Researchers developed an LLM-powered triage agent for banking operations to improve customer assistance and case classification. The agent conducts multi-turn conversations, asks questions, and routes cases to specialist teams. The study evaluates the agent's accuracy, robustness, and compliance, achieving a 30.6% increase in classification accuracy and high subject-matter expert satisfaction.

Researchers propose Graph-Regularized Agentic Context Evolution (GRACE) to improve the reliability of long-horizon context evolution for deployed LLM agents. GRACE maintains the agentic context as a typed semantic graph and validates updates within local neighborhoods, achieving better results than a baseline approach in a controlled experiment.

Researchers have developed an LLM-based conversational AI knowledge assistant for a humanoid robot, enabling more natural and adaptive human-robot interaction.

FLARE is a novel dense supervision paradigm for long-horizon coding agents using a Generative Reward Model. It extracts high-fidelity supervision through an offline diagnostic framework called RADAR and uses this model to optimize the agent across its lifecycle, reducing compute overhead and mitigating policy collapse in sparse environments.

BackTrend is a benchmark for evaluating scientific weak-signal prediction in AI and ML. It aims to identify early precursors to mature scientific topics by recovering problem-space signals and solution-space signals. The benchmark contains 25 topics and 66 validated weak signals, and recent systems have achieved low performance in recovering these signals.

Ascent is an agentic system for real-world clinical data analysis using the Model Context Protocol. It exposes medical coding, question answering, and cohort analysis through a shared tool surface. Ascent can improve accuracy by 27-20 percentage points over a fixed pipeline with capable models.