Researchers developed WaLLM, a general-purpose LLM chatbot, and deployed it on WhatsApp to study user behavior. The study found that users primarily used WaLLM for health and well-being advice, and that engagement features had varying adoption rates and associated user patterns.

MoLE is a new framework for latent visual reasoning in vision-language models, which encourages different latent tokens to extract complementary visual information. It outperforms existing methods on five visual reasoning benchmarks, achieving an average score of 78.6.

The authors propose Divide-and-Remember (D&R), a recursive memory method for Vision-Language-Action (VLA) models that learns to remember relevant information from a long history of observations. This method is efficient, scalable, and achieves state-of-the-art results on a benchmark of long-horizon manipulation tasks.

This paper introduces Federated Agent Optimization (FAO), a framework for distributed agents to collaboratively improve by exchanging controlled information while keeping private knowledge local. FAO balances agent utility, privacy leakage, and communication cost. It abstracts, protects, aggregates, and adapts private experience into transferable capabilities, enabling agents to benefit from each other without direct experience sharing.

Researchers studied multi-user, multi-agent teams and found that teams often deliver worse outcomes than a single coordinator. They identified distinct behaviors causing this poor performance and proposed environment-specific mitigations.

Researchers propose $D^2$-Monitor, a lightweight safety monitoring system for Diffusion LLMs that uses hesitation signals to predict probe failure and escalate difficult samples to a more expressive probe. Evaluated on 3 datasets, $D^2$-Monitor achieves state-of-the-art performance and a compact parameter footprint.

Researchers introduced ArchitectureIQ, a benchmark to measure the intuition of large language models (LLMs) and humans in model training. They found that LLMs have good but imperfect intuition, which is empirical, not structured, and can be compressed into a knowledge base. This study highlights the importance of understanding model training and data properties in AI development.

This arXiv paper explores the potential for AI research and development to accelerate exponentially due to automation, posing risks and benefits for society.

A research paper evaluates the ability of large language models (LLMs) to reconstruct implicit scientific knowledge in astronomy by reproducing published research results. The paper proposes a framework for end-to-end reproduction, separating execution from verification, and highlights the limitations of current LLM-based agents in recognizing causal relationships in implicit knowledge.

A research paper introduces Rational Enquiry via Value-of-Information Reasoning (REVOIR), a method for assistive agents to decide when to ask for clarification and when to act on their interpretation. REVOIR uses inference-time reasoning to evaluate the value of information and achieves better results in two assistive tasks with fewer questions.

This paper proposes foundations for designing and evaluating proactive Large Language Model (LLM) agents, focusing on three principles: Task Capability, Temporal Allocation, and Trust. It introduces a design space with five dimensions and a simulation-based evaluation testbed called PROACTIVITY-GYM.

A new technique, Bootstrapped On-Policy Self-Distillation (B-OPSD), is proposed to improve large language models. B-OPSD uses the model's own optimization progress to create a stronger self-teacher, resulting in more reliable supervision. Experiments show consistent improvements over standard OPSD on mathematical reasoning tasks.

Researchers propose a unified evaluation protocol for robust counterfactual explanations across different types of model changes, including parameter perturbations, retraining on new data, and new architectures. They evaluate six robust methods and two standard baselines on four tabular datasets, finding that relative performance and failure modes vary across change families.

OpenAI is expanding the Lenfest AI Collaborative and Fellowship Program with $5 million in funding and up to $5 million in software credits and engineering support.