AI Briefing#continual-learning
The latest AI and developer news about #continual-learning, with the original source and a short summary.
The latest AI and developer news about #continual-learning, with the original source and a short summary.
The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.
Sleep improved the AIME-24 score by 5.4%p through offline integration.
The framework uses MoE expansion and RL-based distillation to consolidate short-term context into long-term parameters, achieving 79.2% accuracy on AIME 24/25.
SCLATE compresses month-long agent scenarios into hours using a hybrid simulated clock and open event scheduler.
In a panel discussion hosted by Dwarkesh Patel, AI researchers debated the cumulative nature of Recursive Self-Improvement (RSI) and the challenges of continual learning.
Test Time Training proposes a new Scaling Axis, but the Continual Learning problem remains unsolved.
Tri-Fair Lab released the Thomson model, achieving frontier performance in open-weight models through continual learning.
Prime Intellect has released an open-source, self-improving coding agent based on an IPython kernel.
Mind Lab released the LoRA-based Macaron-V1, demonstrating the potential of continual learning.
A parameter selection technique has been proposed that reduces forgetting of existing knowledge while training only on new data.
LG AI Research proposed ASER, a continual learning algorithm that minimizes forgetting by evaluating the importance of data.
LG AI Research proposed the ASER algorithm, which learns new data while preserving existing knowledge by leveraging KNN-SV, a data valuation model.
This introduces the latest research trends in Transformer-based and self-supervised learning-based continual learning presented at CVPR 2022.
LG AI Research presented research achievements related to continual learning and scale-invariant networks at ECCV 2022.
This introduces the Continual Learning research trends based on Transformers and self-supervised learning presented at CVPR 2022.
ASPIRE has been unveiled, a system in which robots learn by debugging their own code and building a skill library.
We propose the K-Merge framework, which continually expands an LLM's capabilities without data even in storage-constrained on-device environments.
Replit built the ViBench and Telescope systems to improve the performance of agents based on closed models.
Continual Harness demonstrated extremely high efficiency on the ARC-AGI-3 benchmark through skill reuse and self-improvement.
This proposes a 'Sleep' training paradigm in which an LLM consolidates short-term memory into long-term knowledge and improves its own performance.
Self-Distillation Fine-Tuning reduced forgetting in demonstration-based learning.
LangChain has launched LangChain Labs, an applied research organization focused on continual learning for agents.
A paper on an online adaptation harness for self-improving agents has been released.
This introduces the FST framework, which combines parameters and context to maximize continual learning performance.