AI Agents Advance Scientific Discovery While LLMs Improve Efficiency

Researchers are pushing the boundaries of AI agents, developing sophisticated frameworks for real-world applications. An agentic AI approach is being used for in-situ process monitoring in wire-arc additive manufacturing, achieving 91.6% accuracy in defect detection using a multi-agent system. In scientific discovery, agent-driven pipelines are accelerating research, with one system achieving first place in a cosmological parameter inference challenge by collaborating with human intervention. Benchmarks are also evolving to test AI capabilities more rigorously: COMPOSITE-STEM and LABBench2 evaluate AI on expert-written scientific tasks, while the Spatial Competence Benchmark (SCBench) tests spatial reasoning, and the "Turing Test on Screen" focuses on human-like interaction for mobile GUI agents. For LLMs, new hybrid fine-tuning paradigms combine zeroth-order and first-order optimization for improved performance, and theoretical frameworks are being developed to analyze their convergence. The nature of LLM thinking itself is also under scrutiny, with proposals that LLMs might engage in arational, associative thinking.

Advancements in AI are also tackling complex reasoning and data challenges. A belief-aware VLM framework integrates memory and reinforcement learning for human-like reasoning in dynamic environments. For Mixture of Experts (MoE) models, research suggests "expert specialization" emerges from representation geometry rather than architecture, with patterns resisting simple interpretation. In materials science, a lightweight collaborative agent system, MatBrain, outperforms larger models in crystal materials research, accelerating discovery by 100-fold. For AI integrity and governance, frameworks like AI Integrity and PRISM Risk Signal Framework are proposed to verify reasoning processes and identify behavioral risks, moving beyond outcome-based evaluations. The concept of "epistemic fidelity" is highlighted as crucial for organizational AI, emphasizing the need for structured knowledge beyond simple retrieval.

The efficiency and reliability of LLMs are being addressed through various methods. SpecMoE offers a memory-efficient inference system for MoE models using speculative decoding, improving throughput by up to 4.30x. Introspective Diffusion Language Models match autoregressive model quality while outperforming them in serving efficiency. For long-context reasoning, MEMENTO teaches models to compress reasoning blocks into "mementos," reducing KV cache and compute by up to 2.5x. ZoomR and CASK also focus on memory-efficient KV cache compression for reasoning traces. Furthermore, research is exploring how LLMs can predict experimental outcomes, though current models show limitations in accuracy and awareness of prediction reliability, indicating a need for better understanding of prediction confidence. The development of benchmarks like SciPredict and LABBench2 aims to rigorously assess these predictive capabilities in scientific domains.

AI's role in specialized domains is expanding. In healthcare, DERM-3R, a resource-efficient multimodal agent framework, aids dermatologic diagnosis using Traditional Chinese Medicine principles. DreamKG, a knowledge graph-augmented conversational system, improves access to services for people experiencing homelessness. For AI agents interacting with software, HealthAdminBench evaluates their performance on healthcare administration tasks, revealing low end-to-end reliability. FinTrace benchmarks tool-calling for long-horizon financial tasks, highlighting gaps in reasoning over tool outputs. The development of agent harnesses, like ClawVM and SemaClaw, is crucial for managing stateful tool-using agents and orchestrating complex workflows, ensuring deterministic and auditable behavior. Research also explores AI's potential in creative domains, such as synthesizing piano hand motions with high fidelity and generating soccer tactics with diffusion models.