issues. Under a set of linear equality constraints, an ILP aims to minimize or maximize a linear objective function, with the important condition that all variables must be integers. Even while ILP is an… →
Input space mode connectivity in deep neural networks builds upon research on excessive input invariance, blind spots, and connectivity between inputs yielding similar outputs. The phenomenon exists generally, even in untrained networks, as evidenced by empirical and theoretical findings. This research expands the scope of input space connectivity beyond out-of-distribution samples, considering all possible inputs.… →
Multi-agent reinforcement learning (MARL) is a field focused on developing systems where multiple agents cooperate to solve tasks that exceed the capabilities of individual agents. This area has garnered significant attention due to its relevance in autonomous vehicles, robotics, and complex gaming environments. The aim is to enable agents to work together efficiently, adapt to… →
Artificial Intelligence (AI) safety has become an increasingly crucial area of research, particularly as large language models (LLMs) are employed in various applications. These models, designed to perform complex tasks such as solving symbolic mathematics problems, must be safeguarded against generating harmful or unethical content. With AI systems growing more sophisticated, it is essential to… →
In artificial intelligence and natural language processing, long-context reasoning has emerged as a crucial area of research. As the volume of information that needs to be processed grows, machines must be able to synthesize and extract relevant data from massive datasets efficiently. This goes beyond simple retrieval tasks, requiring models to locate specific pieces of… →