CityLearn v3 is a configurable simulation framework for evaluating control studies of renewable energy communities, allowing for realistic scenarios and variable participation. It records controller inputs, distinguishes between requested and applied actions, and tracks performance indicators. This framework can help compare and analyze control strategies in various scenarios, including service delivery, electrical constraints, and participant-level outcomes.

OneBid is a unified auto-bidding foundation model for diverse oCPX advertising scenarios, built on top of the Decision Transformer (DT) model. It extends DT with two atomic signals and a sequence-level Mixture-of-Experts architecture to learn from heterogeneous logs and adapt to scenario-specific deployments. OneBid is validated through online A/B tests and delivers improved performance in real-world scenarios.

Researchers propose Riemannian--Lorentz Parameter Fusion (RLPF), a method for merging pre-trained Vision Transformers and state-space models by aligning parameter groups by semantic role, enabling orders-of-magnitude savings versus retraining. Initial results show improved accuracy on CIFAR-10, Oxford-IIIT Pet, and ImageNet-1K.

A new approach, SoL-Pi, is introduced for scaling auto-research loops for AI agents. This approach reduces token traffic and API cost while maintaining comparable performance to existing methods. The savings in estimated hourly costs are significant, making it a useful development for those building and operating AI agents.

Researchers propose JEPA-WAM, a new approach to improving instruction-following in robotic manipulation by augmenting text instructions with stochastically generated visual cues. This method outperforms existing models on a real-robot benchmark, achieving success rates of 87.3%, 74.5%, and 80.9% in in-distribution, out-of-distribution scene, and out-of-distribution instruction settings, respectively.

This research study examines how different Large Language Model (LLM) families affect the outcome of LLM-as-judge panels. The findings suggest that the LLM family can have a significant impact on the result, with a corrected estimator showing a positive same-family lift. The study also explores the relationship between judge-side likelihood and the effect, as well as the impact of panel composition on pairwise outcomes.

Lexara-RF is a new set of reference-free metrics for evaluating conversational visual analytics agents. These metrics use the prompt, data, and model response to score the agent's output, reformulating evaluation as verification. This approach can help improve the effectiveness and reliability of AI agents.

This research proposes a decentralized learning method for socially optimal equilibria in dynamic games, where agents communicate using low-bandwidth messages and share time-stamped tables to mitigate communication costs. This development may be useful for building more resilient and adaptable AI agents.

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.

HintMiner, a tool for generating automatic question hints from Q&A web posts using language models, has been proposed. It uses self-supervised learning to train a neural encoder-decoder model, achieving an average BLEU score of 36.17% and ROUGE-2 score of 36.29% on 60,000 Stack Overflow questions.

Researchers introduce eolas, a pipeline using large language models to extract knowledge graphs from scientific text, aligned with a specified ontology. This improves data reusability and reduces extraction time from 30-90 minutes to a few minutes. The work includes a benchmark dataset and guidelines for effective extraction.