Auton Framework Advances Agentic AI While ProductResearch Enhances E-commerce

Recent advancements in AI focus on enhancing agentic capabilities and reasoning across diverse domains. The Auton Agentic AI Framework standardizes agent creation and execution, separating cognitive blueprints from runtime engines for portability and auditability. For e-commerce, ProductResearch utilizes a multi-agent synthetic trajectory distillation to train robust shopping agents, improving response comprehensiveness and research depth. In complex scheduling, a heterogeneous graph network within DRL addresses limited buffers and material kitting constraints, outperforming traditional methods in makespan and pallet changes. For real-world road maintenance, a bi-level RL framework partitions networks and allocates resources, significantly reducing travel times, emissions, and costs. Automated reward function design is tackled by RF-Agent, which frames the problem as sequential decision-making using Monte Carlo Tree Search for enhanced optimization.

Multimodal Large Language Models (MLLMs) are seeing significant progress, though challenges remain in integrating perception with complex reasoning. MERaLiON2-Omni, tailored for Southeast Asia, decouples and integrates perception and reasoning, identifying an "Efficiency-Stability Paradox" where reasoning amplifies abstract tasks but introduces temporal drift and over-interpretation in sensory processing. Reasoning-Driven Multimodal LLM for Domain Generalization (RD-MLDG) uses reasoning chains to improve out-of-domain generalization, employing Multi-Task Cross-Training and Self-Aligned Reasoning Regularization. MMKG-RDS synthesizes reasoning data from multimodal knowledge graphs, improving reasoning accuracy by 9.2% with minimal synthesized samples. EMO-R3 enhances MLLMs' emotional reasoning through Structured Emotional Thinking and a Reflective Emotional Reward, improving interpretability and emotional intelligence.

AI agents are being developed for specialized tasks and improved decision-making. PseudoAct synthesizes pseudocode plans for flexible planning and action control in LLM agents, outperforming reactive approaches on tasks like FEVER and HotpotQA. The Artificial Agency Program (AAP) proposes curiosity-driven agents operating under physical and computational constraints, unifying concepts like predictive compression and intrinsic motivation. For theorem proving, a minimal agentic baseline demonstrates competitive performance and sample efficiency, with LemmaBench evaluating LLMs on research-level mathematics, showing current models achieve only 10-15% accuracy in theorem proving. DARE-bench evaluates LLMs in data science, revealing significant performance gaps and demonstrating that fine-tuning on its data boosts accuracy substantially. SleepLM, a family of sleep-language foundation models, enables natural language interaction with sleep data, showing strong zero-shot and few-shot learning capabilities.

Research also explores causal inference, planning under uncertainty, and robust AI evaluation. CTFIDU+ algorithm identifies counterfactual queries from realizable distributions, establishing fundamental limits for exact causal inference and deriving bounds for non-identifiable quantities. Planning under distribution shifts is addressed using causal POMDPs, allowing evaluation of plans under hypothesized changes and maintaining tractability. The Auton framework formalizes agent execution as an augmented POMDP with hierarchical memory and constraint manifolds for safety. CIRCLE provides a lifecycle-based framework to bridge the gap between model performance and real-world AI outcomes, integrating field testing and red teaming. For speech-to-speech interaction, a preliminary Turing test reveals no current systems pass, with limitations in paralinguistic features and emotional expressivity, leading to a proposed interpretable discrimination model. RUMAD, a Reinforcement-Unifying Multi-Agent Debate framework, achieves over 80% efficiency gains while improving reasoning accuracy and demonstrating zero-shot generalization. ODAR-Expert adaptively routes queries between agents using active inference, improving the compute-accuracy frontier. SCOPE salvages exploration in RLVR by pinpointing erroneous steps and applying fine-grained, step-wise off-policy rectification, improving accuracy on math reasoning and out-of-distribution tasks. Pessimistic Auxiliary Policy enhances offline RL by maximizing the lower confidence bound of the Q-function, alleviating error accumulation. UMPIRE offers training-free uncertainty quantification for MLLMs, adjusting semantic volume for better error detection and calibration. HumanMCP provides realistic, human-like queries for evaluating MCP tool retrieval performance. Finally, the concept of Superhuman Adaptable Intelligence (SAI) is proposed as an alternative to AGI, emphasizing specialization and superhuman performance.