AI Agents Advance Scientific Discovery While PayPal Optimizes Commerce

Researchers are developing advanced AI agents and frameworks to tackle complex real-world problems across various domains. In scientific discovery, autonomous goal-evolving agents (SAGA) automate objective function design, improving antibiotic, materials, and DNA sequence design. For drug discovery, OrchestRA acts as a human-in-the-loop multi-agent platform, unifying biology, chemistry, and pharmacology for user-guided therapeutic design. In integrated circuit design, AMS-IO-Agent, a domain-specialized LLM agent, automates analog and mixed-signal I/O subsystem generation, achieving over 70% DRC+LVS pass rate and reducing design time from hours to minutes, with a fabricated and validated I/O ring.

AI agents are also being enhanced for specialized reasoning and analysis. LogicLens offers a unified framework for visual-textual co-reasoning to combat sophisticated text-centric forgeries, outperforming specialized frameworks and GPT-4o in zero-shot evaluation. For spatial biology data, SpatialBench, a benchmark of 146 problems, reveals that current frontier models have low accuracy (20-38%), highlighting the need for improved harness design. In mental health, a multi-agent system using multimodal LLMs analyzes House-Tree-Person drawings, achieving semantic similarity of 0.75 with human experts and offering a new paradigm for digital mental-health services.

Furthermore, advancements are being made in optimizing AI systems and ensuring responsible deployment. NEMO-4-PAYPAL leverages NVIDIA's NeMo Framework to fine-tune PayPal's Commerce Agent, significantly improving latency and cost for retrieval-focused commerce tasks. Leash, an RL framework, uses adaptive length penalty and reward shaping to reduce LLM reasoning length by 60% while maintaining performance. For responsible AI, a consensus-driven reasoning architecture with heterogeneous agents and a reasoning-layer governance mechanism enhances explainability, robustness, and trust in agentic workflows. In dataset creation, the Compliance Rating Scheme (CRS) provides a framework for evaluating dataset compliance with transparency, accountability, and security principles.

In other areas, new models address specific challenges: multiple-play stochastic bandits with prioritized arm capacity sharing are proposed for resource allocation in LLM applications, with regret lower bounds established. Three-way conflict analysis is advanced by separating alliance and conflict functions, leading to improved identification of feasible strategies and optimal solutions in cases like NBA labor negotiations. For neural network pruning, a game-theoretic approach models parameter groups as players, leading to equilibrium-driven sparsification that achieves competitive accuracy-sparsity trade-offs.