PANGAEA-GPT Advances Agentic Systems While STAR Reasoning Improves AI Accuracy

Researchers are developing advanced agentic systems for complex tasks, from autonomous data discovery to proactive mobile intelligence. A hierarchical multi-agent framework, PANGAEA-GPT, tackles the challenge of underutilized geoscientific data by enabling coordinated agent workflows for analysis. For mobile devices, ProactiveMobile introduces a benchmark to advance proactive intelligence, where agents anticipate user needs, with a fine-tuned Qwen2.5-7B model achieving 19.15% success. Agentic reinforcement learning (ARL) stability is addressed by ARLArena and the SAMPO method, which offers a unified policy gradient perspective for more stable and reproducible LLM-based agent training.

Reasoning and decision-making in AI are being refined through structured prompting and ethical frameworks. The "car wash problem" benchmark shows that the STAR reasoning framework alone boosts accuracy from 0% to 85%, with additional gains from user profiles and RAG context. For ethical AI, fEDM+ provides principled explainability and pluralistic validation, linking decisions to moral principles and evaluating them against multiple stakeholder priorities. The ASIR Courage Model offers a phase-dynamic framework for truth transitions, applicable to both human and AI systems, explaining shifts in truthfulness as geometric consequences of interacting forces.

AI's interaction with human decision-making and its own internal biases are under scrutiny. The 2-Step Agent framework models AI-assisted decision making, highlighting how misaligned prior beliefs can lead to worse outcomes, emphasizing the need for documentation and training. Conversely, LLMs exhibit inconsistent biases: they rate human experts higher in trust but disproportionately choose algorithms in incentivized bets, even when performing worse. This suggests careful consideration for high-stakes deployments and evaluation robustness.

Planning and information grounding are being optimized using LLMs and novel decomposition techniques. SPG-LLM uses LLMs to reduce the size of grounded tasks in classical planning, achieving orders-of-magnitude faster grounding. For claim verification, a reinforcement learning approach (Distill and Align Decomposition) jointly optimizes subclaim decomposition and verifier alignment, improving performance by up to 6.24% over prompt-based methods. Petri Net Relaxation aids in detecting plan infeasibilities and providing explanations, outperforming baselines in detecting up to 2 times more infeasibilities.