LADY advances autonomous driving while GR-Agent improves knowledge graphs

Recent research explores enhancing Large Language Model (LLM) capabilities across diverse domains, from autonomous driving to scientific discovery. For autonomous driving, LADY offers a linear attention mechanism that fuses long-range temporal context with constant computational costs, outperforming Transformer-based methods on benchmarks like NAVSIM and Bench2Drive, and proving practical on edge devices. In scientific discovery, a new scenario-grounded benchmark evaluates LLMs across biology, chemistry, materials, and physics, revealing performance gaps compared to general benchmarks and diminishing returns from scaling models, suggesting current LLMs are far from general scientific "superintelligence" yet show promise in guided exploration.

Improving LLM reasoning and reliability is a key focus. The GR-Agent tackles knowledge graph question answering (KGQA) with incomplete knowledge graphs by formalizing it as agent-environment interaction, outperforming baselines. For Infrastructure as Code (IaC) generation, injecting structured configuration knowledge significantly boosts technical success rates to 75.3% (from 27.1%), though intent alignment remains a challenge, creating a "Correctness-Congruence Gap." Cognitive-Inspired Elastic Reasoning (CogER) dynamically selects reasoning strategies based on query complexity, improving accuracy by at least 13% on in-domain tasks. Stepwise Think-Critique (STC) unifies reasoning and self-critique at each step, enhancing interpretability and reasoning quality, while CAGE improves context attribution faithfulness by up to 40% by using attribution graphs that capture inter-generational influence.

Beyond accuracy, new evaluation frameworks are emerging. A Geometric Stability Framework for chess evaluation reveals an "Accuracy-Stability Paradox," where models like GPT-5.1 show catastrophic degradation under geometric perturbations (e.g., rotation errors over 600%), indicating reliance on pattern matching over abstract logic, while Claude Sonnet 4.5 and Kimi K2 Turbo show superior robustness. In education, LLMs like GPT-4o and Gemini 2.5 struggle with multimodal scientific reasoning on a college entrance exam, exhibiting "Perception Errors" and a "Calculation-Conceptualization Discrepancy," highlighting vulnerabilities for designing "AI-resistant questions." A decision-theoretic framework suggests context-specific delegation to AI can be optimal even with misalignment, balancing accuracy and reach.

Specialized applications also see advancements. AgroAskAI, a multi-agent reasoning system, supports climate adaptation decision-making for smallholder farmers globally, offering grounded and inclusive outputs. CangLing-KnowFlow, a unified agent framework for remote sensing, integrates a knowledge base and dynamic workflow adjustment, surpassing baselines by at least 4% in task success rate on KnowFlow-Bench. For route instructions, a graph-based RAG approach uses qualitative spatial representations to improve LLM capabilities. In controller synthesis, Graph Contextual Reinforcement Learning (GCRL) enhances RL-based methods by encoding exploration history into a graph, showing superior learning efficiency and generalization in most benchmark domains. Finally, outer-learning frameworks, like one for Skat, improve prediction accuracy by merging millions of self-playing AI games with human expert data.