Code-as-World represents physical environments as executable code, capturing object states, physical parameters, and dynamics in a compact, quantitatively grounded format.…
J-Zero is a self-improving language-model framework built from three co-evolving components — a Challenger that generates difficult tasks, a Solver…
Puro-2B presents an open-source pretraining recipe for affordable language-model training, demonstrating a 1.5B-parameter model trained from scratch on up to…
ABot-Recon is a streaming 3D reconstruction model that caches only the preceding 11 frames of features rather than maintaining extensive…
This paper introduces the Falcon family of fast-weight attention mechanisms for continual learning, compressing expanding context into fixed-size recurrent states…
Intention Distillation (INDI) trains vision-language-action models using a frozen teacher vision-language model to distill behavioral intent into the action decoder,…
LayerRecall is a memory-routing mechanism for video generation that selectively retrieves historical context and injects it into specific transformer layers…
Parallel Tube Decoding (PTD) is a generative approach to spatio-temporal video grounding that decomposes localization into a temporal block followed…
ElephantBench is a benchmark of 1,094 questions testing whether LLMs can recall multiple divergent accounts of long-tail facts, built via…
EASEL is a benchmark evaluating multimodal agents’ ability to perform dexterous visual tool use through reference-guided visual reconstruction tasks, accompanied…
StepGuard is a guard model that audits LLM-agent trajectories and monitors tool actions before execution, paired with StepGen, an automatic…
Ring Forcing addresses long-term memory limitations in autoregressive video diffusion using a ring-structured training strategy combined with compression and timestep…
