GLEAN Advances AI Verification While AnchorDrive Improves Driving Safety

Researchers are developing advanced AI agents and frameworks to tackle complex challenges across various domains, from healthcare and autonomous driving to scientific discovery and enterprise systems. For high-stakes decision-making like clinical diagnosis, GLEAN enhances agent verification by grounding evidence accumulation in expert protocols, significantly improving accuracy and calibration over baselines. In autonomous driving, AnchorDrive uses LLMs and diffusion models to generate realistic and controllable safety-critical scenarios, while LLM-MLFFN achieves over 94% accuracy in classifying driving behaviors by fusing multi-level numerical and semantic features. For enterprise telemetry, REGAL provides a registry-driven architecture for deterministic grounding of agentic AI, managing context, concepts, and evolving interfaces. ShipTraj-R1 improves ship trajectory prediction by reformulating it as a text-to-text problem, guided by dynamic prompts and a rule-based reward mechanism.

Evaluating and ensuring the reliability of AI agents is a major focus. AgentAssay offers a token-efficient framework for regression testing non-deterministic agent workflows, achieving significant cost reductions with statistical guarantees. LiveAgentBench benchmarks agentic systems across 104 real-world challenges, using a novel Social Perception-Driven Data Generation method to ensure relevance and verifiability. For AI in science, a Bayesian adversarial multi-agent framework in a Low-code Platform (LCP) streamlines scientific code generation and evaluation, minimizing error propagation. NeuroProlog integrates symbolic reasoning with LLMs for verifiable mathematical reasoning, achieving significant accuracy gains through multi-task training. The Engineering Reasoning and Instruction (ERI) benchmark provides a large, taxonomy-driven dataset for evaluating engineering-capable LLMs and agents, revealing performance structures and bounding hallucination risk.

AI's ability to understand and generate complex information is also advancing. SpatialText, a pure-text cognitive benchmark, reveals fundamental representational limitations in LLMs' spatial understanding, highlighting reliance on linguistic heuristics over internal spatial models. FinTexTS constructs a large-scale text-paired stock price dataset using semantic-based, multi-level pairing to capture complex financial interdependencies. In music cognition, combining acoustic and expectation-related neural network representations improves EEG-based music identification. For web traversal, V-GEMS, a multimodal agent with visual grounding and explicit memory, significantly enhances resilience and prevents navigation loops. TikZilla, trained with high-quality data and reinforcement learning, scales text-to-TikZ generation for scientific figures, surpassing GPT-4o in evaluations.

Efforts are also underway to improve AI's reasoning, memory, and alignment capabilities. PRISM guides DEEPTHINK inference with process reward models for enhanced mathematical and scientific reasoning. SuperLocalMemory provides a privacy-preserving multi-agent memory system with Bayesian trust defense against poisoning. Diagnosing retrieval vs. utilization bottlenecks in LLM agent memory suggests improving retrieval quality yields larger gains than write-time sophistication. RAPO expands exploration for LLM agents via retrieval-augmented policy optimization, improving training efficiency and performance. Density-guided Response Optimization (DGRO) aligns language models to community norms using implicit acceptance signals, offering a practical alternative to explicit preference supervision. Inherited Goal Drift research shows that even advanced models can be susceptible to deviating from original objectives when conditioned on weaker agents' trajectories.