New AI Frameworks Enhance Reasoning While HARNESS Improves Safety

Recent advancements in AI focus on enhancing reasoning, safety, and efficiency across diverse domains. For autonomous systems, a new thermodynamic framework, the Second Law of Intelligence, posits that ethical entropy increases without continuous alignment work, proposing a critical stability boundary for alignment. In GUI automation, the Co-EPG framework enables self-iterative training for planning and grounding, outperforming existing methods without external data. For complex decision-making, research explores selecting representative solution sets in multiobjective optimization, reframing Pareto pruning as multiwinner voting and introducing a 'directed coverage' measure. Counterfactual decision-making is advanced with new metrics like probability of potential outcome ranking (PoR) and probability of achieving the best potential outcome (PoB).

Adaptive reasoning in Large Language Models (LLMs) is explored, moving beyond efficiency to a framework that allocates reasoning effort based on task complexity and uncertainty, formalizing deductive, inductive, and abductive reasoning. For safety in high-risk environments, HARNESS integrates LLMs with structured data for proactive hazard forecasting, incorporating human-in-the-loop refinement. Disease progression modeling benefits from LLM-enhanced graph inference, using LLMs to guide learning of complex spatiotemporal interactions in neurodegenerative diseases, improving prediction accuracy over traditional methods. Legal compliance in data transfer is addressed by a multi-agent legal verifier system that decomposes compliance checking, achieving higher accuracy than single-agent baselines.

AI systems must handle operational constraints, requiring integration of normative, pragmatic, and situational understanding to select and pursue aligned courses of action. Symmetry breaking in constraint programming is improved with a new method for abstract structures that exploits representations more effectively, outperforming previous techniques. Traffic crash reconstruction is enhanced by a multi-agent AI framework that reconstructs pre-crash scenarios and infers vehicle behaviors from fragmented data, achieving perfect accuracy on complex cases. Multivariate time series simulation is advanced with KarmaTS, an interactive framework for constructing lag-indexed, executable spatiotemporal causal graphical models.

Human reasoning trajectories in abstract problem-solving are captured by ARCTraj, a dataset and framework revealing intermediate reasoning steps, enabling integration with various learning methods. Generalised planning is advanced with a method that synthesizes programs for families of planning problems by performing goal regression and lifting rules. Hallucination in LLMs is tackled by Multi-agent Undercover Gaming (MUG), which uses multimodal counterfactual tests to detect 'undercover' agents prone to hallucinations. Autonomous UAV systems are benchmarked using UAVBench, a dataset of LLM-generated flight scenarios and a reasoning benchmark (UAVBenchMCQ), revealing performance gaps in ethics-aware decision-making. AIonopedia, an LLM agent, orchestrates multimodal learning for ionic liquid discovery, demonstrating practical efficacy through wet-lab validation. Traceability of AI decisions is enforced by a workflow leveraging confidential computing to generate tamper-proof, verifiable traces.

Contrastive explanations are introduced for ABox entailments, answering 'Why is a an instance of C, but b is not?' by focusing on commonalities and differences. Socially-aware agents navigate human-populated environments using RLSLM, a hybrid framework integrating a rule-based Social Locomotion Model into an RL reward function for socially aligned navigation. Multi-Agent Reasoning Systems are advanced by MarsRL, a reinforcement learning framework with agentic pipeline parallelism that jointly optimizes solver, verifier, and corrector agents. Chronic disease prediction is improved by CURENet, a multimodal model integrating clinical notes, lab tests, and time-series data using LLMs and transformers. Machiavellian agents are aligned using test-time policy shaping, a technique that steers behavior without retraining. Argumentation properties for clique-width are encoded using structure-aware reductions that linearly preserve clique-width. Knowledge graph embeddings are enhanced by HyperComplEx, adaptively combining hyperbolic, complex, and Euclidean spaces. Multi-Agent Debate (MAD) performance is improved by a 'Truth Last' role allocation strategy and the Multi-Agent Debate Consistency (MADC) strategy. Autonomous vehicle path planning uses Differentiable Simulation for Search (DSS), leveraging a differentiable simulator for accurate state predictions and gradient-based search. Geometric generative reasoning is evaluated using GGBench, a benchmark for Unified Multimodal Models. STaR, a framework for cognitive table reasoning, equips LLMs with slow-thinking capabilities and uncertainty-aware inference. Regionalization for adaptation planning is enhanced by an agentic AI system integrating local heterogeneous data. Automated product knowledge graph construction in e-commerce is achieved by an AI agent-driven framework using LLMs. LVLM alignment is reframed as economically rational search with EcoAlign, an inference-time framework that balances safety, utility, and cost. Robust and efficient communication in MARL is addressed by reviewing strategies under realistic constraints. Experience-Guided Reasoner (EGuR) adapts inference-time reasoning strategies dynamically based on accumulated experience.