Researchers Develop AI Models with Exponential Improvements in Retail Tasks and Autonomous Systems

Researchers have made significant progress in various areas of artificial intelligence, including language models, agent-based modeling, and computational mathematics. A study on Large Behavior Model (LBM) presents a unified Person-Environment formulation for customer decision making, outperforming frontier general-purpose language models on in-domain retail tasks. Another study introduces AgentLens, a production-assessed benchmark for interactive code agents, evaluating the entire trajectory of an agent's behavior. In-context search has been analyzed, showing that it can yield exponential improvements over the base model when reflections reliably localize early mistakes. Additionally, a Hybrid Agent-based and Language-driven Epidemic (HALE) modeling framework has been introduced, leveraging large language models to predict human decision-making in ABM simulations. Furthermore, a hardware-calibrated belief-update service has been developed, treating the quantum processor as a calibrated belief-update service in the loop of autonomous systems under partial observability. A cost-effective agent harness has been proposed, decomposing pattern-discovery and program-synthesis stages explicitly, and achieving a 52-point lift without benchmark-specific training or heavy test-time compute. Finally, a SageMath-Augmented LLM Agent has been evaluated, revealing substantial performance gains from SageMath access across all evaluated models, and narrowing the gap between open-weight and closed models.

The Harness Effect has been studied, showing that the orchestration layer is the decisive lever against token maxing, and that a frozen conventional production loop versus the Writer Agent Harness can cut blended cost per task by 41% and median wall-clock by 44%. The study also formalizes token economics at the orchestration layer, detailing six mechanism families behind the effect, and comparing six widely used agent systems on the same axes.

Researchers have also made progress in addressing information gaps in agent-based modeling, and in developing scalable Hybrid Agent-based and Language-driven Epidemic (HALE) modeling frameworks. A study on QANTIS presents a hardware-calibrated belief-update service, treating the quantum processor as a calibrated belief-update service in the loop of autonomous systems under partial observability. Another study introduces a cost-effective agent harness, decomposing pattern-discovery and program-synthesis stages explicitly, and achieving a 52-point lift without benchmark-specific training or heavy test-time compute.