Researchers have proposed the Cylindrical Representation Hypothesis (CRH) to address the instability and unpredictability observed in steering large language models, an issue not fully explained by the existing Linear Representation Hypothesis (LRH). CRH suggests that overlapping concept contributions lead to a sample-specific axis-orthogonal structure, comprising a central axis for concept generation and a surrounding normal plane for steering sensitivity. This framework identifies intrinsic uncertainty at the 'sensitive sector' level within the plane, providing a principled explanation for fluctuations in steering outcomes. Experiments verify the existence of this cylindrical structure and demonstrate CRH's practical utility in interpreting real-world model steering behavior, with code available on GitHub from mbzuai-nlp. Why it matters: This research from MBZUAI offers a crucial theoretical advancement in understanding and potentially improving the control and reliability of large language models.
Researchers from MBZUAI have introduced WR-Arena, a new comprehensive benchmark designed to evaluate World Models (WMs) beyond traditional next-state prediction and visual fidelity. WR-Arena assesses WMs across three core dimensions: Action Simulation Fidelity, Long-horizon Forecast, and Simulative Reasoning and Planning, using a curated task taxonomy and diverse datasets. Extensive experiments with state-of-the-art WMs revealed a significant gap between current models' capabilities and human-level hypothetical reasoning. Why it matters: This benchmark provides a critical diagnostic tool and guideline for developing more robust and intelligent world models capable of advanced understanding, forecasting, and purposeful action, particularly for AI research in the region.
MBZUAI's Institute of Foundation Models (IFM) has developed PAN, a novel interactive world model capable of long-horizon simulation. PAN uses a Generative Latent Prediction (GLP) architecture, coupling internal latent reasoning with generative supervision in the visual domain. The model evolves an internal latent state conditioned on history and natural language actions, then decodes that state into a video segment using a Causal Swin-DPM mechanism for smooth transitions. Why it matters: PAN represents a significant advancement in AI's ability to simulate and predict evolving environments, enabling more steerable and coherent long-term video generation and opening new possibilities for interactive AI systems.
MBZUAI has launched the Institute of Foundation Models (IFM) with a new Silicon Valley Lab in Sunnyvale, CA, joining existing facilities in Paris and Abu Dhabi. The launch event showcased PAN, a world model for simulating diverse realities with multimodal inputs. The IFM lab is also advancing K2-65B and JAIS AI systems. Why it matters: This expansion enhances MBZUAI's global presence and connects it with a critical AI ecosystem, supporting the UAE's economic diversification through advanced AI technologies.
MBZUAI is previewing PAN, a next-generation world model designed to simulate diverse realities and advance machine reasoning. PAN allows researchers to test AI agents in simulated environments before real-world deployment, enabling them to learn from mistakes without real-world consequences. It facilitates complex reasoning about actions, outcomes, and interactions, crucial for reliable AI performance in dynamic environments. Why it matters: PAN represents a significant advancement in AI by enabling comprehensive simulation and testing of AI agents, which can revolutionize fields like disaster management and healthcare where real-world experimentation is risky.
Krishna Murthy, a postdoc at MIT, researches computational world models to enable robots to understand and operate effectively in the physical world. His work focuses on differentiable computing approaches for spatial perception and interfaces large image, language, and audio models with 3D scenes. Murthy envisions structured world models working with scaling-based approaches to create versatile robot perception and planning algorithms. Why it matters: This research could significantly advance robotics by enabling more sophisticated perception, reasoning, and action capabilities in embodied agents.
Mingyu Ding from UC Berkeley presented research on endowing robots with human-like commonsense and physical reasoning capabilities. The talk covered multimodal commonsense reasoning integrating vision, world models, and language-based task planners. It also discussed physical reasoning approaches for robots to infer dynamics and physical properties of objects. Why it matters: Enhancing robots with these capabilities can improve their ability to generalize across everyday tasks, leading to greater social benefits and impact.
MBZUAI Professor Ivan Laptev is working to bridge the gap between data-driven AI systems and embodied agents (robots). He notes challenges in robotics including data scarcity, the need to generate new data through actions, and the requirement for real-time operation. Laptev aims to transfer innovations from computer vision to robotics, addressing these challenges to improve robots' ability to interpret and respond to the complexities of the real world. Why it matters: Overcoming these hurdles is crucial for advancing robotics and enabling robots to effectively interact with and navigate dynamic real-world environments.