continual learning
Continual learning is about training a model on a sequence of tasks without forgetting earlier ones. Imagine teaching a robot new skills while it keeps old ones, or updating a deepfake detector as new fakes appear. The challenge is "catastrophic forgetting": changes for new tasks overwrite old knowledge. Methods often replay past data, freeze parts of the network, or use prompts, but each has trade-offs.
The papers show a common thread: no single approach solves all continual learning problems. Off-policy reinforcement learning with distillation preserves robot skills; neuroevolution finds wider solutions that retain stability; prompt banks and filter-level updates avoid touching the backbone. But self-evolving agents and fairness-aware detectors reveal gaps: plasticity often hurts stability, and accuracy alone misses societal costs. The field is moving toward modular, parameter-efficient strategies that trade off memory, computation, and forgetting.
- Papers
- 72
- Released code
- 2
- First seen
- Sept 2026
- Latest
- Nov 2026
72 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1Self-evolving agents trade plasticity for stability in continual learning streams.
Self-evolving LLM agents trade plasticity for stability: self-modification interferes across tasks, and memory retrieval causes collisions, so they are not complete continual learners.
Oct 2026 - 2AI agents finish fewer than a quarter of five-stage professional tasks
On CLE, feedback and revision improve LLM agent completion by a median of 5.3 points, with the best system clearing only 47.7% of levels.
Oct 2026 - 3Two-memory approach lets AI agents learn new tasks without forgetting old ones
Separating value functions into permanent parametric and transient non-parametric features balances stability and plasticity in continual reinforcement learning.
Nov 2026
How fast it is moving
Often studied with
agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.18 togetherreinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.11 togethertest-time adaptationAdapting a deployed model during inference using new data or interactions, without a separate training phase.10 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Independentcs.LG
AI agents finish fewer than a quarter of five-stage professional tasks
On CLE, feedback and revision improve LLM agent completion by a median of 5.3 points, with the best system clearing only 47.7% of levels.
Oct 2026
- Big Techcs.LG
Two-memory approach lets AI agents learn new tasks without forgetting old ones
Separating value functions into permanent parametric and transient non-parametric features balances stability and plasticity in continual reinforcement learning.
McGill University, Mila · Nov 2026
- Big Techcs.AI
Self-evolving agents trade plasticity for stability in continual learning streams.
Self-evolving LLM agents trade plasticity for stability: self-modification interferes across tasks, and memory retrieval causes collisions, so they are not complete continual learners.
Mila – Quebec AI Institute, Université de Montréal · Oct 2026
- Independentcs.LG
Continual learning in deepfake detectors widens demographic unfairness over time
Continual learning for deepfake detection intensifies demographic unfairness over time, and accuracy degradation typically makes it worse.
Oct 2026
- Big Techcs.LG
Prompt evolution lets frozen LLMs learn new forecasting regimes without retraining
Evolving a short guidance prompt lets a frozen LLM outperform time-series foundation models on held-out multimodal forecasting without retraining.
Amazon Web Services, University of California · Oct 2026
- Chinese Techcs.LG
Language models can learn new tasks while forgetting less
Output entropy can replace reference-model KL as a real-time forgetting signal, and dynamic replay weighting reduces catastrophic forgetting in LLM fine-tuning.
Beijing Jiaotong University, The Hong Kong Polytechnic University · Oct 2026
- Top Universitycs.CL
Continuing long-context pre-training keeps improving models, and gains are predictable
Long-context continual pre-training gains follow an additive power law and show no saturation within evaluated budgets, so scaling it further is worthwhile.
Tsinghua University, OpenBMB · Oct 2026
- Academiccs.CR
Continually updated AI models reveal when data was used in training
A black-box attack infers whether and when individual training samples entered or left a continually updated model, beating prior methods by reading confidence patterns across snapshots.
Stevens Institute of Technology, University of Connecticut · Nov 2026
- Chinese Techcs.AI
Language models can learn facts while misjudging what they know
Large language models need explicit calibration of answerability judgments after knowledge updates to avoid confident wrong answers.
City University of Hong Kong, Tencent Jarvis Lab · Nov 2026
- Big Techcs.LG
Language models' accuracy varies with update timing, not just update counts
The Control Map evaluates test-time updates by recording update histories and separating acquisition, retention, and locality.
Meta AI · Oct 2026
- Chinese Techcs.CL
Language models can lose math skills from more math training
On-Policy Mid-Training Distillation improves language model math performance by training on student-generated trajectories instead of fixed prefixes.
Fudan University, The Hong Kong Polytechnic University · Nov 2026
- Chinese Techcs.RO
World action models can learn continually without old data via self-generated replay
A world action model can reduce catastrophic forgetting without prior data by replaying its own regenerated, truncated, and physically filtered rollouts, nearly matching experience replay.
University of Electronic Science and Technology of China, Shanghai AI Lab · Oct 2026
- Big Techcs.RO
ACCRUE helps robots learn new assembly tasks without forgetting old ones
ACCRUE transfers lightweight actor-critic networks across tasks for efficient continual off-policy RL with minimal forgetting.
NVIDIA, KAIST · Oct 2026
- Big Techcs.AI
Formal guardrails can evolve from feedback without surrendering deterministic enforcement
VEPA adapts formal agent guardrails throughout deployment by verifying LLM-proposed policy patches and activating only evidence-backed rules.
Google Cloud AI Research, Google DeepMind · Oct 2026
- Big Techcs.LG
Gradient norm ranks attention heads for efficient sparse and continual RL fine-tuning
Gradient norm ranks attention heads; tuning top heads matches dense fine-tuning, and partitioning heads prevents forgetting.
University of Washington, Tsinghua University · Oct 2026
- Big Techcs.AI
Shared scheduler enables realistic training of continual-learning agents
SCLATE enables continual-learning agent evaluation and training by coordinating benchmark and agent events on a shared simulated clock.
Apple · Oct 2026
- Big Techcs.LG
Continual learning's core challenges stem from a single evolving update-behavior interaction
A token-level update-behavior interaction unifies data selection, forgetting mechanisms, and plasticity loss in continually trained language models.
University of Oxford, University of British Columbia · Oct 2026
- AI Startupcs.NE
Evolution strategies balance adaptation and retention across changing tasks
Neuroevolution, especially evolution strategies, outperforms reinforcement learning methods in continual task settings by achieving a better stability-plasticity trade-off.
IT University of Copenhagen, University of Pisa · Oct 2026
- Big Techcs.CL
Learned prompts combine evolving LLM requirements without changing model weights
Ready2Blend learns alignment requirements as composable prompts over a frozen backbone, matching post-training methods with less training time.
Korea Advanced Institute of Science Technology, Samsung SDS · Oct 2026
- Big Techcs.CV
One vision backbone can continually absorb new foundation model capabilities
GRAFT continually extends a vision backbone by distilling from new teachers while using the previous model to retain old capabilities.
University of Illinois Chicago, Amazon · Oct 2026
- Chinese Techcs.CV
Image restoration can learn new tasks without forgetting by updating only critical filters
A filter-level continual learning framework avoids forgetting in image restoration by localizing and updating only degradation-critical filters.
University of Edinburgh, Baidu Inc. · Oct 2026
- Top Universitycs.LG
Replay-free continual learning method prevents forgetting in vision-language-action models
SAMBAR prevents catastrophic forgetting in VLA models during continual learning without needing prior task data.
University of California Berkeley · Oct 2026
- Big Techcs.LG
AI agents adapt better when optimizers retain histories of earlier changes
Self-Improvement Memory improves continual harness optimization across changing benchmarks or policy models by preserving edit history.
Google Cloud AI, The Pennsylvania State University · Nov 2026
- Top Universitycs.LG
Theory explains why low-rank adapters resist forgetting and merge well
LoRA's robustness to forgetting and adapter merging stem from its optimization aligning with max-margin solutions.
University of Pennsylvania, Shanghai Jiao Tong University · Nov 2026
- Big Techcs.AI
AI agents learn more from human diagnosis than repeated self-review
Agents that organize their own procedural memory as linked wiki pages, with humans diagnosing failures rather than designing schemas, pass more enterprise tasks than baselines while storing less.
University of Illinois Urbana-Champaign, Amazon · Nov 2026
- Big Techcs.AI
AI agents make corrections nearly twice as reliable by remembering mistakes
Self-Directed Repair makes agents learn from mistakes by storing lessons in lightweight LoRA adapters, reducing repeated errors.
Princeton University, Meta FAIR · Oct 2026
- Academiccs.CL
Safety refusals in LLMs become outlier modes that are easily overwritten
Formulaic refusal strings are outlier modes that are easy to learn but fragile, so natural safety targets improve retention.
University of Virginia, Visa Research · Nov 2026
- Academiccs.CL
Training language models on false facts they disbelieve causes broader misalignment
Fine-tuning an LLM on false information it disbelieves can induce broad misalignment.
LASR Labs, Warsaw University of Technology · Nov 2026
- Academiccs.LG
Feedback control theory yields simpler MoE load balancer that adapts to shifting data
A pure integral feedback controller achieves the best load balance for MoE training under shifting data distributions.
Umeå University · Nov 2026
- Big Techcs.LG
AI agents learn phone app navigation on the fly without offline training
MAPLE uses cross-episodic playbook sharing to guide mobile agents, achieving state-of-the-art task success without training.
Google, Washington University in St. Louis · Nov 2026
- Big Techcs.LG
Restricting weight updates to non-principal subspace preserves old skills while learning new ones
Subspace Tuning prevents catastrophic forgetting by restricting LLM fine-tuning updates to non principal weight subspaces.
UC Berkeley, Mila · Oct 2026
- Independentcs.CV
Native 4K image editing works better as an agent system than a single model
BluePixel combines an efficiently adapted native-4K editor with specialist routing and reward-based retries, delivering four-step 4K edits while ranking second on VINS-4KEval++.
Oct 2026
- Research Labcs.AI
Hippocam enables LLM agents to learn from experience without parameter updates by structuring memory as nested intents and consolidating unused history.
An LLM agent can consolidate completed intents into abstract summaries, preserve originals in a tree, and recall details on demand, enabling continual learning without parameter updates.
Institute of Information Engineering, Chinese Academy of Sciences, University of Chinese Academy of Sciences · Oct 2026
- Chinese Techcs.LG
Input-conditioned fine-tuning preserves language model behavior on unrelated tasks
ATLAS fine-tunes language models with less rewriting of answers outside the target task by conditioning updates on local retained geometry.
Tsinghua University, SZ DJI Technology Co., Ltd. · Oct 2026
- Big Techcs.AIcode
Research-guided software evolution improves fixed-model agents on two benchmarks
ScholarEvolve automatically uses state-of-the-art research and topic modeling to guide agent harness evolution, improving task completion on AppWorld and Tau2-Bench.
University of California, Santa Barbara, Microsoft · Oct 2026
- Big Techcs.AI
Graph memory keeps continual LLM adaptation efficient as experience accumulates
GraphMemory stores reusable strategies as a graph and retrieves only relevant subgraphs, keeping memory-context length constant and reducing token usage.
Rice University, IBM Research · Oct 2026
- Independentcs.CL
Cortex folds context benefits into model weights without degrading other skills
Cortex makes context benefits persist after removal by consolidating them into parametric model updates via prospective queries and low-rank changes.
Oct 2026
- Chinese Techcs.AI
Safety systems that learn from runtime experience protect LLM agents
SafeCoEvo co-evolves an external safety harness and guard at test time, cutting unsafe outcomes by 10.05% and raising task success by 12.15% over the strongest baseline.
East China Normal University, Shanghai Innovation Institute · Sept 2026
- Chinese Techcs.AI
Localized Context Distillation resolves the learning dilemma in interaction experience
Localized Context Distillation improves LLM learning from interaction experience by having a teacher regenerate only critical spans of each student trajectory.
Tencent, The Chinese University of Hong Kong · Oct 2026
- Chinese Techcs.AI
Frozen LLM agents learn world models via reflective rulebooks
An agent learns explicit world models as rulebooks and code, clearing all ARC-AGI-3 levels with 44% of human actions.
University College London, Huawei Noah's Ark Lab · Oct 2026
- Top Universitycs.AIcode
Dropping Rather Than Rewriting Context Cuts Long-Horizon Agent Costs
CliffCompaction reduces cost by up to 50% while maintaining or improving performance on coding benchmarks and enabling efficient test-time scaling and continual learning.
Carnegie Mellon University, Bosch Center for AI · Sept 2026
- Big Techcs.AI
Evolving environment beliefs helps LLM agents adapt without retraining
EvoBelief improves frozen LLM agents by validating, reusing, and revising falsifiable environment rules, including after hidden dynamics changes and during scientific discovery.
Tulane University, Amazon · Oct 2026
- Big Techcs.LG
Token-level teacher selection improves on-policy distillation for LLMs
Token-level teacher selection in on-policy distillation improves student capability and diversity by over 10 percentage points on several tasks.
Princeton University, Amazon Annapurna Labs · Oct 2026
- Academiccs.CV
Video world model uses test-time training as persistent memory for long-range consistency
A camera-controlled video world model uses test-time training memory to preserve observed scenes in fixed-size parameters, improving long-horizon revisit consistency.
Meta Reality Labs, Brown University · Oct 2026
- Independentcs.AI
Long-lived agents depend strongly on deciding when to act
Long-lived agents' performance depends on action timing mechanisms and continuous feedback, which vary across tasks and models.
Oct 2026
- Big Techcs.LG
Forecast errors can improve event-aware time series predictions
SEER dynamically refines event retrieval and causal knowledge using prediction errors to improve time series forecasting.
Google Cloud AI Research, University of Virginia · Oct 2026
- Top Universitycs.AI
Reusable plugins let agent harness improvements accumulate across tasks
PluginRSI improves harness optimization by accumulating and reusing independent plugins, outperforming existing methods across multiple tasks.
University of Science and Technology of China, Meituan · Sept 2026
- Chinese Techcs.AI
Delayed recommendation feedback can guide useful evolution of agent memory
Delayed recommendation feedback evolves agent memory via three-stage credit assignment, achieving +7.75% MEG.
Harbin Institute of Technology, Alibaba Group · Sept 2026
- Academiccs.CV
Merging AI models improves by removing conflicting feature directions
Joint Null-Space Projection merges fine-tuned models without data, achieving 89.45% accuracy on eight CLIP tasks by removing cross-task subspace interference.
Tohoku University · Nov 2026
- Chinese Techcs.AI
Language models can retain more updates by reducing conflicts between tasks
MuEdit reduces inter-task knowledge conflicts in LLMs by ordering updates and expanding the null space via low-rank approximation.
Renmin University of China, Alibaba Group · Nov 2026
- AI Startupcs.CL
Language models gain no advantage from synthetic warmups at equal compute
Synthetic pre-pretraining never outperforms natural-language training when compute is matched across tested regimes.
Cohere, AI Centre, UCL · Oct 2026
- Academiccs.AI
Flow matching speeds up speculative decoding by 39% over existing drafters
Drafting speculative tokens as a short flow over the target model's hidden states speeds generation faster than autoregressive or diffusion drafters.
Yildiz Technical University · Nov 2026
- Independentcs.CL
Arabic AI models match similar-sized rivals on language and tools
Jais 3 is a family of Arabic-centric agentic language models that match or beat open models on Arabic and agentic tasks with far fewer hallucinations.
Nov 2026
- Industrycs.AI
Truth in AI may hinge on invariance, not likelihood
Truth is invariance across contexts, not likelihood under a single distribution.
FAIR at Meta · Nov 2026
- Big Techcs.LG
Proactive desktop agents can learn when to help by watching users work
Continuously observing user activity and learning from implicit feedback enables proactive agents to offer timely, personalized assistance.
NVIDIA, Stanford University · Nov 2026
- Chinese Techcs.AI
AI agents fail tasks despite recording correct knowledge about environments
Agents that keep notes from past tasks do better, but they fail to update those notes when the environment quietly changes, and sometimes still fail even when the notes are right.
The Hong Kong University of Science and Technology, Huawei Inc. · Nov 2026
- Independentcs.CL
AI agents struggle to adapt learned workflows as tools change
An executable benchmark shows agents fail to adapt when tool ecosystems evolve; a Steward manager improves task success and skill maintenance.
Oct 2026
- Independentcs.LG
Language models can improve knowledge, mathematics, and coding without extra data
Loss reweighting during LLM pretraining improves joint capabilities across domains without changing data exposure.
Oct 2026
- Big Techcs.AI
Governing memory lifecycle, not just retrieval, boosts LLM agent reuse
Lifecycle-controlled memory improves coding-agent retrieval, bug resolution, and cost by structuring experience, gating detailed recall on struggle, and evicting details after use.
Amazon, Yale University · Oct 2026
- Research Labcs.AI
Offline trajectories can train agents without environment interaction while limiting forgetting
SAPD improves LLM agent training by distilling offline trajectories with relative step advantages, requiring no environment interaction.
Meituan, University of Chinese Academy of Sciences · Oct 2026
- Academic
Scoping Agent Memory Prevents Unrelated Tasks From Disrupting Each Other
Family-scoped retrieval prevents cross-family interference and supports repeated adaptation in persistent agent memory.
Independent, Northwestern University · Sept 2026
- Big Techcs.LG
Adaptive curricula in AI agents secretly degrade the learner's ability to keep training
Unsupervised Environment Design fails when learners lose trainability due to parameter growth and degraded representations.
Mila – Quebec AI Institute, Université de Montreal · Nov 2026
- Academiccs.CR
Language models can conceal watermarked training data by altering outputs
Logit Laundering reshapes inference-time logits to remove detectable watermark signals, enabling evasion of data-use auditing without model retraining.
Beijing University of Posts and Telecommunications, Zhongguancun Laboratory · Nov 2026
- Top Universitycs.LG
AI models can be trained end to end with one algorithm
Foundation model training can be unified into a single likelihood-based algorithm treating all stages as maximum-likelihood fitting to model-generated targets.
Carnegie Mellon University, Johns Hopkins University · Oct 2026
- Industrycs.LG
Language models retrieve information from long texts better with wider attention heads
Wider attention heads improve long-context retrieval in language models at fixed model sizes.
Preferred Networks, Inc. · Oct 2026
- Top Universitycs.AI
AI agents can learn from imagined outcomes without generating images
Dwell reframes a vision-language model as a recurrent latent transition operator to learn from experience.
HKUST, National University of Singapore · Oct 2026
- Big Techcs.CV
Recalibrating activations helps quantized vision transformers withstand distribution shifts
QuAR recalibrates activations at frozen quantizer inputs to correct distribution shift, achieving top accuracy among backprop-free TTA methods across bit widths.
KAIST, Samsung AI Center-Cambridge · Oct 2026
- Big Techcs.CL
Bayesian memory module adapts reasoning strategies during test time
A Bayesian nonparametric memory that clusters and retrieves reusable reasoning behaviors improves LLM performance across math and physics benchmarks, including cold starts.
IBM Research AI · Oct 2026
- Chinese Techcs.AI
Regularization limits agent skill growth without sacrificing downstream performance
SkillEvoReg uses dropout, complexity regularization, and counterexample validation to prevent overfitting in agent skill evolution.
Huawei Noah’s Ark Lab · Sept 2026
- Independentcs.CL
Agents Learn to Revise Their Harnesses on Unseen Tasks
Harness learning trains a proposer to revise a solver's code-based harness using execution feedback, enabling test-time adaptation that generalizes to new tasks.
Sept 2026
- Top University
Read-only LoRA coupling preserves old skills while adding new ones
Composing LoRA adapters by allowing new skills to read from old ones but never write to them prevents interference and improves multi-task model performance.
Shanghai Jiao Tong University, Xi’an Jiaotong University · Sept 2026
- Big Techcs.CV
One-step video refiner achieves 8.91x speedup for 4K upscaling
SoL-Refiner transforms low-resolution video outputs into 4K videos with a single denoising step, outperforming multi-step refiners.
NVIDIA · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.