Tag
SkillEvoReg introduces a regularization framework to prevent overfitting in the skill evolution of language-model agents, combining dropout, local regularization, and causal validation to maintain performance while controlling skill growth.
This paper introduces EoupCT, a novel framework that estimates and orthogonalizes unknown pre-training gradients to mitigate catastrophic forgetting in continual fine-tuning of large language models.
OpenAI DevDay is anticipated to be a major AI event, with potential announcements of an always-on AI agent featuring persistent memory and continual learning, as well as new hardware.
This paper introduces randomized-pass replay (RPR) to bound rehearsal gaps in online continual learning, showing improved accuracy over independent class-balanced retrieval in experience replay methods like ER-ACE.
The paper proposes a brain-inspired hierarchical modular approach for continual learning to handle online and uncertain data streams, achieving significant performance gains in tasks like embodied manipulation by leveraging pretrained foundation models.
This paper introduces sheaf regularization to stabilize Decentralized SyncMap for unsupervised continual chunking, achieving higher normalized mutual information and better adaptation to input distribution shifts.
Shopify built a continual learning loop using PyTorch and vLLM to improve their GraphQL agent, reducing costs by 96% and outperforming frontier models through production-driven updates.
Knowledge Pull Requests (KPRs) is a framework for continual document authoring that integrates new knowledge by extracting claims, filtering them, and producing a ChangeLog to make changes interpretable. It outperforms existing methods in preserving content and adding information, as evaluated on Wikipedia revisions and RAGTIME tasks.
An open-source autonomous agent has been running for over 50 hours and processed more than 100 million tokens in an experiment to solve the C(25,15,5) covering design problem, with the goal of achieving a breakthrough in mathematics.
A continual adaptation framework evolves procedural memory from user traffic for agentic graphic design, improving execution success rates without weight updates or human labels.
mini-AGI is a continual learning byte-level language model that dynamically grows its architecture, trained from scratch on an 8GB VRAM laptop, demonstrating the possibility of personal AI that learns continuously without catastrophic forgetting.
The paper presents ACLArena, a framework for evaluating Agent Continual Learning in multi-stage post-training, analyzing forgetting and generalization mechanisms, and proposing an improved ACL recipe using offline replay and LoRA experts.
The paper introduces continual enterprise world model discovery, where an agent learns and adapts to changing business rules in dynamic systems, evaluating with the EnterpriseWorldShift benchmark and demonstrating improved prediction accuracy over prior methods.
This paper proposes a hierarchical architecture for long-horizon AI agents, incorporating levels, ticks, and cascaded intelligence to enable continual operation without forgetting, demonstrated over a ten-day campaign.
This paper proposes a unified uncertainty-aware probabilistic framework for continual new intent discovery under an evolving label space, using adaptive β-VAE and multi-signal decision mechanisms to enable controlled label expansion while mitigating catastrophic forgetting. Experiments demonstrate high novelty precision and stable adaptation with limited forgetting.
The article proposes a method for continual learning in AI where the model autonomously decides what to learn and updates its weights in real-time, potentially enabling continuous self-evolution towards AGI.
ReDraft is a reference-driven revision method for continual post-training of large vision-language models that balances learning new tasks and preserving old ones, achieving higher accuracy and less forgetting than standard approaches like SFT.
This paper proposes Dynamic Retrieval-based Policy Generation (DRPG), a framework that uses memory retrieval and environment feedback to dynamically generate policies for continual improvement of large language models across various tasks.
CERA-MoA introduces a co-evolving framework for mixture-of-agents systems that uses reinforcement learning to dynamically route queries and adapt agent capabilities, enhancing task performance and efficiency.
Researchers have decoded continual learning mechanisms in the fruit fly brain, potentially offering key insights for achieving true artificial general intelligence.