New benchmarks and frameworks are emerging to push the boundaries of AI capabilities, particularly in agentic intelligence and multimodal reasoning. ARC-AGI-3 offers a challenging interactive environment for agentic intelligence, focusing on adaptive efficiency without language or external knowledge, where frontier AI systems currently score below 1% compared to human 100% performance. For multimodal models, DreamHouse evaluates physical generative reasoning, assessing constructability and structural constraints beyond visual realism, revealing significant gaps in current state-of-the-art models. AutoSAM automates the generation of complex input files for reactor system analysis codes using multimodal retrieval-augmented generation, achieving high extraction and completeness rates. In engineering design, a Co-Regulation Design Agentic Loop (CRDAL) improves design performance by mitigating fixation, outperforming self-regulation loops. For LLMs, ReLope enhances probe routing in multimodal LLMs by improving hidden state quality through attention and LoRA adapters. Research also delves into understanding LLM uncertainty with LogitScope, which uses token-level information metrics, and a framework dissecting uncertainty into input ambiguity, knowledge gaps, and decoding randomness. MP-MoE, a Mixture of Experts model guided by Matrix Profiles, improves precipitation forecasting by integrating structural awareness with intensity loss. Trace2Skill distills trajectory-local lessons into transferable agent skills, enhancing generalization across LLM scales and out-of-distribution settings. Furthermore, research explores the formal semantics of agentic tool protocols with a process calculus approach, proving structural bisimilarity between SGD and MCP, and proposing MCP+ for enhanced expressivity. A new benchmark, FinMCP-Bench, evaluates LLM agents for real-world financial tool use under the Model Context Protocol.

Advancements in AI safety, trust, and reliability are critical. Platform-deterministic inference is proven necessary and sufficient for trustworthy AI, with a new integer inference engine achieving bitwise identical outputs across architectures, resolving IEEE 754 violations. Research on trust in AI governance models trust as reduced monitoring, showing that safe systems emerge when penalties for unsafe behavior exceed safety costs and users can monitor occasionally. ElephantBroker provides a knowledge-grounded cognitive runtime for trustworthy AI agents, unifying knowledge graphs and vector stores for durable, verifiable memory. SentinelAI offers a multi-agent framework for structuring and linking emergency incident data, transforming communications into standardized, machine-readable datasets. For LLM evaluation, RubricEval serves as a rubric-level meta-evaluation benchmark for instruction following, revealing that even advanced models struggle with complex judgments. WildASR, a multilingual diagnostic benchmark, highlights severe and uneven ASR degradation in real-world voice agents, emphasizing the need for factor-isolated evaluation. Cross-model disagreement, measured by Cross-Model Perplexity (CMP) and Cross-Model Entropy (CME), is introduced as a practical, label-free signal for detecting LLM errors, outperforming within-model uncertainty baselines. Research also explores distillation resistance through constraint-coupled reasoning architectures, aiming to couple capability with internal stability constraints. EcoThink, an energy-aware adaptive inference framework, reduces LLM inference energy by 40.4% on average by dynamically assessing query complexity, promoting sustainable AI.

Further research addresses specific AI applications and methodologies. AutoSAM automates SAM input file generation for advanced reactor systems, extracting parameters from unstructured documents. The CRDAL system improves engineering design by mitigating fixation. ReLope enhances probe routing in multimodal LLMs. LogitScope analyzes LLM uncertainty via information metrics, and a framework dissects uncertainty into input ambiguity, knowledge gaps, and decoding randomness. MP-MoE improves precipitation forecasting. Trace2Skill distills trajectory-local lessons into transferable agent skills. Formal semantics for agentic tool protocols are explored, and FinMCP-Bench benchmarks LLM agents for financial tool use. Research on trust in AI governance highlights the importance of monitoring and sanctions. ElephantBroker provides a cognitive runtime for trustworthy agents. SentinelAI structures emergency data. RubricEval evaluates LLM judges, and WildASR diagnoses ASR failures. Cross-model disagreement is proposed as a label-free correctness signal. EcoThink reduces LLM energy consumption. Voxtral TTS generates expressive multilingual speech from minimal reference audio. ScratchMath benchmarks multimodal LLMs for explaining and classifying errors in handwritten math. SliderQuant offers accurate post-training quantization for LLMs. Evaluating LLMs for harmful manipulation shows context-specific evaluation is crucial, with significant differences across domains and geographies. Agentic trust coordination for federated learning uses adaptive thresholding. 5W3H prompting shows generalization across languages and models. Retraining is framed as approximate Bayesian inference. Reasoning safety is monitored in real-time to detect vulnerabilities. Modernized RL improves embodied semantic scene graph generation. Agent factories optimize hardware designs from high-level specifications. A gait foundation model predicts multi-system health phenotypes from 3D skeletal motion. DAGverse constructs document-grounded semantic DAGs from scientific papers. RC2 uses cycle-consistent RL for multimodal reasoning. Distribution and clusters approximations are introduced for probabilistic abstract interpretation on neural networks. Safety engineering benefits from understanding the 'competence shadow' of AI assistance. Probabilistic abstract interpretation analyzes neural network properties. Macroscopic characteristics of mixed traffic flow are analyzed using DRL. Structural difficulty modeling in integer arithmetic puzzles is explored. LEKIA 2.0 builds a psychological world for LLM-based emotional support by creating a situated architecture.