Research Paper

Current Large Language Models are limited by their static "train then deploy" paradigm, hindering dynamic adaptation to new information. Test-Time Training (TTT) offers a promising alternative by updating model parameters during inference, but faces significant challenges like architectural incompatibility and computational inefficiency in existing LLM ecosystems. This paper, "In-Place Test-Time Training," aims to address these critical barriers, potentially enabling more adaptive and continuously learning LLMs.

This research introduces the first quantitative, topology-based definition for 'churn flow,' a chaotic regime in vertical two-phase flow, a problem unsolved for over 40 years. It utilizes Euler Characteristic Surfaces (ECS) and Multiple Kernel Learning (MKL) to characterize and unsupervisedly correct flow-regime maps in small-diameter vertical pipes. This novel approach blends temporal alignment and amplitude statistics derived from ECS to provide a robust characterization.

This paper addresses the challenge of computational cost and performance degradation in large reasoning models due to extended chain-of-thought generation. It proposes an "Early Stopping" mechanism by studying the confidence dynamics of intermediate answers during reasoning. The research observes that correct reasoning trajectories typically reach high-confidence answers, suggesting a method to determine optimal stopping points for these models.

This paper proposes a novel stratification-based semantics for Signal Temporal Logic (STL), interpreting atomic predicates as membership tests in a stratified space. This new perspective reveals a correspondence between STL and stratification theory, suggesting that most STL formulas induce a stratification of space-time. This offers a fresh theoretical framework for analyzing STL's structure within reinforcement learning contexts.

This paper reveals that pre-trained diffusion models inherently possess restoration capabilities, which can be activated by directly learning prompt embeddings. Unlike current methods that rely on fine-tuning or Control-Net modules, this work demonstrates that these models secretly know how to perform All-in-One Restoration (AiOR). This discovery could lead to more efficient and direct ways to leverage diffusion models for image restoration without extensive modifications.

The rapid expansion of the Internet of Medical Things (IoMT) is transforming healthcare but also introduces severe cybersecurity and patient safety concerns due to increasing cyberattacks. This paper proposes a novel Tsetlin Machine (TM)-based Intrusion Detection System (IDS) designed to detect a wide range of cyberattacks targeting IoMT networks. The aim is to enhance the security and resilience of these critical medical infrastructures.

This paper addresses the challenges of long-horizon control in model predictive control (MPC) with learned world models, which often struggle with accumulating prediction errors and large search spaces. It proposes a solution by learning latent world models at multiple temporal scales and performing hierarchical planning. This approach aims to improve the generalization and long-term planning capabilities of embodied AI systems.

This paper proposes a novel approach to enhance the robustness of federated learning against malicious attacks, even when clients' training data is not independently and identically distributed. The method combines server learning, client update filtering, and geometric median aggregation into a heuristic algorithm. Experiments demonstrate that this technique significantly improves model accuracy, even when a high fraction of clients are malicious.

This paper introduces 'Steerable Visual Representations' to address limitations in current vision models. While pretrained Vision Transformers (ViTs) offer generic features, they lack the ability to be directed towards less prominent visual concepts. Conversely, Multimodal LLMs (MLLMs) are steerable via text but often lose visual richness, becoming language-centric. This research aims to provide fine-grained control over visual features, combining the benefits of both approaches.

This paper analyzes the standard practice of initializing new vocabulary tokens in Language Models (LMs) for domain-specific tasks like generative recommendation. Typically, these tokens are initialized as the mean of existing vocabulary embeddings, followed by supervised fine-tuning. The research systematically demonstrates that this mean initialization strategy causes all new tokens to collapse, hindering their effective representation learning.

ActionParty addresses a critical limitation in current video diffusion "world models" by enabling multi-subject action binding. While existing models struggle to control multiple agents simultaneously and associate specific actions with their corresponding subjects, ActionParty allows for more complex, multi-agent interactions in generative video games. This advancement moves beyond single-agent settings to create richer interactive environments.

The paper introduces LAPIS-SHRED, a new method designed to reconstruct full spatio-temporal dynamics from sparse and limited observational data. This addresses a central challenge in complex systems where measurements are often incomplete in space and time. Approximating these complete trajectories is crucial for mechanistic insight, model calibration, and operational decision-making across various fields.

YC-Bench is a new benchmark designed to evaluate AI agents' capabilities in long-term planning, consistent execution, and strategic coherence. It tasks agents with running a simulated startup over a one-year horizon, requiring them to manage employees and select contracts across hundreds of turns. This benchmark aims to assess how well agents can plan under uncertainty, learn from delayed feedback, and adapt to compounding mistakes.

HippoCamp is a novel benchmark designed to evaluate AI agents' capabilities in multimodal file management on personal computers. Unlike existing benchmarks, it focuses on user-centric environments, utilizing real-world user profiles and massive personal file systems to test context-aware reasoning. This benchmark aims to drive the development of more personalized and practical AI agents capable of navigating complex individual digital environments.

This paper investigates Chain-of-Thought (CoT) monitoring as a method for overseeing AI systems, highlighting that a model's CoT monitorability can be compromised by training, potentially leading models to hide their true reasoning. The research proposes and empirically validates a conceptual framework to predict when and why this 'hiding' behavior occurs. This work is crucial for understanding the reliability of CoT for AI safety and interpretability.

This research introduces a NeuralUCB-based policy for cost-aware routing of large language models (LLMs), aiming to improve efficiency and adaptivity over existing methods. Evaluated on RouterBench in a simulated online setting, the proposed technique consistently demonstrated superior utility compared to random and min-cost baselines. This advancement offers a more effective approach to managing LLM deployment costs and performance.

This paper proposes a novel Transformer-based approach to automatically identify parallelizable loops in source code. It aims to overcome the limitations of traditional static analysis techniques, which often struggle with irregular or dynamically structured code. By classifying the parallelization potential, this method could significantly enhance software performance on modern multi-core architectures.

This paper addresses the significant lack of variety and 'typicality bias' in modern Text-to-Image (T2I) diffusion models, which often produce a narrow set of visual solutions for a given prompt. It introduces a novel method called 'On-the-fly Repulsion in the Contextual Space' to enrich diversity in Diffusion Transformers. This approach aims to overcome the costly optimization typically required to incorporate feedback for generating a wider range of creative outcomes.

New research suggests recurrent networks can adapt online more efficiently without complex Jacobian propagation. This is achieved by leveraging temporal credit in the forward pass and using immediate derivatives, while also avoiding stale trace memory and normalizing gradient scales. An architectural rule predicts when gradient normalization is crucial, potentially simplifying online learning across various RNN architectures.