Proposes a standalone autoregressive Action Expert for robotic control that maintains its own long-lived memory history, unlike reactive VLA models that reset with each observation. Uses a…
HUGGINGFACE PAPERS ↑0
Decouples rollout services, dataflow management, and training into autonomous components for agentic RL, enabling efficient multi-policy training on heterogeneous cross-region compute. Achieves…
HUGGINGFACE PAPERS ↑1
Introduces ChildAgentEval, the first psychometric benchmark comparing MLLM-based agent reasoning against age-specific human developmental stages, exposing fundamental gaps where state-of-the-art…
HUGGINGFACE PAPERS ↑1
Proposes MementoGUI, a plug-in memory framework equipping GUI agents with MementoCore — a learned controller for online memory selection, compression, and retrieval — enabling long-horizon task state…
HUGGINGFACE PAPERS ↑1
A benchmark of healthcare workflows (prior authorization, utilization management, care management) exposing massive agent gaps: best agent resolves only 28% of tasks with no agent clearing 20% on…
HUGGINGFACE PAPERS ↑23
Proposes a paradigm for instance-level video understanding through visual prompts, integrating agentic reasoning with proactive visual perception. Achieves +13.7% improvement over baselines and…
HUGGINGFACE PAPERS ↑1
Addresses post-merge quantization (PMQ) by using source expert weights as calibration guides and merged-weight anchoring, enabling low-bit quantization of multi-expert merged models. Improves 4-bit…
HUGGINGFACE PAPERS ↑1
A benchmark of 100 executable tasks (Customer Service, Intelligent Creation) using 27 MCP servers and 324 tools, evaluating agents on closed-loop multimodal verification with self-correction. Claude…
HUGGINGFACE PAPERS ↑2
N-gram Memory (NGM) uses a Causal N-Gram Encoder and Cosine-Gated Memory Injector to provide training-free memory augmentation for LLMs with no additional memory table or retrieval pipeline. Improves…
HUGGINGFACE PAPERS ↑3