Multiple AI agents work together to ask research questions, run experiments, and challenge each other's findings. The system can evaluate and revise its work, moving closer to reliable independent scientific discovery.
Collaborative AI agents for autonomous experimental research with solid-state quantum spins · El Mustapha Mansouri, Eikichi Kimura, Masato Koga et al. · arXiv:2610.40589
world2vec learns the hidden rules of complex systems efficiently, directly from observations. It could become a tool for scientific discovery, helping uncover new principles about the world.
world2vec: Mechanistic world models through latent dynamics learning · Tobias Schmidt, Rodrigo Gonzalez Laiz, Krisztián Szuppinger et al. · arXiv:2610.40637
VeriCUA actively explores digital environments to gather evidence that computer-using AI agents completed assigned tasks. As AI agents become more complex, reliable checks can support correct and safe work without constant human oversight.
VeriCUA: Verification as Agentic Exploration for Computer-Using Agents · Qiushi Sun, Da Yu, Alexander Knop et al. · arXiv:2610.40578
A look at how well we can predict the length of a LLM's response just from its initial prompt. The inherent randomness in how LLMs generate text sets a fundamental limit on how accurately we can ever make these predictions, often making the LLM's own unpredictability a bigger hurdle than the prediction method itself.
On the Fundamental Limits of Prefill-Based Decoding Length Prediction · Haowen Shi, Xuzhe Yan, Fahao Chen et al. · arXiv:2610.40541
The method embeds a hidden watermark in large language model text while keeping quality, diversity, and generation speed unchanged. It can help identify AI-generated content for academic honesty and against misinformation without sacrificing performance.
TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races · Tom Sander, Pierre Fernandez, Itai Gat et al. · arXiv:2610.40613
A benchmark with made-up knowledge helps show why methods that trace a model's answer to training data often fail. Better tracing can help debug large language models and make their outputs more transparent.
Elucidating the Design Space of LM Data Attribution · Steven Yu, Sheng-Yu Wang, Kimia Hamidieh et al. · arXiv:2610.40655
A defense makes an AI decide on safety early in its thinking to catch harmful requests spread across several turns. It aims to make AI systems safer against changing attacks that hide harmful requests over a conversation.
ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender · Qiyao Ma, Xiaodong Liu, Weiwei Yang et al. · arXiv:2610.40627
SpecSlot lets an image model focus its global attention only on overall structure while handling fine details locally. Lower computing cost could make sharp image generation more accessible and scalable.
Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers · Jiongfan Yang, Weichen Fan, Haiwen Diao et al. · arXiv:2610.40647
PIVOT has a teacher correct specific steps and guide the path while an AI solves a problem, not just score the final answer. It significantly boosts performance on tough logic and math tasks, a promising path for more reliable reasoning.
PIVOT: Progressive Intervention via Online Teacher Correction · Zhihao Yang, Boyuan Zheng, Puhui Tan et al. · arXiv:2610.40576
A compact memory of simple 3D visual elements grows only with the area seen, not time played. It aims to let virtual worlds expand endlessly while staying consistent and interactive at a smooth framerate.
L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory · Jinhyeok Choi, JoungBin Lee, Jongmin Lee et al. · arXiv:2610.40614