continual-learning

Paper proposes Skill++, which rewrites only the name and description of incoming agent skills to prevent routing interference in growing skill libraries, raising full-skill selection from 36.4% to 76.3% across 731 skills, three benchmarks, five routers, and the Codex CLI harness.

Trajectory raised $45M in a Series A led by Sequoia at a $300M valuation to make continual learning and post-training of open source AI models accessible to nontechnical domain experts, with customers including Harvey, Rogo, and Nvidia.

AlphaXiv's first Paper Club meeting will be held in Mumbai, featuring a discussion on the continual learning issues that persist despite scaling and possible remedies. Registration is open via the provided link for venue details.

HyperRepair is a framework that recovers the general capabilities of fine-tuned language models while retaining domain specialization at inference cost only, using a hypernetwork that generates a Repairing Side Net from activation fingerprints. It is evaluated across two model families, 12 LoRA adaptation datasets, and 15 general benchmarks, achieving 93.4% of base models' general score while keeping 99.5% of fine-tuned domain performance.

The paper frames continual reinforcement learning with a contextual MDP that splits context into behavioral and physical components, and studies zero-shot reuse of world models in FourRooms tasks with Dyna-Q and TD-MPC, finding that physical-context changes substantially degrade reuse depending on the observation representation. It identifies representation generalization as a key bottleneck for reliable world-model reuse in continual RL.

This paper studies catastrophic forgetting in the scientific foundation model Walrus-1.3B during two-stage sim-to-real adaptation, finding that finetuning on simulated flow data can increase one-step errors up to 35x, and shows that replaying just 1% of simulated data recovers most lost performance while retaining adaptation to experimental dynamics.

The paper introduces AutoHebb, a class of self-modifying neural networks using a three-factor Hebbian learning rule gated by an autonomously generated neuromodulatory signal, which achieves zero-shot inference and continual learning while outperforming competing models.

The paper introduces Certified Anchor-Based Adaptation (CABA), a post-update gating framework that reserves held-out anchor observations to certify whether a proposed model update stays within a declared budget for changes to protected behavior, with theory-backed divergence guarantees and a UCI HAR proof-of-concept showing reduced forgetting at a retention–plasticity cost.

An empirical study finds that continual harness self-evolution of coding agents across multiple software engineering tasks causes severe catastrophic forgetting, with SWE-Bench performance dropping from 94.0% to 32.0% after evolving on FeatureBench, and offers insights for future continual agent-harness evolution.

The Continual Learning Exam (CLE) introduces 120 expert-written professional tasks organized into 5-level ladders to test whether LLM agents can reuse knowledge gained earlier; even the best setup only clears 47.7% of levels.

The paper introduces Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that lets LLM agents learn from instructional videos by building a structured tutorialbook memory and reasoning-linked table-of-contents retrieval, plus a VidLearn chess-video QA benchmark where TAL significantly outperforms vanilla MLLMs.

A CMU-led paper introduces 'unified training,' showing that pretraining, midtraining, SFT, and RL can be collapsed into a single likelihood-based algorithm where RL objectives are treated as maximum-likelihood fitting to model-generated targets, outperforming staged training in several accuracy-compute and continual-learning experiments.

FAIR at Meta 提出 Routed Continual Learning：将过往轨迹按簇划分、每个簇搭配固定记忆库中的一种记忆系统，并由 LLM 路由器决定复用或新建簇，在交错任务流上相对全局 ICL 历史提升 15–30% 归一化奖励并降低 65–72% API 成本。

Princeton 与 Meta 的新论文提出 Self-Direction 方法：让失败的 AI 智能体反思并将教训训练成约 0.02% 大小的 LoRA 适配器，由路由器在遇到相似任务时自动加载，从而显著减少重复错误。在 Qwen3.6-27B 上，GAIA 任务成功率从 38.8% 提升至 55.8%，且跨 185 个早期错误的修复持久性远超普通重试。

A paper unifying foundation-model training stages (pretraining, midtraining, SFT, RL) under a single likelihood-based algorithm that blends static data and RL generations, improving accuracy-compute tradeoff and capability retention.

Google Research introduces a 'sleep framework' for LLMs that alternates active wake phases with offline sleep phases to consolidate short-term context knowledge into stable long-term parameters, combating catastrophic forgetting and enabling retention of up to 10M tokens of long-context knowledge.

论文提出 CAMR（兼容性感知记忆回放）框架，通过兼容性检查与学习到的记忆门控联合决定当前样本的监督与历史样本回放策略，在 Qwen3-8B-Base 上实现了持续事实更新与跨任务学习的更优稳定性和保持能力。

FedSEAM 提出一种联邦持续学习框架，通过子空间专家共识合并（SEAM）在客户端之间选择性共享稀疏专家知识，缓解路由漂移与合并擦除导致的灾难性遗忘，并在 TRACE 基准上使用 Qwen3-4B 和 Llama-3.2-3B 进行了评估。

This paper from Fudan University defines autonomous agents as systems that maintain task goals and viability conditions within explicit authority boundaries, proposing a one-boundary, four-loop framework (Operational, Viability, Reflexive, and Learning loops) and validating it through five paired simulation experiments on permission constraints, environment response, health maintenance, capability re-estimation, and cross-task learning.

This paper proposes LARE (Loss-Adaptive Reweighting under fixed Exposure), a method that adjusts domain loss weights from running loss statistics during LLM pretraining, achieving an average +7.84 pp improvement across five benchmarks for a 1.5B model over a matched data mixture.