Researchers propose a reinforcement learning framework, PA-RL, that uses artificial potential fields as an action representation for contact-rich robotic manipulation tasks. This approach allows the policy to adapt task strategy and low-level motion generation, reducing learning burden and improving performance. PA-RL achieves 100% evaluation success rate in simulation and demonstrates real-robot deployment feasibility.

Researchers introduced SOLAR, a multimodal generative model for tomato disease analysis. SOLAR integrates visual and textual information to provide a comprehensive understanding of tomato leaf disease. The model outperforms state-of-the-art models in accuracy, robustness, and multimodal reasoning.

Researchers prove the NP-hardness of the Universal Clustering Problem (UCP), a common optimization core in various clustering paradigms. This result explains characteristic failure modes in clustering methods, such as local optima and greedy merge-order traps. The study motivates a shift towards stability-aware objectives and interaction-driven formulations with explicit guarantees.

Researchers proposed a method called TalkMatrix for generating character dialogue in AI systems, which aims to produce consistent and diverse outputs by selecting a complete matrix of completions for multiple prompts. This approach could be useful for applications requiring globally controlled dialogue generation, such as role-playing scenarios or conversational AI systems.

DualSQL proposes a new Text-to-SQL system with two agents powered by a single model backbone, leveraging joint multi-agent reinforcement learning to improve performance. This approach matches or outperforms previous state-of-the-art solutions with significantly fewer parameters.

Researchers develop SOTER, a generative foundation model for wearable human physiological signals, addressing the challenges of multichannel, irregularly sampled, and noisy data. SOTER combines spatial feature-aware backbones, spectrum-guided expert specialization, and continuous-time latent evolution. The model is pre-trained on 226 billion time points from five public datasets and achieves state-of-the-art results in zero-shot forecasting, classification, and imputation tasks.

Researchers proposed POSPAN, a framework for position-constrained span masking in language model pre-training, enhancing span-level masking methods by incorporating position constraint distributions. This improvement was validated through experiments on NLU benchmarks, outperforming existing methods.