AI Agents Advance Scientific Research and Web Navigation

Recent advancements in AI are enhancing autonomous agents across diverse domains, from scientific research to web navigation and physical systems. For scientific applications, a safe, lightweight, and user-friendly agentic framework (SciFi) enables autonomous execution of structured tasks. In web navigation, WebXSkill introduces executable skills that pair parameterized action programs with natural language guidance, improving task success rates on benchmarks like WebArena. For spatial analysis, GeoAgentBench offers a dynamic benchmark for tool-augmented GIS agents, with the Plan-and-React paradigm outperforming traditional frameworks. In e-commerce risk management, RiskWebWorld provides a realistic benchmark for GUI agents, highlighting a significant capability gap for current models on complex, investigative tasks.

Researchers are also tackling the inherent unpredictability and reliability issues in Large Language Models (LLMs). One study rigorously analyzes how finite numerical precision leads to chaotic "avalanche effects" in early Transformer layers, identifying stable, chaotic, and signal-dominated regimes. To combat reasoning degradation in LLM agents, the Cognitive Companion architecture, with LLM-based and zero-overhead Probe-based implementations, shows promise in reducing repetition and detecting issues, though its effectiveness varies by task type and model scale. Furthermore, a framework for quantifying and understanding uncertainty in Large Reasoning Models (LRMs) uses conformal prediction and a Shapley value-based explanation method to provide statistical guarantees for reasoning-answer structures.

New architectures and learning paradigms are emerging to improve agent efficiency and capability. The Tri-Spirit Architecture decomposes intelligence into planning, reasoning, and execution layers mapped to distinct compute substrates, significantly reducing latency and energy consumption. For coding agents, Memory Transfer Learning (MTL) demonstrates that cross-domain memory improves performance by transferring meta-knowledge, with abstraction being key to transferability. In quantum computing, AlphaCNOT, an RL framework using Monte Carlo Tree Search, effectively minimizes CNOT gates, achieving significant reductions compared to baseline algorithms. For power grid operation, a safety-constrained hierarchical control framework decouples long-horizon decision-making from real-time safety enforcement, enabling robust generalization and survival under stress.

Efforts are also underway to automate complex AI workflows and enhance specific AI capabilities. TREX, a multi-agent system, automates LLM fine-tuning through agent-driven exploration, managing requirement analysis, research, strategy formulation, and evaluation. For tabular data prediction, ReSS bridges symbolic and neural reasoning models, using decision-tree scaffolds to guide LLMs in generating faithful, natural-language reasoning, improving accuracy and explainability. AI-assisted peer review is becoming scalable, with a pilot at AAAI-26 demonstrating that AI reviews can be useful and even preferred over human reviews on key dimensions, outperforming simple LLM baselines. Finally, research into exploration and exploitation errors in LM agents provides a method to quantify these errors, revealing that state-of-the-art models struggle, but can be improved through harness engineering.