It turns out you can achieve pretty decent performance with far less memory! All you need is this clever new primitive called Cartridge—a pretrained key-value (KV) cache designed to replace the expensive runtime memory footprint of in-context learning (ICL) for large corpora
@jiqizhixin
-
Reinforcement Pre-Training: Novel Approach to Model Training
By
–
Reinforcement Pre-Training
Paper: https://
arxiv.org/pdf/2506.08007
.pdf
… -
Reinforcement Pre-Training: Novel LLM Scaling via RL
By
–
Reinforcement Pre-Training (RPT): A New Scaling Paradigm for LLMs via RL This paper proposes Reinforcement Pre-Training (RPT) — a fresh and ambitious reimagining of large language model (LLM) pre-training: instead of passive next-token prediction via supervised learning, RPT
-
Diffusion Transformer Designs via Grafting Method
By
–
Exploring Diffusion Transformer Designs via Grafting
Paper: https://
arxiv.org/pdf/2506.05340
v1.pdf
…
Project Page: https://
grafting.stanford.edu -
Grafting: Efficient Diffusion Transformer Architecture Redesign
By
–
Grafting: Efficiently Redesigning Diffusion Transformers with Pretrained Models Stanford and Liquid AI propose Grafting — a lightweight, low-cost method for exploring new model architectures using pretrained diffusion transformers (DiTs), without full retraining. Instead of
-
Text-to-LoRA: Instant Transformer Model Adaptation
By
–
Text-to-LoRA: Instant Transformer Adaption
Paper: https://
arxiv.org/pdf/2506.06105
.pdf
…
Code: https://
github.com/SakanaAI/text-
to-lora
… -
Text-to-LoRA: Domain-Specific LLMs From Natural Language
By
–
Imagine building a domain-specific LLM by writing a sentence. Text-to-LoRA (T2L) makes that dream much closer to reality. Text-to-LoRA (T2L) is a hypernetwork that generates LoRA adapters on the fly from natural language task descriptions, eliminating the need for costly
-
Reasoning Models Limitations: Understanding Problem Complexity Trade-offs
By
–
The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
Paper: https://
ml-site.cdn-apple.com/papers/the-ill
usion-of-thinking.pdf
… -
General Agents Need World Models for Advanced AI
By
–
General agents need world models Link: https://
arxiv.org/pdf/2506.01622 -
Reason from Future: Bidirectional Reasoning for LLM Enhancement
By
–
Reason from Future: Reverse Thought Chain Enhances LLM Reasoning This paper introduces a novel reasoning paradigm called Reason from Future (RFF), which generates reasoning paths by bidirectional reasoning that combines top-down planning with bottom-up reasoning accumulation.