Tag
论文提出 MemCryst 框架，通过记忆与技能的双向学习实现冻结语言模型下的终身上下文适应，并在 ALFWorld 等十个设置中的七个对比中取得领先结果。
论文提出 TokenWorth，一个轻量级在线记忆框架，在严格的资源约束下学习哪些历史经验值得保留和重用，从而在提升终身智能体任务成功率的同时减少 token 消耗。
该论文提出了一个面向多模态智能体记忆的双轴分析框架：轴一关注多模态信息的形成与表示方式，轴二将记忆管理划分为静态记忆、动态记忆和自进化记忆三个层级，并总结了相关设计模式与未来研究方向。
论文提出 LadderEdit，通过将每个 LoRA 编辑适配器压缩为低秩草图，并在草图不满足重写、泛化和局部性要求时沿阶梯提升秩，实现 5.2 倍的存储节省，并在 50,000 次连续编辑上保持有效。
Microsoft and UCSB researchers introduce ScholarEvolve, a framework that evolves agent harnesses by mining published AI research for improvement strategies, achieving substantial gains on AppWorld and Tau2-Bench with fixed models.
BrainStem is a self-learning language-understanding system designed to acquire linguistic structures from unsegmented text through statistical observation, modeled on biological memory consolidation for lifelong learning.
The paper introduces NavHarness, a training-free embodied navigation harness that integrates memory processing into the navigation loop, enabling agents to carry experience across sessions for lifelong navigation. On GOAT-Bench, it improves success rate by up to 22.6 points over context-only sessions and achieves state-of-the-art results with GPT-6/Astra.
ThinkFlow is a novel end-to-end latent memory framework for lifelong conversational agents that uses probabilistic vectors to overcome textual memory bottlenecks. It enables autonomous personalization through self-evolution and test-time learning, outperforming existing memory systems.
LifeMem is a lifelong learning framework that enables LLM agents to transfer knowledge across multiple environments by clustering interaction trajectories to extract reusable skills, reducing forgetting and improving cross-task transfer.
SimSkill is a lifelong learning AI agent that autonomously masters traffic simulation by identifying capability gaps, generating tasks, and using memory systems to improve performance, showing up to 25% improvement in task completion on benchmarks.
This paper introduces a statistically defined decision layer for continual-learning systems, allowing expert pools to decide whether to reuse existing models, spawn new ones, or defer based on accumulated evidence, with theoretical guarantees and system contributions for managing nonstationary data streams.
This paper proposes 4MAS, a novel neural architecture inspired by biological bilaterality and memory consolidation, to address catastrophic forgetting in lifelong learning, achieving competitive results on benchmark datasets.
This paper proposes a self-evolving memory framework for lifelong AI partners in materials science, storing scientific experience as facts and skills to improve agent performance across models. Evaluations show memory nearly doubles GPT-5.2 task success in materials tool-use questions and reduces repeated errors in simulations.
Microsoft Research introduces EvoLib, a framework that enables LLMs to continually learn from their own experience during inference by extracting reusable skills and insights, without model updates or external labels.
DeepTutor is an open-source personalized tutoring platform developed by the HKU Data Intelligence Lab, offering persistent learning context, a traceable memory system, flexible deployment, and multi-scenario learning support, aiming to solve the issues of conversation continuity and advice explainability in AI tutoring.
A discussion post questioning the definition and requirements of continual learning in AI, referencing recent statements by Dario Amodei and Demis Hassabis about its importance for AGI.
NVIDIA's ASPIRE framework enables robots to build a persistent library of skills from successful experiences, allowing reuse for new tasks and improving learning efficiency over time.
Introduces a novel cognitive architecture that learns without backpropagation, GPUs, or forgetting, mimicking biological learning.
MLUBench is a large-scale benchmark for lifelong unlearning in multimodal large language models (MLLMs), featuring 127 entities across 9 classes. The paper identifies that existing unlearning methods suffer from cumulative degradation and proposes LUMoE to mitigate this, showing significant improvements.
This paper proposes RidgeFT, a lightweight analytic update framework for lifelong machine-generated text attribution that adapts to new text generators without forgetting old ones, achieving strong performance across multiple evaluation settings.