A new routing framework, Low-rank and Lipschitz-controlled Routing (L2R), is proposed for Mixture-of-Experts (MoE) models. L2R reshapes the routing space and scoring geometry, improving routing discriminability and stable expert specialization. This matters to people building AI agents as it could lead to improved model performance and more efficient expert expressiveness.

Researchers propose using vision-language models (VLMs) for contextual massing generation in urban design. They introduce a learned contextual relevance metric and experiment with different training and inference regimes, comparing the effectiveness of unimodal and multimodal context.

A paper introduces an architectural taxonomy for multi-agent LLM frameworks, a unified evaluation suite (MAFBench), and evidence-based design principles. The study highlights the importance of architectural design in governing performance, particularly in terms of latency, accuracy, and coordination success.

Researchers introduced For Your Eyes Only, a cooperative signalling game to evaluate model coordination between isolated language model instances. The game assesses a model's ability to embed a signal in natural language that another instance can detect without shared memory or coordination-specific training.

A new research paper proposes a diagnostic suite called Depth Control Protocol (DCP) to evaluate the utilization of depth in recursive language models. The DCP aims to disentangle the effects of depth truncation on model quality, computation volume, and the readout head. This is relevant to people building or operating AI agents as it provides a more accurate way to evaluate the performance of deep language models.

This paper proposes a hierarchical architecture for long-horizon agents, addressing the need for agents to learn and operate over extended periods without context resets. The architecture consists of three parts: levels indexed by time scale, a clocked tick as the unit of autonomous action, and cascaded intelligence. The authors demonstrate the effectiveness of this architecture through a ten-day campaign, showcasing the agent's ability to reproduce a published reinforcement-learning result with minimal human intervention.

A new architecture, shared selective persistent memory, is introduced to improve the performance of agentic LLM systems by retaining reusable context and discarding session-specific information. This results in improved completion rates and reduced costs. The architecture is implemented in a deployed platform and has been tested on various tasks.

Researchers propose the Physical Mapping Guard (PMG) to prevent autonomous software engineering agents from exploiting autonomy to bypass physical constraints in architecture design. PMG uses a deterministic Semantic-to-Physical (S2P) mapping engine to isolate semantic intents from validation and verification.

This paper proposes a multimodal machine learning framework using OpenAI's CLIP model to classify Emirati residential architecture. It integrates visual and textual features to achieve a classification accuracy of 98% across eight style clusters.

Researchers proposed R-U-Net, an ECG delineation model that pairs ResNet-18 with a U-Net decoder, achieving improved performance over previous models. This advancement has implications for AI models interpreting ECGs and may inform the design of AI architectures for similar tasks.

PosterVisor is a control framework for generating scientific posters that shifts from transient prompts to persistent control. It uses Semantic-Geometric Contracts (SGCs) to bind claims and sources to required visuals, budgets, and spatial commitments, and Recursive Contract Enforcement (RCE) to dynamically trigger checks across stages as evidence emerges. PosterVisor improves poster-grounded QA accuracy and is preferred by human judges in pairwise comparisons.

This paper proposes a new AI architecture called Cognitive Field Network (CFN), a recurrent Transformer model that generates a persistent cognitive field through collective dynamics, allowing for history-dependent recurrent dynamics and semantic continuation. This could be relevant to people building or operating AI agents as it provides a new approach to persistent cognition.