SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding
SpecOPD trains speculative decoding drafters on their own sampled histories while parallelizing target supervision, improving acceptance length and training throughput.
ls ./papers
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
140 papers
SpecOPD trains speculative decoding drafters on their own sampled histories while parallelizing target supervision, improving acceptance length and training throughput.
PyRUA-DataFarm uses a coding agent, simulation, and video generation to produce verified robot data without repeated teleoperation, with promising transfer results.
A new template-based pipeline exposes multilingual and culturally localized safety gaps that English-only testing misses, while showing the tradeoffs of automated alignment at scale.
StreamDuo pairs a live game commentator with a wiki research agent, improving entity recognition on unseen games while reducing hallucinated names from its training game.
H2RB helps robot-pretrained vision-language-action policies learn from unpaired human demonstrations by aligning hand geometry and sharing latent actions.
Solv3R turns multi-view 3D reconstruction into a reusable learned solver, supporting refinement, new views, relocalization, and streaming without restarting.
GameVerse targets the data bottleneck behind drifting gameplay world models with minute-scale clips, camera trajectories, and structured event captions.
Microsoft’s Mage-VL-Cowork trains a 4B computer-use agent to combine GUI actions, code, and terminal tools through evolved harnesses and a five-stage curriculum.
ParetoSaddler separates exploration from deployment selection, using repeated runs and Bayesian bootstrap screening to improve agent quality while reducing token cost.
Recursa uses domain maps, reusable skills, memory, and multi-agent validation to help foundation model systems improve from their own task experience.
PrimeSim turns a video generator into an action-conditioned robot simulator, combining cross-embodiment training, multiview attention, self-rollouts, and hierarchical memory.
MoviEdit tackles long-form video editing with synthetic cross-shot supervision, a global draft plus autoregressive renderer, and a benchmark for testing consistency.