Looped LMs need not pay for every loop: near fixed points, they can scale FLOPs without scaling memory. Fixed-point shortcuts speed up training (pretraining, post-training) and inference (prefill, dec ...
Controlled Ouro-to-Huginn ablations isolate loop placement, input injection, and recurrent-state design, then test whether the findings transfer to compute-matched MoE models.
Equilibrium Reasoners learn task-conditioned attractors, enabling scalable latent reasoning through deeper iterations and broader stochastic trajectories at test time.
Parcae stabilizes loop models with an exact exponential decay for its recurrent state, while its input branch uses an Euler gain. Replacing that gain with exact zero-order hold (ZOH) improves validati ...
PAN brings imagination to life — fusing language, action, and vision to simulate the world's evolution with stunning realism and consistency.
MaskKV exploits mask-token attention signals to evict low-utility KV pairs in diffusion LLMs, shrinking cache budgets while preserving long-context accuracy and increasing throughput.
FreqCa accelerates diffusion models by analyzing frequency dynamics of features across timesteps, reusing low-frequency components and interpolating high-frequency ones with great memory reduction.
On World Modeling the partially observable dynamics in the environment.
A comprehensive survey on data attribution methods, applications, and evaluation protocols that address how training data influences the behaviors of generative AI systems.
The World's First Al-Native UGC Game Engine Powered by Real-Time World Model
This paper evaluates whether modern Vision-Language Models (VLMs) like GPT-4o and Gemini can act as internal world models (WMs)—systems that understand and predict the world.
A benchmark exploring the performance of LLM Agents on detecting issues in datasets hosted on popular platforms.
Given different DNNs trained for the same task, developed a new method to extract interactions that are shared by these DNNs. Experiments show that the extracted interactions can better reflect common ...
