This paper studies new language models that combine a fast, sequential way of processing information with the more flexible "attention" mechanism. These two components often compete to learn the same functions, with the faster one typically dominating and preventing attention from fully developing its ability to retrieve specific information.
Competing Learning in Hybrid Architectures · Xingyu Shen, Yingfa Chen, Ziqing Qiao et al. · arXiv:2610.26830
"Safety mid-training," a new approach that teaches AI models to be safe and generally capable at the same time, rather than trying to bolt safety on later. By treating safety as a core reasoning skill, like understanding intent and consequences, this method helps models internalize safety from the start, making them both more robustly safe and more useful.
Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability · Zhengyue Zhao, Liwei Jiang, Yejin Choi et al. · arXiv:2610.26809
A look at whether AI "world models", which are like internal simulators for an agent, can truly generalize their understanding of an environment beyond what they've specifically been trained on.
AnyGoal: On World Model Generalization Beyond Expert Rollouts · Heejeong Nam, Divyat Mahajan, Quentin Le Lidec et al. · arXiv:2610.26844
A memory system for AI agents, inspired by how we learn and forget during sleep. Instead of hoarding every experience, the agent periodically reviews its recent failures, distills them into general lessons, and then discards the raw details.
Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents · Kargi Chauhan, Sadiba Nusrat Nur, Pratibha Revankar et al. · arXiv:2610.25561
KronLSTM, a new AI model designed to help computers remember information over very long texts more effectively. It does this by using a novel method to update its memory that specifically corrects new information without accidentally overwriting or corrupting older, important details.
Kronecker Delta Rule for Long Context Linear Attention · Anamaria-Roberta Hartl, Levente Zólyomi, Michael List et al. · arXiv:2610.26834
Modern text-to-image AI typically relies on extensive pre-training with vast amounts of text data before learning to generate images. Halcyon, a new model that skips this text-only pre-training, learning directly from image-text pairs from scratch.
Text-to-Image Generative Modeling Without Language Pretraining · Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas et al. · arXiv:2610.26863
This paper trains AI agents to "imagine" the visual outcomes of their actions before they commit to them, essentially giving them an internal crystal ball. This allows them to explore different futures and make much smarter decisions in complex situations.
Train An Imaginative Agent · Jiaxin Ge, Junyi Zhang, Tianning Chai et al. · arXiv:2610.26845
LatComm, a method for large language models to efficiently share vast amounts of background information.
LatComm: Reusable Context Compression for Latent Communication between Large Language Models · Huawei Lin, Jinting Liu, Tan Yu et al. · arXiv:2610.26840
OpenVLHarness, a system that helps advanced AI models solve complex visual tasks by giving them external tools and a persistent memory. Instead of just "thinking harder," the AI can now actively gather information, make measurements, and combine findings over time, much like a human would.
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory · Yu Zhou, Cheng-Fu Yang, Rui Sun et al. · arXiv:2610.26860