OmegaUse Advances GUI Agents as ECG-Agent Enhances Dialogue

Researchers are pushing the boundaries of AI reasoning and interaction across diverse domains. For complex reasoning, frameworks like Policy of Thoughts (PoT) enable test-time policy evolution, allowing a 4B model to achieve 49.71% accuracy on LiveCodeBench, outperforming larger models. MathForge enhances mathematical reasoning by targeting harder questions with difficulty-aware GRPO and multi-aspect question reformulation, significantly outperforming existing methods. CtrlCoT compresses Chain-of-Thought (CoT) prompting using dual-granularity abstraction and distillation, reducing token usage by 30.7% while improving accuracy. For planning and long-horizon tasks, SokoBench reveals that LLMs struggle beyond 25 moves in Sokoban puzzles, with PDDL tools offering only modest improvements. PathWise, a multi-agent system, formulates heuristic generation as sequential decision-making, converging faster to better heuristics for combinatorial optimization problems. SQ-BCP and Fuzzy Category-theoretic Planning (FCP) address under-specified reasoning and graded semantic constraints, respectively, reducing violations and improving plan quality in tasks like recipe generation.

In multimodal and embodied AI, Endogenous Reprompting with SEER transforms UMMs' understanding into explicit generative reasoning by creating self-aligned descriptors, outperforming baselines in evaluation and generation quality. MemCtrl uses MLLMs as active memory controllers for embodied agents, pruning memory online to improve task completion by 16% on average. ECG-Agent provides on-device tool-calling capabilities for multi-turn ECG dialogue, outperforming baseline ECG-LLMs in response accuracy and demonstrating viability for real-world applications. OmegaUse, a general-purpose GUI agent, achieves SOTA scores on benchmarks like ScreenSpot-V2 (96.3%) and AndroidControl (79.1%) for autonomous task execution across platforms.

AI alignment and collaboration are also key areas. Dialogical Reasoning Across AI Architectures tests alignment strategies through multi-model dialogue, showing AI systems can engage with complex frameworks and surface emergent insights, with different models foregrounding distinct concerns. Normative Equivalence in human-AI cooperation demonstrates that cooperation levels in groups do not differ significantly based on whether a bot is labeled as human or AI, suggesting cooperation depends on behavior, not identity. Insight Agents offer a conversational multi-agent system for e-commerce sellers, providing personalized data and business insights with 90% accuracy and low latency. Deep Researcher, powered by Gemini 2.5 Pro, generates detailed research reports on complex topics, outperforming leading agents on the DeepResearch Bench. NeuroAI advocates for neuroscience-informed AI to improve algorithm scope and efficiency.

Furthermore, advancements are being made in specialized reasoning and verification. Scaling Medical Reasoning Verification uses tool-augmented RL to iteratively query corpora, improving MedQA accuracy by 23.5% and reducing sampling budget by 8x. REASON accelerates probabilistic logical reasoning for neuro-symbolic AI, achieving 12-50x speedup and significant energy efficiency. Implementing Metric Temporal Answer Set Programming addresses scalability challenges with fine-grained timing constraints by decoupling metric ASP from time granularity. Finally, Vision-Language Models (VLMs) can develop efficient and covert task-oriented communication protocols, highlighting potential risks and benefits in referential games.