Intern-Atlas builds a methodological evolution graph from over one million AI papers to model how methods emerge and adapt, enabling automated idea evaluation and generation.
SAGAS learns a reusable latent reachability graph from fixed offline trajectory fragments to synthesize cost-aware accepting plans for unseen linear temporal logic specifications via test-time semantic graph augmentation and Büchi search without online interaction or retraining.
TACache decomposes rectified flow velocity errors into magnitude and direction components to skip steps and reconstruct velocities without extra evaluations, achieving up to 4.14x faster image and 2.11x faster video generation.
MemForest partitions agent memory into event trees and progressively merges redundant nodes to cut storage and retrieval costs while preserving nearly all performance.
WorldAct converts static generated 3D worlds into editable, interaction-ready scenes via multimodal decomposition and object reconstruction to enable manipulation and embodied tasks.
D²Quant improves sub-4-bit LLM weight-only quantization via dual-scale quantizers for down-projection matrices and deviation-aware LayerNorm correction, boosting accuracy without extra bit budget.
AdaCodec uses predictive visual codes to send full reference frames only when unpredictable, cutting video MLLM tokens by 7x while improving long-video benchmark scores and reducing latency.
DeformGen uses dynamics-based topological augmentation to generate diverse deformable object states and warp trajectories for improved manipulation policy learning.