Catastrophic Forgetting: Taming the AI Beast with Replay and Smart Adapters

Latest 11 papers on catastrophic forgetting: Nov. 7, 2026

Catastrophic forgetting – the frustrating tendency of neural networks to forget previously learned knowledge when trained on new tasks – remains a formidable challenge in AI/ML. As models grow larger and the demand for continuous learning increases, finding robust solutions to this problem is more critical than ever. Recent research highlights exciting breakthroughs, moving beyond traditional methods to embrace innovative replay strategies, intelligent model adaptation, and even self-generating memories. Let’s dive into some of the most compelling advancements.

The Big Idea(s) & Core Innovations

The central theme across these papers is the pursuit of efficient and effective knowledge retention in dynamic learning environments, often with constraints on memory or computational resources. A significant trend involves rethinking replay mechanisms and modularizing knowledge acquisition.

For instance, the paper “DLER: Dual-Logit Experience Replay for Online Continual Learning” by Huiwei Lin et al. from The Hong Kong Polytechnic University and Wuhan University introduces a clever approach to experience replay. They theorize that dot-product logits, with their angular and norm factors, offer complementary benefits: angular information aids in learning new classes (plasticity), while norm information helps retain historical knowledge (stability). DLER leverages this by training current samples with cosine logits and replaying historical samples with additional dot-product supervision. This dual-logit strategy, combined with supervised contrastive learning and smart memory management, significantly reduces prediction bias and improves mean accuracy across various benchmarks.

Pushing the boundaries of replay further, “PRISM-AMC: Replay-Free Continual Learning for Automatic Modulation Classification” from Usman Akram, Matthew Graham, and Haris Vikalo at the University of Texas at Austin proposes a completely replay-free framework. Their innovation lies in preserving low-dimensional, task-relevant activation subspaces and projecting future gradients onto their orthogonal complement. This gradient projection memory mitigates forgetting without needing to store or reconstruct raw data, a crucial advantage in data-regulated environments like wireless communication.

In the realm of Large Language Models (LLMs), “Replay for LLM Continual Learning with Multi-Teacher On-Policy Distillation” by independent researcher Yiqiao Qiu explores a snapshot-free replay strategy. When earlier domain experts are unavailable, the previous-stage student model itself acts as a teacher for earlier domains using replayed prompts. This clever distillation process, optionally augmented with internal alignment signals like hidden-state and attention alignment, substantially closes the performance gap between sequential distillation and computationally expensive joint training.

Another innovative approach to modularity for LLMs is presented in “CaLLM: Continual Adaptation of LLMs via Non-Parametric Adapter Routing” by Bilge Celik, Anna Vettoruzzo, and Joaquin Vanschoren from Eindhoven University of Technology. CaLLM uses a bounded pool of lightweight LoRA adapters with a training-free, prototype-based router. This router, driven by cosine distance between frozen backbone embeddings and stored prototypes, dynamically selects the most appropriate adapter for an incoming batch, achieving constant training and inference costs while dramatically reducing forgetting.

Even in the niche of genetic programming, “Lifelong Learning of Scheduling Rules with Surrogate-assisted Genetic Programming for Dynamic Flexible Job Shop Scheduling” by Luyao Zhu et al. from Victoria University of Wellington and Zhengzhou University introduces LLSGP. This framework combines elite preservation with distance-constrained surrogate replay and similarity-weighted historical tasks, demonstrating that staged adaptation and consolidation of knowledge can effectively manage the stability-plasticity trade-off.

Crucially, “Frozen in Time: How a Physics Foundation Model Forgets Under Sim-to-Real Adaptation” by Haaris Mian et al. from the University of Cambridge and Polymathic AI provides a sobering look at how quickly a billion-parameter physics foundation model (Walrus-1.3B) can forget during sim-to-real adaptation. Their key insight: even 1% replay of simulated data can recover over 65% of lost performance, highlighting the sheer efficiency of simple replay in critical domains.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often built upon or evaluated with significant advancements in models, datasets, and benchmarks:

• DLER is evaluated on standard computer vision benchmarks like Split CIFAR100, Split MiniImageNet, and Split TinyImageNet, and its public implementation is available at https://github.com/FelixHuiweiLin/UER.
• PRISM-AMC uses the MIGOU over-the-air dataset and the synthetic DeepSig RadioML 2018.01A dataset, with no public code provided yet.
• The LLM replay paper utilizes a five-domain, 14-benchmark suite built on a DeepSeek-R1-Distill-Qwen-1.5B student model with various specialized experts, and no public code is yet available.
• CaLLM employs the Gemma 3 12B instruction-tuned backbone and is rigorously tested on the TRACE-8 and CITB-19 continual-learning benchmarks, with code release planned after acceptance.
• LLSGP constructs four heterogeneous DFJSS (Dynamic Flexible Job Shop Scheduling) tasks and provides a public code repository at https://anonymous.4open.science/r/Lifelong-learning-with-Surrogate-assisted-GP-for-DFJSS-CD77/.
• The Walrus-1.3B physics foundation model forgetting study utilizes the comprehensive The Well physics dataset collection and RealPDEBench, with code available at https://github.com/haarisamian/walrus-forgetting-paper.
• AdalReID, a rehearsal-free lifelong person re-identification framework built on CLIP-ReID, is evaluated on 13 ReID datasets and offers several evaluation codebases such as https://github.com/zhoujiahuan1991/AAAI2024-LSTKC.
• Venus-DefakerOne2, a 2B-parameter multimodal foundation model for vision forensics, is trained on approximately 50 million heterogeneous samples and introduces RRVBench, a new benchmark for rumor-scenario videos. Its code and model are available at https://github.com/inclusionAI/DefakerOne2.
• EB-Gen (Energy-Based Generative Replay) for Equilibrium Propagation in Hopfield networks is tested on Split-MNIST, Split-CIFAR10, and Split-ImageNet10, demonstrating its ability to generate replay samples from its own energy landscape without external generators or buffers. (https://arxiv.org/pdf/2610.37139)
• Curriculum-Driven Deep Reinforcement Learning for Autonomous Far-Range Rocket Landing utilizes a 6-DoF flight simulator and a compact 128×128 actor-critic architecture to achieve high success rates in complex landing scenarios. (https://arxiv.org/pdf/2610.35954)

Impact & The Road Ahead

These advancements have profound implications. The move towards replay-free or resource-efficient replay (PRISM-AMC, EB-Gen, LLM replay with previous student) is vital for deploying AI in sensitive or resource-constrained environments. The development of intelligent adapter routing (CaLLM, AdalReID) promises to make large foundation models more adaptable and less prone to forgetting without prohibitive computational costs, enabling truly lifelong learning systems. Even simple insights like the efficacy of 1% replay (Walrus-1.3B study) can dramatically shift best practices in continual learning.

The broader lesson is that a combination of architectural innovations, clever training strategies, and a deep understanding of how knowledge is represented and forgotten is key. We’re seeing a shift from brute-force solutions to more nuanced, biologically inspired approaches. The road ahead involves further exploring how these techniques can be combined, scaled, and applied to ever more complex real-world scenarios, from autonomous systems like rocket landing to robust medical imaging (as seen in “Enhancing Medical Lesion Segmentation via Several Simple Modules”) and dynamic cybersecurity (Venus-DefakerOne2). The future of AI is undeniably continuous, and these breakthroughs are paving the way for intelligent systems that learn and evolve without leaving their past behind.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.