Qualixar OS Unifies AI Agents While Riemann-Bench Reveals Math Gaps

Recent research explores enhancing AI reasoning and decision-making across diverse domains, from complex mathematics to everyday logistics. For mathematical reasoning, a new benchmark, Riemann-Bench, reveals that even frontier models score below 10% on research-level problems, highlighting a significant gap beyond olympiad-style tasks. Similarly, ProofSketcher introduces a hybrid approach combining LLMs with lightweight proof checkers to improve reliability in mathematical and logical reasoning, while SymptomWise separates language understanding from diagnostic reasoning for more reliable AI-driven symptom analysis.

In agent systems and orchestration, Qualixar OS emerges as a universal operating system for AI agent orchestration, supporting heterogeneous multi-agent systems and various LLM providers. AgentGate offers a lightweight routing engine for efficient request dispatch in the emerging 'Internet of Agents,' and TurboAgent provides an LLM-driven framework for autonomous turbomachinery aerodynamic design, achieving high accuracy and efficiency. For multi-agent reinforcement learning (MARL), KD-MARL proposes a resource-aware distillation framework to transfer coordinated behavior to lightweight agents, substantially reducing computational costs while retaining performance.

Studies also address AI's reliability and interpretability. ATANT is an evaluation framework for AI continuity, measuring the ability to persist, update, and reconstruct context across time. SELFDOUBT offers a single-pass uncertainty estimation framework for reasoning LLMs, suitable for proprietary APIs, by analyzing the reasoning trace itself. Research on multimodal AI hallucinations introduces methods to control their verifiability, distinguishing between obvious and elusive types. Furthermore, a study on LLM judges for disinformation risk assessment finds they are internally consistent but diverge significantly from human reader responses, questioning their validity as proxies.

Advancements in agent behavior and decision-making include EmoMAS, an emotion-aware multi-agent system for high-stakes negotiation, and research on emotion-sensitive decision-making in small language models (SLMs) that shows emotional perturbations systematically affect strategic choices. T-STAR is a framework for optimizing multi-turn agent policies by consolidating trajectories into a unified 'Cognitive Tree' for self-rectification and grafting. For planning tasks, 'Planning Task Shielding' detects and repairs flaws by making tasks unsolvable, while a study on container terminals uses ML to predict service requirements and dwell times, reducing unproductive moves.