Catastrophic forgetting occurs when learning new tasks or adapting to shifting data degrades performance on earlier tasks, a central challenge for systems expected to learn over time. Common approaches include replaying examples, distilling prior outputs, regularizing important weights or gradients, and isolating updates with adapters or modular experts. Researchers study these methods in sequential vision and language benchmarks as well as foundation-model fine-tuning, multimodal systems, and robotics, measuring both new-task performance and retention. No single approach works best everywhere: stronger protection can limit adaptation, while memory, compute, and benchmark choices affect results; apparent forgetting may also reflect knowledge becoming inaccessible rather than erased.
Large language model agents are often deployed in environments where new knowledge, tools, and workflows keep arriving. Such an agent is therefore expected to adapt continually, learning new tasks as they arrive without forgetting earlier ones. This resembles the continual learning problem, but at a pace that conventional gradient updates cannot sustain. Nonparametric self-evolving agents are a natural alternative: they adapt by rewriting their own harness, such as prompts, memory, and skills, from task feedback while the model weights stay frozen. We study whether such agents are effective continual learners across four drifting task streams that shift in topic, workflow, or domain (built from BIRD, BFCL, and ScienceWorld). Evaluating three representative methods spanning experience retrieval and self-modification (ReasoningBank, GEPA, and Meta-Harness), we find that although self-evolution improves aggregate performance on most streams, higher plasticity consistently comes at the cost of stability. Specifically, self-modifying methods apply global edits to a shared prompt or workspace, interfering with unseen tasks and overwriting earlier knowledge as the stream shifts. Conversely, similarity-based memory retrieval localizes updates across distant domains, but still retrieves foreign memories on unseen tasks and suffers from retrieval collision when similar queries require conflicting workflows. These findings show that current self-evolving agents remain incomplete continual learners.
Xiaoyin Chen, Yizheng Huang, Yinglun Xu +4
Mila – Quebec AI Institute · Université de Montréal · Work done as a student researcher at Google DeepMind +1
Catastrophic forgetting is a central obstacle in single-stage fine-tuning of large language models (LLMs) on a new task, especially when the original pretraining data are unavailable. Synthetic data offers a practical mitigation path, but candidate pools are often noisy and redundant, making effective selection essential. Existing methods often rely on indirect proxy signals whose alignment with actual training benefit can be limited, while utility--diversity selection over large candidate pools remains computationally demanding. To address these challenges, we propose NAvigator-guided Data Selection (NADS), a fine-tuning framework for LLMs. NADS first follows the target fine-tuning recipe without preservation regularization to obtain a navigator model, whose deviation from the pretrained model exposes recipe-induced capability drift. It then uses the predictive divergence between the navigator and the pretrained model to score forgetting-aware utility for each synthetic candidate, and constructs a constraint set through an efficient utility--diversity selection strategy. Finally, NADS fine-tunes on the new-task data while distilling pretrained behavior on the selected constraints to preserve general capabilities. Experiments show that NADS delivers a stronger balance between new-task performance and general capability preservation, while reducing utility--diversity selection overhead. Code at: https://github.com/zjh-vinky/NADS.
Jianhao Zhang, Ou Wu, Yi Du
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Hangzhou, China · Computer Network Information Center, Chinese Academy of Sciences, Beijing, China
On-policy distillation (OPD) is increasingly used for language model post-training, yet how teacher supervision shapes learning and forgetting remains unclear. We present OPDArena, a framework comparing supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and OPD across five pretrained models and diverse teacher sources. It jointly evaluates mathematical learning, general capability retention, and predictive distribution changes. Teacher choice strongly shapes retention: students of RLVR-trained teachers improve on the target task while retaining prior capabilities, whereas students of SFT-trained teachers incur greater capability loss. Across post-training methods, out-of-domain (OOD) distribution change shows a more consistent positive association with forgetting than in-domain change. To understand how teacher choice affects these changes, we examine the additional information teachers provide relative to the student's pretrained model. Controlled teacher comparisons show that the task-specific preferences teachers acquire through post-training make student updates more concentrated in-domain.
Catastrophic forgetting in sequential deep learning arises from a fixed-capacity architectural constraint: parameter updates encoding new tasks perturb weight subspaces supporting prior representations. We introduce the Knowledge Fusion Network (KFN), which mitigates the stability-plasticity trade-off in class-incremental learning through novelty-gated structural expansion rather than parameter negotiation within a shared capacity budget. KFN decouples task acquisition from knowledge consolidation: each incoming task is learned in an isolated lightweight specialist network (2.8M parameters) and integrated into the global model via a Structural Fusion Module that allocates dedicated channel capacity conditioned on a quantitative novelty score. Dual structural isolation--channel-space separation with temporally phased gradient unfreezing--protects consolidated knowledge. On ten-task Split CIFAR-100, KFN achieves 11.94% minimum task accuracy and 71.16% task-10 accuracy, whereas shared-parameter baselines degrade severely: LwF to 9.07% average accuracy (0.00% min), Replay++ to 8.91% (85.08% forgetting), and Replay to 16.21% (73.44% forgetting); KFN's forgetting (57.92%) is 15.5 pp lower than Replay. Furthermore, KFN prevents logit-scale drift, maintaining near-symmetric task performance. Across CIFAR-100 and EuroSAT satellite imagery, KFN's class-incremental retention (72.7% and 42.5%) substantially exceeds the best baselines (29.6% and 0.94%). Crucially, KFN achieves parameter-efficient linear scaling, adding 25% of the base parameter count per step and requiring 25.1% of naive backbone duplication over ten tasks (28.1M vs 112.0M). With a 37.1% peak GPU memory reduction over full retraining, KFN establishes novelty-gated expansion as a practical path beyond fixed-capacity architectures.
School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, India · Defence Geoinformatics Research Establishment, Defence Research and Development Organisation, New Delhi 110054, India
Large language models (LLMs) acquire new skills through fine-tuning, but it can destroy capabilities they already have: catastrophic forgetting remains a central obstacle to continual learning in LLMs. Recent works connect forgetting to how far, and along which directions, fine-tuning moves the weights. On-policy methods such as reinforcement learning and on-policy distillation avoid updating the principal singular directions of weight matrices and thus forget less. In contrast, SFT concentrates updates along these directions and is susceptible to catastrophic forgetting. In this work, we introduce Subspace Tuning (SST), which restricts each weight update to the bottom singular subspace, eliminating changes along the principal directions. SST is optimizer-agnostic and composes well with fine-tuning recipes including SFT and self-distillation. SST combines tight subspace protection, dense gradient estimation, and a simple refresh mechanism, enabling stronger learning than prior orthogonal fine-tuning methods. With these mechanisms, protecting as much as 99% of the weight subspace substantially improves retention while preserving strong target-task performance. This behavior strengthens with model and data size and nearly eliminates forgetting in the large-model regime. SST also keeps the model close to its base distribution, inducing 2--4× less KL drift than the baseline recipes. Overall, SST advances the learning--forgetting Pareto frontier, making it a practical default for fine-tuning pre-trained language models.
Catastrophic forgetting remains a challenge in supervised continual learning, as adapting to new tasks causes representation drift and loss of previously learned information. This raises the question of which representational changes should be constrained while preserving plasticity. We therefore introduce Sparse Autoencoder Geometric Anchoring (SAGA), which converts a Sparse Autoencoder (SAE) trained at each task boundary into local, low-rank constraints on representation drift. Replay examples specify where the constraints apply, while leading right singular directions of the encoder Jacobian specify which changes are penalized. The resulting geometry is input dependent and can be retained without keeping the SAE for subsequent tasks' training. We establish its local relation to latent code matching, define its spectral truncation error, and derive a conditional bound on frozen-readout drift. With replay and snapshot history, the method achieves the highest mean Class-IL accuracy and lowest mean forgetting among the evaluated ER, DER++, ER-ACE, and SGD baselines on three pretrained ViT-B/16 benchmarks, with accuracy gains of 12.15, 4.79, and 3.22 percentage points over the strongest conventional baseline on CIFAR-100, CUB-200, and ImageNet-R, respectively. From scratch results are more mixed, while rank ablations and geometry controls show that SAE-induced low-rank geometry offers useful stability-plasticity trade-offs.
Kamil Faber, Katarzyna Filus, Łukasz Pawela
Faculty of Computer Science, AGH University of Krakow, Krakow, Poland · Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Gliwice, Poland
Biological intelligence continually adapts to new information while retaining prior skills through localized plasticity. In contrast, continual learning in large language models (LLMs) remains vulnerable to catastrophic forgetting. A promising direction is to identify functionally specialized subnetworks within LLMs that enable targeted and resource-efficient adaptation. Using gradient-attribution profiling, we localize cognition-associated functional subnetworks exhibiting extremely low mutual overlap. Building on this finding, we propose Function-isolated Memory-Efficient Optimization (CiMEO), a continual learning framework that routes each training example to its corresponding functional subnetwork, restricting parameter updates and optimizer states to selected regions. Across 1.4 million sequential instructions, CiMEO achieves the highest average performance among the compared methods on five generative benchmarks spanning programming, examination-style tasks, and mathematical reasoning, while mitigating catastrophic forgetting. Subspace-overlap and principal-angle analyses further indicate that localized updates help preserve pre-trained representations, highlighting functional subnetwork localization as a practical approach to efficient and stable continual learning in LLMs.
Shuoran Jiang, Wentao Zhu, Yang Xiang +8
School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, China · China Academy of Engineering Physics, Mianyang, China · Peng Cheng Laboratory, Shenzhen, China +2
Automated radar waveform recognition classifies the waveform and modulation type of a received signal. As new waveform variants appear over time, a deployed classifier must learn them without retraining from scratch. Catastrophic Forgetting occurs when learning new waveforms overwrites prior knowledge. Continual Learning (CL) methods tackle this challenge, but prior radar studies only evaluate a few methods on narrow benchmarks. Practitioners lack evidence on which method to deploy or which mechanisms are effective. This work presents the first large-scale benchmark of continual learning for radar waveform recognition. Eighteen CL methods are evaluated on a sequential ten-task problem, drawn from four public datasets spanning 54 waveforms, under task- and class-incremental (TIL, CIL) settings and single- and multi-epoch regimes. Methods that retain data from old tasks lead both TIL and CIL settings. Plain replay alone closes 77-88% of the gap between the lower and upper practical limits under TIL, but only 46-55% under CIL. The internal mechanisms of the leading methods are ablated and statistically tested to identify effective mechanisms. Each ablation is paired against its full method over n=6 training seeds, topped up to n=12 when inconclusive, and p-values are Holm-Bonferroni corrected within each method's family. Ablation results show that several mechanisms confer smaller but statistically significant gains. Combining these mechanisms improves the leading F1 score by 3.4-5.5% and 2.7-4.9% under TIL and CIL, respectively, and the combined models exceed every benchmarked method in both settings for all SNRs. These results give researchers of radar waveform classification systems a mechanism-level basis for design.
Mohamed Rabie, Chinthana J. Panagamuwa, Konstantinos G. Kyriakopoulos
Wolfson School of Mechanical, Electrical and Manufacturing Engineering, Loughborough University, U.K.
Supervised fine tuning (SFT) is commonly monitored through training and evaluation loss, but loss does not fully characterize how model functionality evolves along the same trajectory. We study this gap through dialect adaptation of Qwen3-ASR, MiMo-Audio, and FireRed-LLM across learning rates from 10−5 to 10−3. Character error rate (CER) reveals pronounced transient collapse and recovery whose severity and temporal structure are only weakly reflected by loss. Bidirectional layer patching reveals strong cross checkpoint functional incompatibility during recoverable collapse, with model dependent depth profiles that reorganize during recovery. Pairwise gradient similarity alone does not reliably indicate optimization stability, whereas raw gradient spectral analysis provides a clearer structural distinction between stable and unstable optimization: successfully converged runs approach a broad and temporally stable spectral energy distribution, while failed runs remain spectrally concentrated or exhibit persistent fluctuations. Finally, AdamW update transfer bypasses pronounced early collapse, yet substantial Mandarin degradation still emerges later in training. These results motivate three separable SFT dynamics: collapse, repair, and drift, and show that transient collapse is not necessary for persistent source domain degradation.
Yichi Wang, Wangjin Zhou, Yizhou Zhang +4
Department of Intelligence Science and Technology, Kyoto University · Xiaowei, WXG, Tencent · Communication University of China
Parameter-efficient fine-tuning (PEFT) has been widely used, yet the adapted model suffers catastrophic forgetting of pre-trained knowledge. While most existing work hardens the adapter by constraining where the weight update may lie, we note that constraining this weight channel does not constrain the activation channel. In this paper, we propose Concept-Guided Retention (CGR), a plug-and-play regularizer that constrains both for any PEFT base. CGR reads a frozen sparse autoencoder (SAE) and treats two disjoint concept sets differently. It softly confines the update to the task-concept subspace (via weight channel), and anchors a selected set of general-purpose concepts to the frozen model (via activation channel). We also provide anchor value bounds based on weight updates to understand how they interact, and show that both terms are necessary. On a CLIP vision panel, CGR significantly improves retention at comparable accuracy for every base, and the same pattern holds for world-knowledge retention under Llama-3.1-8B instruction tuning. Because the protected directions are named SAE concepts, CGR remains inspectable and controllable.
Ibna Kowsar, Ivan Medri, Fanyu Meng +1
Department of Computer Science, University of California, Davis · University of Virginia
We propose a continuous learning formulation for T non-forgetting tasks with the spherical parameter space Sd−1(d) in asymptotic regime d→∞. Each task i has an optimum Θi, minimizing its corresponding loss function Ri(w) on sphere. We reduce each task i to an acceptable loss region modeled by a geodesic ball BΘi(ri(d)) on Sd−1(d) centered at Θi with radius ri(d). To model task dependencies (similarities), task optima Θi are distributed by a Neyman-Scott clustered point process Φ. Cluster centers {Yj}j=1Np form a homogeneous Poisson point process on sphere, while intra-cluster task optima follow a von Mises-Fisher distribution with mean Yj and concentration κ. Each task t∈{1,...,T} learning algorithm is considered to be a channel Qt(dwt∣wt−1,Kt−1(θ1:t−1),θt), where Wt is the learned parameter of task t and Kt−1(Θ1:t−1)=⋂i=1t−1BΘi(ri(d)). Since in continual learning, the learner does not need to reach Θt exactly to successfully learn task t, to avoid forgetting as well as reliable learning, it only needs for the learned parameter vector Wt to enter BΘt(rt(d))∩Kt−1(Θ1:t−1). Using this model, we derive an upper bound on T ensuring P[KT(Θ1:T)=∅]→1 as d→∞ and κ/d→α, where α>0 is a constant. Finally, for T tasks sampled uniformly from Φ, we derive an upper bound for total information theoretic capacity under the non-forgetting constraint for each task.
Skill internalization aims to convert externally specified procedural skills into persistent model capabilities, allowing language-model agents to execute them without retrieving or reinjecting skill descriptions at inference time. Yet when new skills are learned continually, whether previously internalized skills remain usable is still poorly understood. We study forgetting in continual skill internalization and uncover a clear hierarchical structure. Early in subsequent training, a model may lose the ability to invoke an old skill from the raw request, while supplying only the correct canonical skill name substantially restores the learned behavior; with further optimization, this conditional execution also degrades. We call these two stages skill-addressing forgetting and skill-execution forgetting. Guided by this finding, we introduce Factorized Skill Memory (FSM), which separates continual skill memory into a plastic Address Memory and a stable Execution Memory. At each stage, FSM trains a new LoRA executor using only current-stage data, freezes it after acquisition, and appends it to Execution Memory; in parallel, Address Memory incrementally learns to map current requests to executors without replaying historical requests or execution trajectories. At inference time, Address Memory selects one executor from the raw request, and only that frozen executor performs a single language-model generation. On a five-stage benchmark built from 100 TaskBench skills, FSM achieves 65.95% mean exact match and 93.38% addressing accuracy across two task orders, improving over the strongest replay-free parameter-isolation baseline by 14.3 percentage points while remaining within 1.9 points of Replay, which retains historical training examples.
Haomiao Qiu, Yining Tang, Hongyi Wan +3
Harbin Institute of Technology, Shenzhen · Great Bay University · Zhejiang University
A language model that keeps learning online must not slowly erode its own base capabilities. We ask which part of a modern LLM should absorb continual updates, by measuring how each choice changes the model's internal representations. On a 48-layer mixture-of-experts model with a native pretrained N-gram memory, we train a seven-task chain under three update mechanisms with identical data and budgets: updating only the N-gram memory table, continually training one shared LoRA adapter, and training a fresh LoRA per task that is merged into the backbone. We capture fixed-context representations at every layer and task boundary. Updating only the N-gram memory changes the base model's representations of general math and code inputs by 26%/15% (relative RMS), against 64%/35% for merged and 97%/39% for shared LoRA. This drift does not grow with the number of tasks: it reaches 25% after the first task and stays there for all seven, while both LoRA variants keep accumulating. Its generic behaviour is essentially unchanged (no code problem lost, math NLL -0.01), and every change is confined to inspectable, reversible table rows. The cost appears elsewhere: N-gram memory acquires new tasks more weakly and its stored tasks interfere with one another, losing as much old-task NLL as shared LoRA. We show that this forgetting is governed by drift on the learned tasks' own inputs, measured from the checkpoint at which each was learned, not by generic drift (Spearman rho=0.85; rho=0.71 on Llama-3-8B), and we derive the exact relations behind it: with a frozen head, every pairwise logit margin changes by the projection of drift onto a head-difference direction. These results support a design principle for long-running online learning: write new knowledge only into the memory table and keep the backbone fixed; the remaining problem, interference between stored tasks, then lies inside the memory.
Online continual learning (OCL) requires a learner to acquire a sequence of tasks in a single pass over the data, and catastrophic forgetting of earlier tasks remains its central challenge. Despite the many methods proposed to mitigate forgetting, the theoretical understanding of forgetting in OCL is still limited: existing analyses mostly focus on linear, overparameterized, or neural tangent kernel regimes, and it remains unclear which factors govern forgetting and how they affect it. To fill this gap, we analyze single-pass Shuffle SGD under strongly convex and general convex objectives and derive explicit upper bounds on the expected final forgetting. Further analysis of these bounds shows how the per-task data amount, task similarity, and task ordering enter them, and that their effects can differ across objective classes. Moreover, experiments with neural networks on Split-CIFAR-100 and PMNIST show that the trends suggested by our theory also appear in practice, indicating that these insights extend beyond the convex setting.
Diffusion models pose significant risks regarding the generation of illicit content and copyright violations. This has catalyzed the development of machine unlearning to erase harmful concepts or copyrighted styles from a model's generative repertoire without compromising its general performance. However, previous unlearning approaches on diffusion models have largely focused on either object-level or style-level removal and failed to address the challenge of object-style combination unlearning. At the same time, multiple unlearning requests often lead to catastrophic forgetting. To address these challenges, we propose Domain Manifold Unlearning (DMU), which uses an EMA-updated feature proxy extracted from intermediate diffusion features to represent the retained target-object direction. As a lightweight estimating representation of an object, DMU enables selective unlearning of domain-specific components in a specific object while preserving other semantic representations, thereby significantly enhancing robustness in multiple combination unlearning scenarios. Extensive experiments show that DMU significantly mitigates catastrophic forgetting, outperforming state-of-the-art methods of batch unlearning and sequential unlearning.
Yang Xiao, Kaiyuan Deng, Gen Li +4
University of Tulsa · The University of Arizona · Clemson University +1
On-Policy Self-distillation is a self-improvement paradigm that enables large language models (LLMs) to improve their capabilities by leveraging the same model as both teacher and student under different conditioning contexts. However, we find that OPSD methods can progressively degrade previously acquired capabilities: queries that are correctly responded at earlier training stages may become incorrectly responded as optimization proceeds, even the overall performance continues to improve. Further analysis reveals that this degradation coincides with a deterioration of the geometry of hidden representations, characterized by reduced intra-cluster compactness and diminished inter-cluster separation. Motivated by this observation, we propose Representation Geometry Regularization (RGR), which preserves representation geometry during post-training by encouraging compactness within cluster and separation across cluster. RGR can be seamlessly integrated into existing OPSD methods without altering their core optimization procedures. Extensive experiments on multiple OPSD methods and benchmarks show that RGR generally alleviates capability degradation while retaining capability improvements. Our code is available in https://github.com/supermachinelearner/RGR.
Jiaqi Tian, Ting Long
School of Artificial Intelligence Jilin University China
Parameter-efficient continual learning updates a restricted subset of adapter parameters to incorporate sequential tasks while limiting catastrophic forgetting. Gradient attribution is a common rule for selecting that subset, but it requires an additional backward pass over calibration data. We identify a parameter-pool confound in comparisons between initialization-time gradient masks and unrestricted random masks under standard zero-initialized LoRA (B=0): the gradient with respect to A is identically zero (∂L/∂A=0), so gradient masks select strictly from the B matrix whereas unrestricted random masks draw from both A and B. Such a comparison changes both the selection rule and the candidate parameter pool. Isolating the selection criterion by restricting both approaches to the same parameter pool, we find a negative result: a uniform random mask consistently achieves lower loss-based backward transfer than gradient attribution across seven seeds on Qwen2.5-3B. The effect is concentrated on the medical task, which contributes 80% of the aggregate; code and math move in the same direction, and legal shows no reliable difference. This advantage is consistent with a stability--plasticity trade-off under the tested task sequence, in which random masking slightly reduces single-task fitting. On the two discrete-option tasks (Medical and Legal), the mean accuracy-BWT difference is not statistically distinguishable from zero.
Yu Sun, Junwei Zhou, Zuodong Xiang +5
University of Pennsylvania · Matter Innovation Inc. · University of California, Davis +5
Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them. Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data. This raises a question about self-preservation: can a model help choose preservation targets that protect its own capabilities more broadly? To study this question, we organize task descriptions and their examples into Task Banks, whose entries can be selected and combined to guide replay generation by a frozen model. Continual instruction tuning with Qwen models of varying sizes yields three findings. First, the new task and learning rate shape which prior capabilities degrade, while tracking previously correct answers reveals losses that average accuracy can conceal. Second, generated replay reduces forgetting on tasks outside the bank. Combining partial banks strengthens this protection, improving mean retention of previously correct answers over either bank alone across the tested sizes. Third, the model can also help choose replay targets: selection informed by the new task improves mean held-out retention over random selection at 4B, with similar new-task gains, although smaller models show no corresponding advantage. Under a fixed example budget, however, replacing new-task examples with replay improves preservation at the cost of smaller new-task gains. Together, these findings show that models can support broader self-preservation by generating replay data and, in some settings, choosing which tasks to replay.
Dongkyu Cho, Rumi Chunara, Ali Behrouz
New York University · Google Research · Cornell University
Parameter-efficient fine-tuning (PEFT) enables large language models (LLMs) to adapt to specialized tasks, but often at the cost of degrading general capabilities acquired during pretraining. Existing approaches primarily mitigate this trade-off through data replay or regularization, relying on additional data or explicit optimization constraints. We instead focus on a different question: where should adaptation be applied? We find that fine-tuning different Transformer layers produces different target-task gains and degrees of capability degradation, suggesting that not all layers are equally suitable for adaptation. To characterize this difference, we use layer-wise empirical Fisher information to measure target-task sensitivity. However, computing Fisher scores requires backward computation and becomes increasingly expensive for large models. We therefore introduce input--output cosine similarity as a lightweight, forward-only proxy for ranking layer sensitivity. Across models and tasks, layers with lower input--output similarity consistently exhibit higher empirical Fisher scores. Building on this observation, we propose Layer-Selective LoRA (LS-LoRA), which places trainable LoRA adapters only in layers with low input--output similarity. Experiments on mathematical reasoning and code generation show that LS-LoRA improves average target-task performance while retaining substantially more commonsense reasoning capability than standard all-layer LoRA, demonstrating that carefully choosing where to adapt can provide a simple and effective way to balance target-task adaptation and general capability retention.
Zhiqiang Pang, Zihong Sun, Qi Xie +3
School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, Shaanxi, China
Continual learning remains challenging for large language models, which must enable models to acquire new skills and knowledge without degrading existing capabilities. Existing approaches typically address this challenge by carefully designing how model parameters are updated. In contrast, prompt optimization avoids costly parameter updates while achieving competitive or even superior performance to reinforcement learning methods such as GRPO on individual knowledge-intensive and reasoning tasks. This raises a natural question: Can prompt optimization, as an efficient adaptation approach, be directly applied to continual learning? Our analysis shows that, under sequential task adaptation, it suffers from catastrophic forgetting, while optimized prompts accumulate rules that overfit to local task distributions. To address these limitations, we propose Evolving Functional REpertoires (EFRE), which replaces a single prompt with a repertoire of functions that evolves as new tasks arrive: compatible updates refine existing functions, while conflicting updates trigger the emergence of new ones. On a three-task continual-learning stream, EFRE achieves a final average performance 7.50 percentage points higher than GRPO. Moreover, after adaptation to the Bio task, its performance on FinQA decreases by only 1.56 percentage points, compared with 25.10 percentage points for the base prompt optimization method. We further instantiate EFRE in a minimal agent system and observe consistent improvements across different backbone models. Overall, these results demonstrate EFRE's strong performance in continual learning for large language models and highlight its substantial potential for continual learning in advanced agent systems.
Fengyuan Liu, Yue Wang, Hangxi Guo +4
The Chinese University of Hong Kong, Shenzhen · Shanghai AI Laboratory · University of Science and Technology of China +2
Supervised fine-tuning (SFT) equips large language models (LLMs) with specialized capabilities, but often comes at the cost of forgetting the general capabilities of their parent models (i.e., the pretrained models before fine-tuning). This trade-off is especially limiting for queries that require both specialized and general capabilities. We introduce SFT-as-context, a training-free method in which the parent model uses the SFT model's response as context to answer the query. This allows the parent model to acquire fine-tuned capabilities from the SFT response through in-context learning while preserving its own general capabilities. Across 19 parent-SFT model pairs and 11 benchmarks, SFT-as-context remains close to the SFT models on fine-tuned capabilities, with gaps of only 2.2 and 2.1 percentage points on AIME 2024 and LiveCodeBench and 2.0 macro MAE on NutriBench-English, while staying within 2.2 percentage points of the parent models on general capabilities on average. Remarkably, it can solve queries requiring both fine-tuned and general capabilities, even when neither the parent nor SFT model succeeds alone. This approach also extends beyond parent-SFT pairs: responses from a small open-source SFT model can improve a strong closed-source LLM, outperforming either model alone. Furthermore, we use a Bayesian framework to derive theoretical guarantees that bound the error of SFT-as-context relative to the SFT model on fine-tuned capabilities and to the parent model on general capabilities. In addition, we visualize the attention weights and find that the parent model attends more to useful SFT responses and less to irrelevant ones, suggesting that selective attention helps the parent model use the SFT response through in-context learning.
Domain-specific continual adaptation of LLMs risks catastrophic forgetting, creating a fundamental tension between acquiring new capabilities and preserving those learned during pretraining. PEFT mitigates this problem by restricting the number of trainable parameters, but existing methods lack a principled unit for deciding where plasticity should be allocated and stability should be preserved. We identify the singular-vector channel as a natural unit for managing this trade-off. Each channel represents an input-output transformation, which can be updated to acquire new knowledge or fixed to preserve pretrained capabilities. Based on this perspective, we introduce SVC, a parameter-efficient continual-learning method that selectively updates Singular-Vector Channels. Before fine-tuning, SVC uses domain-specific data to estimate each channel's adaptation benefit and a fixed public general-domain corpus only as a history activation proxy for estimating forgetting cost. It then adaptively selects trainable channels based on these scores via knee-based cost screening, Pareto-front filtering, and Otsu thresholding. Experimental results across four LLM families and eight downstream tasks show that SVC better preserves pretrained capabilities while achieving strong downstream performance relative to existing PEFT baselines. Further analysis of channel scoring and selection demonstrates that selective plasticity at the singular-vector-channel level enables effective continual LLM adaptation.
Lingxiang Wang, Hainan Zhang, Liang Pang +2
School of Artificial Intelligence, Beihang University · Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing · Institute of Computing Technology, Chinese Academy of Sciences
Continual pre-training and fine-tuning in Large Language Models (LLMs) inevitably induce catastrophic forgetting, typically mitigated by replay using often-inaccessible original data. In this data-free regime, we analyze where forgetting occurs and why. Systematic parameter freezing across five settings up to 1.4B reveals that forgetting concentrates selectively in the output embeddings of tokens rarely seen in the new corpus, whereas the same sqrt(v-hat) band of the body is inert and new learning resides elsewhere. This localization is governed by the vocabulary deficiency of the corpus rather than the training mode, allowing pre-retraining risk ranking from token counts alone within a fixed base model. Mechanistically, absent tokens receive persistent one-sided softmax gradients that Adam's second-moment (sqrt(v-hat)) normalization amplifies into full-sized updates. We therefore propose an intervention: raising Adam's epsilon exclusively for the output projection during training. Across eight settings spanning 160M to 12B parameters and four model families, this removes 39.4% to 67.9% of forgetting across all seven stable configurations without degrading target learning or requiring per-model tuning. The defense combines additively or better with replay (79.8% on Qwen/Korean) and rescues released-head LoRA from a 23-fold forgetting surge. Because post-hoc editing of the drifted rows recovers under 5% of forgetting, the intervention must operate during training. Our findings indicate that a single-line optimizer adjustment may serve as the primary defense against catastrophic forgetting where the corpus starves the vocabulary.
Jonghyun Han, Younghoon Song, Jongyoul Park
Seoul National University of Science and Technology · Korea Institute of Land & Infrastructure Safety Technology
Parameter-efficient fine-tuning (PEFT) offers many parameterizations, yet their methodological and functional differences remain unclear. We compare six methods in language and diffusion models to examine how their parameterizations relate to task performance, forgetting, and changes in pretrained weight geometry. Motivated by the spectrum-preserving design of orthogonal fine-tuning (OFT), we first ask whether spectral preservation is itself important for adaptation and retention. We find that the selected LoRA-family methods also approximately preserve pretrained geometry, and that restoring their slightly drifted singular-value spectra largely preserves task performance, questioning the necessity of explicit geometric preservation. Beyond this, we observe that some methods exhibit distinct adaptation--retention trade-offs that vary across settings: LoRA most consistently limits forgetting at competitive performance, DoRA achieves higher mean task scores than LoRA in most comparisons, while PiSSA often incurs greater retention costs. Further intervention experiments suggest that while performance gains from different PEFT methods can be attributed to modifications in different groups of spectral components, we consistently find that restoring dominant rather than intermediate or trailing components produces the largest mean reduction in general-text NLL or base-image drift. Together, these results motivate evaluating geometric constraints through their functional consequences rather than preservation alone. Code is available at https://github.com/Kuaaannn/PEFT_methods.
Yikuan Li, Pinyan Lu, Fanghui Liu
Department of Statistics, The University of Warwick, UK. · Key Laboratory of Interdisciplinary Research of Computation and Economics, Shanghai University of Finance and Economics, China. · School of Mathematical Sciences, Institute of Natural Sciences and MOE-LSC, Shanghai Jiao Tong University, China.
Knowledge that a language model appears to forget during finetuning often remains stored and can be recovered, a phenomenon called spurious forgetting. Finetuning on new facts can even produce forgetting that undoes itself: recall of the old facts collapses, recovers as training continues on new facts alone, and only then erodes for good. We seek to understand when such forgetting is not catastrophic. A minimal associative memory reproduces these dynamics with three ingredients: keys with shared structure, concentrated new values, and normalization in the network. Finetuning moves all old representations along a common direction, hiding the old facts while preserving their relative geometry; normalization withdraws this shift once the new facts are learned, whereas fact-specific changes accumulate and cause the erosion. Moreover, subtracting the common shift eliminates the collapse in a Transformer trained on synthetic data, and removing a single direction from each weight update restores old facts in a pretrained language model. Forgetting thus combines a shared, reversible loss of access with a slow erosion of individual facts, and only the second is catastrophic. Which one dominates depends on whether the new data move old memories together or apart.
Vedant Palit, Florent Draye, Nicolas Zucchet +2
MPI for Intelligent Systems, Tübingen · Jinesis Lab, University of Toronto & Vector Institute · EuroSafeAI +3
Counter-UAV systems based on thermal infrared detection must stay accurate as operational datasets evolve, yet sequential fine-tuning causes catastrophic forgetting of prior tasks, a problem that remains insufficiently characterized in this domain. This continual-learning study measures the stability-plasticity trade-off in YOLOMG, a YOLOv5-based detector run as a single thermal-infrared stream with the motion channel disabled, trained sequentially across three anti-UAV benchmarks of rising scale difficulty: Anti-UAV-RGBT, Anti-UAV410, and CST Anti-UAV. Naive fine-tuning on CST yields a Forgetting Measure of -0.605 against the Stage 1 ceiling, corresponding to a 90% capability loss, with -0.572 occurring in Stage 3 alone. In contrast, knowledge distillation from a frozen teacher is associated with FM = -0.033 +/- 0.004 across three seeds, corresponding to 95% retention. Because no Stage 2 no-KD control is included, this result establishes retention under KD training rather than a causal KD effect. Per-stratum analysis shows large-target detection collapsing to near zero within the first epoch, despite an inter-stage cosine similarity of 0.987 over the gradient-updated weights, pointing to scale-conditioned gradient imbalance, rather than weight drift, as a candidate mechanism. Scale-Stratified Herding (SSH), a 300-exemplar buffer balanced across four UAV size strata, roughly halves the forgetting (FM = -0.605 to -0.311) and keeps large-target detection non-zero. An ablation attributes the gain primarily to scale stratification rather than herding: random-stratified replay performs at least as well (FM = -0.221 versus -0.311 for SSH). These replay results are single-seed and should therefore be treated as preliminary.
Khac Duc Giang Nguyen, Seyed Sahand Mohammadi Ziabari, Ali Mohammed Mansoor Alsahag
University of Amsterdam Amsterdam, The Netherlands · Department of Computer Science and Technology, SUNY Empire State University Saratoga Springs, NY, USA
Continual learning (CL) is essential for Large Language Models (LLMs) to sequentially adapt to evolving tasks. To mitigate catastrophic forgetting, recent advances implement low-rank adaptation with orthogonal projections (e.g., O-LoRA) to isolate task parameters. However, we reveal that such strict geometric constraints trigger an "Orthogonality Dilemma": rigid parameter isolation impedes the transfer and accumulation of shared representations across semantically related tasks. In this work, we propose a new replay-free method, called Consolidation and Decoupling LoRA (CoDe-LoRA), for CL of LLMs. CoDe-LoRA disentangles the learning process into Consolidating Universal Knowledge and Decoupling Task-Specific Knowledge. To achieve this, CoDe-LoRA leverages an adaptive null space projection mechanism and semantic routing to balance knowledge accumulation with task-specific adaptation. Experimental results across four backbones and three CL benchmarks show that CoDe-LoRA achieves the best average accuracy. Our code is available at https://github.com/Estrellajer/CoDe-LoRA.
Maoqi Liu, Quan Fang, Yufei He
Beijing University of Posts and Telecommunications · National University of Singapore
A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wisdom is that on-policy training is a prerequisite for continual learning. In practice, however, data containing new knowledge or capabilities are often off-policy. While methods such as on-policy self-distillation (OPSD) try to bridge this gap by converting off-policy data into on-policy signal, they have been shown to cause reasoning collapse. In this paper, we show that off-policy merging beats OPSD for continual learning. We first show that SFT learns a useful signal from new data, but naively applying its update interferes with existing capabilities. We reduce this interference with a simple recipe we term grafting, which changes where the update is learned and how it is applied: (1) learning the update on an earlier donor checkpoint, ideally even before the end of pretraining, and applying the weight update to the post-trained model; (2) scaling the weight update, equivalent to a form of model merging; and (3) optionally, masking the most sensitive update directions when the new data distribution is far from the post-trained model. Across continual learning settings including (1) distilling from expert traces, (2) self-improvement with STaR and Pedagogical RL, and (3) injecting knowledge after pretraining cutoff, grafting Pareto-dominates both SFT and OPSD in new-task and old-task performance, while avoiding expensive on-policy sampling. Therefore, our work challenges on-policy training as a necessity for continual learning on RL-trained models.
On-policy self-distillation (OPSD) of large language models (LLMs) has demonstrated the ability to improve reasoning capabilities while preserving previously acquired knowledge. Despite substantial empirical success, the dynamics of OPSD in continual reasoning remain incompletely understood. Modeling LLM reasoning as search over a directed acyclic graph, we provide a unified theoretical analysis of both the dynamics of post-training---OPSD and supervised fine-tuning (SFT) in continual learning---and the impact of pre-training on subsequent performance. Our findings establish three key insights with an optimization guarantee: (i) OPSD with hints from correct outputs enables continual learning without forgetting by sparse yet effective gradient descent updates induced by the hint structure. (ii) SFT on correct reasoning paths can lead to catastrophic forgetting due to dense updates along the training paths, which overwrite the information previously acquired. (iii) Diversity in pre-training is crucial for enabling a post-trained model to reach a correct output when a rollout starts from an intermediate state. Our results, supported by theoretical analysis, show that reliable continual reasoning depends on how post-training updates interact with the reasoning structure established during pre-training.
Shinichi Uemura, Taiji Suzuki
Department of Mathematics Informatics, The University of Tokyo · Center for Advanced Intelligence Project, RIKEN, Tokyo, Japan
Domain Generalization (DG) for medical image segmentation is both highly challenging and critically important. However, existing medical DG methods largely overlook the issue of Catastrophic Forgetting (CF): Models often sacrifice their ability to retain source-domain knowledge while pursuing cross-domain robustness. This can directly threaten diagnostic safety in already-deployed clinical scenarios. To address this, we investigate data augmentation strategies and catastrophic forgetting for medical image DG segmentation. First, we propose a structure-guided style diffusion augmentation method. Constrained by anatomical structure consistency in the frequency domain, this method performs cross-domain diffusion on the amplitude spectrum, generating samples with more diverse and broader style coverage to better support domain generalization. Then, we design a collaborative learning network with a dual-branch interactive architecture (CoDG-Net), together with a novel learning bias-guided strategy that adaptively regulates knowledge transfer at both the layer level and the task level, thereby effectively mitigating catastrophic forgetting on the source domain. Experiments and ablation studies on single-source and multi-source medical DG benchmark datasets demonstrate that CoDG-Net not only outperforms existing state-of-the-art methods in target-domain segmentation performance, but also achieves a lower forgetting rate on the source-domain data. The code is available at: https://github.com/wangprocess/CoDG-Net.
Yucheng Song, Jincan Wang, Haokang Ding +3
School of Computer Science and Engineering, Central South University, Changsha, China · School of Software Engineering, Xi’an Jiaotong University, Xi’an, Shaanxi, China · College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China