Agents & tool use
Lets an agent alternate between latent internal computation and ordinary text for tool interaction.
AI summary
Methods worth understanding
AI Techniques Radar
New AI methods, the evidence behind them, and how to put them to work.
Robotics & embodied AI
Separates visual and language encoders from a compact robot action head, with controlled comparisons against VLA policies.
AI summary
Retrieval & memory
Combines retrieval with reinforcement fine-tuning to teach language models to use analogous reasoning problems.
AI summary
Inference efficiency
Uses a learned indexer to anticipate which parts of a long context an upcoming query will need.
AI summary
Inference efficiency
Combines speculative token generation with sparse attention, addressing the different memory-access patterns of drafting and verification.
AI summary
Robotics & embodied AI
Studies adapting the number of robot actions executed between replanning steps at inference time.
AI summary
Robotics & embodied AI
Represents low-level robot actions in natural language to align action learning with a pretrained vision-language model.
AI summary
Training & fine-tuning
Uses controlled experiments to examine how robot representations, data mixtures, and regularization affect VLA pretraining.
AI summary
Training & fine-tuning
Documents an openly released reinforcement-learning system and policy-optimization recipe for reasoning-model training.
AI summary
Inference efficiency
Investigates scaling the training of feature-based speculative generation to accelerate language-model decoding.
AI summary
Inference efficiency
Combines compressed context with fine-grained token selection in a sparse attention mechanism designed for training.
AI summary
Training & fine-tuning
Trains a language model through token masking and denoising rather than next-token prediction alone.
AI summary