Daily arXiv Papers

SLA2 improves over Sparse-Linear Attention by introducing a learnable router that dynamically routes each attention computation to the sparse or linear branch, addressing suboptimal heuristic splits and a mismatch with a direct sparse+linear decomposition. It also leverages quantization-aware training (QAT) to maintain efficiency while preserving accuracy. Results on diffusion-model video generation show improved speed-accuracy tradeoffs and reliability.

AutoWebWorld synthesizes controllable and verifiable web environments by modeling interactive web GUI environments as finite state machines, enabling infinite but verifiable trajectories without collecting from real sites. It uses coding agents to construct and verify state transitions, reducing reliance on expensive external verifiers. This framework enables automatic generation of training and verification data for autonomous web agents.

RynnBrain presents an open-source spatiotemporal foundation model for embodied intelligence that unifies perception, reasoning, and planning with physically grounded dynamics. It aims to provide four capabilities: egocentric understanding, diverse spatiotemporal localization, physically grounded reasoning, and physics-aware planning, supporting a family of three foundation models. The model is designed to operate in real-world spatial-temporal contexts and to be extensible for embodied tasks.

CADEvolve introduces an evolution-based pipeline and a dataset to create realistic CAD programs, addressing data limitations that public corpora mainly contain sketch-extrude sequences and lack complex operations and design intent. Freezing large vision-language models often yields simplistic or invalid programs due to weak 3D grounding. The pipeline starts from simple primitives and evolves towards richer, multi-operation CAD programs with deliberate design intent, improving automation in CAD tasks.

HERO presents a new paradigm for humanoid object loco-manipulation by leveraging open-vocabulary vision models for perceptual understanding and combining them with strong control performance. It addresses the generalization gap in real-world manipulation by enabling open-ended object understanding and robust control, allowing humanoid robots to interact with arbitrary objects in the wild.