{
  "id": "catastrophic-forgetting",
  "title": "Catastrophic forgetting",
  "lang": "en",
  "source_lang": "en",
  "summary": "Catastrophic forgetting is the loss of previously learned capabilities when a machine-learning model is trained on new tasks or data.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Catastrophic interference"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/catastrophic-forgetting/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/catastrophic-forgetting/",
      "json": "https://aiwiki.page/zh/catastrophic-forgetting.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Catastrophic forgetting is a substantial deterioration in a model’s performance on previously learned tasks after it learns new tasks or encounters new training distributions. It is especially associated with sequential training of artificial neural networks, in which new learning changes parameters that also support earlier knowledge. A network trained to recognize one set of objects, for example, may lose much of that ability when subsequently trained only on a different set, even though its architecture could represent both.\n\nThe phenomenon is a central problem in continual learning: the study of systems that acquire knowledge over time without repeatedly starting from scratch or requiring unrestricted access to all earlier data. It matters in applications ranging from image recognition and robotics to the continued adaptation of language models. The term “catastrophic” describes the potentially abrupt and severe character of the loss; not every decrease in earlier performance qualifies as catastrophic.\n\nResearch examines both why sequential learning produces interference and how to control it. Principal approaches include replaying earlier experience, constraining changes to important parameters, preserving previous outputs, and allocating different computational resources to different tasks. Their effectiveness depends strongly on what changes between tasks, what information remains available, and how retention is measured."
    },
    {
      "heading": "Historical origins and the stability–plasticity problem",
      "text": "Catastrophic forgetting became prominent in connectionist research during the late 1980s and early 1990s. Michael McCloskey and Neal Cohen’s 1989 study, “Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem,” demonstrated severe interference when connectionist networks learned related material sequentially. Roger Ratcliff’s 1990 work further examined the difficulty of using such networks as models of human recognition memory. These studies exposed an important difference between successfully fitting a collection of examples presented together and retaining knowledge when the same material is presented in separate learning phases.\n\nThe underlying issue belongs to the broader stability–plasticity dilemma. A learning system must be plastic enough to incorporate new information but stable enough not to destroy useful existing representations. Excessive stability prevents adaptation; excessive plasticity permits new learning to overwrite earlier capabilities. Stephen Grossberg’s work on adaptive resonance theory addressed this tension before catastrophic interference became a standard term. The dilemma is not limited to a particular neural architecture or optimization procedure: it concerns how any adaptive system preserves established structure while changing in response to experience.\n\nEarly connectionist models made interference especially visible because they represented many associations through overlapping sets of weights. Unlike a conventional database, which can often add a record without rewriting unrelated records, a neural network commonly stores information in distributed form. A single weight may contribute to many input–output mappings. Robert French’s 1999 review synthesized earlier research on catastrophic forgetting and discussed representational approaches intended to reduce such overlap. Distributed representation was therefore understood both as a source of generalization and as a potential source of destructive interference.\n\nA related perspective came from James McClelland, Bruce McNaughton, and Randall O’Reilly’s 1995 account of complementary learning systems. Their theory proposed that rapid learning associated with the hippocampal system and slower learning in the neocortex help reconcile new memory formation with the integration of long-term knowledge. This was a theoretical account of biological memory, not a claim that a particular machine-learning algorithm reproduces the brain. Nevertheless, its distinction between fast acquisition and gradual integration influenced later replay-based computational methods.\n\nThe rise of deep learning renewed interest in the problem. Networks could achieve strong results when trained on large, repeatedly shuffled datasets, but that training arrangement concealed difficulties arising in sequential environments. During the 2010s, continual-learning research developed more explicit benchmarks, evaluation protocols, and method families. Catastrophic forgetting consequently became a practical concern in the adaptation of deployed models as well as a longstanding theoretical issue about distributed memory."
    },
    {
      "heading": "Mechanisms of interference in shared parameters",
      "text": "A neural network generally learns by minimizing a loss function through parameter updates computed using backpropagation. Suppose that \\(\\theta\\) denotes its parameters and that \\(L_A(\\theta)\\) and \\(L_B(\\theta)\\) measure error on an earlier task \\(A\\) and a later task \\(B\\). After learning \\(A\\), training exclusively on \\(B\\) optimizes \\(L_B\\), not the combined objective \\(L_A+L_B\\). Unless the two objectives have compatible solutions along the optimization trajectory, improvement on \\(B\\) can increase error on \\(A\\).\n\nThe local geometry makes this interference explicit. Let \\(g_A=\\nabla_\\theta L_A\\) and \\(g_B=\\nabla_\\theta L_B\\). A gradient-descent step on the new task is \\(\\Delta\\theta=-\\eta g_B\\), where \\(\\eta\\) is the learning rate. For a sufficiently small step, the change in the old loss is approximately\n\n\\[L_A(\\theta-\\eta g_B)-L_A(\\theta) \\approx -\\eta\\,g_A^\\top g_B.\\]\n\nIf the gradient inner product is negative, the update that improves the new task locally worsens the old one. If it is positive, the same update may help both. This relationship explains why task similarity can produce positive transfer in some circumstances and interference in others.\n\nThe first-order picture is incomplete near a minimum of the old task, where \\(g_A\\) may be small. Curvature then matters. A second-order expansion gives\n\n\\[\\Delta L_A \\approx g_A^\\top\\Delta\\theta+ \\frac{1}{2}\\Delta\\theta^\\top H_A\\Delta\\theta,\\]\n\nwhere \\(H_A\\) is the Hessian of the old loss. Moving in a direction with high positive curvature can damage old performance even when the first-order term is negligible. Parameter-protection methods attempt, with different approximations, to identify directions or coordinates in which change is especially costly.\n\nInterference also occurs at several representational levels. Earlier layers may change the features supplied to later classifiers; later layers may alter decision boundaries among classes; normalization statistics may become inappropriate for older inputs. In class-incremental recognition, training predominantly on newly introduced classes can bias output scores toward those classes. A model may retain useful old features yet still misclassify old examples because its scores are no longer calibrated across the full class set. Conversely, retaining an old output layer is insufficient if its input representations have changed substantially.\n\nCatastrophic forgetting is therefore not a single microscopic event equivalent to erasing a stored record. It is an observable behavioral loss that can arise from interacting changes in features, decision boundaries, optimization state, and output competition. It is also distinct from overfitting: a model can generalize well within the new task while forgetting earlier tasks. Likewise, poor performance under distribution shift does not establish forgetting unless there is evidence that subsequent learning degraded a previously acquired capability."
    },
    {
      "heading": "Continual-learning settings and task assumptions",
      "text": "The difficulty of retaining knowledge depends on what a learner must predict and what information it receives. Three widely used settings are task-incremental learning, domain-incremental learning, and class-incremental learning. These labels describe different experimental assumptions rather than interchangeable names for sequential training. A method that performs well in one setting may solve only part of the problem posed by another.\n\nIn task-incremental learning, the system is told which task an input belongs to at prediction time. For example, separate tasks may distinguish different pairs of handwritten digits, and a task identifier selects the appropriate output head. The learner must preserve the ability to solve each task, but it need not determine which task is relevant. This external information can substantially simplify inference and enables methods that maintain task-specific components.\n\nIn domain-incremental learning, the meaning of the outputs remains broadly stable while the input distribution changes. A digit recognizer might encounter different writing styles, scanners, or image transformations while continuing to predict the same digit labels. The system generally cannot rely on an explicit task identifier during prediction. Retaining performance across earlier domains requires preserving useful invariances without preventing adaptation to the latest domain. This setting connects catastrophic forgetting with domain adaptation, but the questions differ: adaptation emphasizes performance in the new domain, whereas continual learning also measures what happens to old ones.\n\nIn class-incremental learning, new categories arrive over time and the model must choose among all categories learned so far. A classifier might first learn cats and dogs, then cars and bicycles, while ultimately predicting among all four without a task label. This requires both retaining earlier distinctions and establishing comparisons between old and new categories. Separate output heads do not by themselves solve the problem unless the model can also select or reconcile them. The absence of old-class examples during later training makes class imbalance and output bias particularly important.\n\nOther distinctions concern the timing and accessibility of experience. Offline continual learning may permit multiple passes through each incoming dataset; online continual learning commonly limits exposure to a stream, sometimes to a single pass. Some experiments provide clear task boundaries, whereas task-free settings require learning under gradual or unannounced change. Recurring tasks, delayed labels, and overlapping distributions further complicate the picture.\n\nThese assumptions determine what counts as a valid solution. Keeping one complete model for every known task may be effective when task identities are available and storage can grow, but it does not meet a fixed-memory, task-free requirement. Similarly, unrestricted retraining on all historical examples can reduce forgetting, yet removes the data-access constraint that motivates much continual-learning research. Meaningful comparisons therefore specify task identity, boundary information, replay access, training passes, and resource limits."
    },
    {
      "heading": "Measuring forgetting and interpreting benchmarks",
      "text": "Evaluation normally tracks performance on earlier tasks throughout a learning sequence. Let \\(R_{i,j}\\) denote the accuracy on task \\(j\\) after training through task \\(i\\), for \\(j\\leq i\\). The resulting performance matrix distinguishes a model that never learned an early task from one that learned it well and later lost it. A single final accuracy score cannot make that distinction. Evaluation should use held-out test sets rather than the examples used to update the model.\n\nA common measure is final average accuracy after \\(T\\) tasks:\n\n\\[A_T=\\frac{1}{T}\\sum_{j=1}^{T}R_{T,j}.\\]\n\nThis summarizes final competence across the observed sequence but does not isolate forgetting. One frequently used forgetting measure compares each earlier task’s best previous performance with its final performance:\n\n\\[F_T=\\frac{1}{T-1} \\sum_{j=1}^{T-1} \\left( \\max_{i\\in\\{j,\\ldots,T-1\\}}R_{i,j}-R_{T,j} \\right).\\]\n\nDefinitions vary across studies, including whether negative values are retained when later learning improves old-task performance. Such conventions must be stated before numerical results are compared.\n\nBackward transfer measures how later learning changes performance on earlier tasks. One common definition is\n\n\\[\\operatorname{BWT} = \\frac{1}{T-1} \\sum_{j=1}^{T-1} \\left(R_{T,j}-R_{j,j}\\right).\\]\n\nNegative backward transfer indicates deterioration relative to performance immediately after learning each task; positive backward transfer indicates improvement. Forward transfer concerns whether earlier learning helps a later task, often relative to an untrained or separately trained baseline. A method can suppress forgetting by barely learning anything new, so retention must be evaluated together with new-task acquisition.\n\nCommon benchmarks include Split MNIST, Permuted MNIST, and class-incremental partitions of CIFAR-100. Split MNIST divides digit categories into sequential groups. Permuted MNIST uses different fixed pixel permutations to construct tasks. These benchmarks are inexpensive and useful for controlled experiments, but their artificial structure may favor methods that exploit known boundaries or task-specific organization. ImageNet-derived sequences, video streams, language tasks, and reinforcement-learning environments introduce different forms of complexity and should not be treated as equivalent tests.\n\nResource accounting is equally important. A replay buffer measured only in numbers of examples may conceal large differences in bytes, particularly for images, audio, or multimodal data. An expanding network should report added parameters, and a generative replay method should include the cost of its generator. Training time, inference cost, number of passes, and hyperparameter-selection procedures can alter apparent rankings. Task order and random initialization also affect outcomes, making repeated runs important.\n\nFinally, a joint-training baseline helps show the cost of sequential data access, but it is not a mathematical upper bound under every optimization procedure. A frozen model provides a complementary reference for stability but says little about adaptation. Reliable interpretation requires both kinds of comparison rather than a single headline forgetting score."
    },
    {
      "heading": "Rehearsal, memory selection, and generative replay",
      "text": "Replay methods reduce forgetting by reintroducing information about earlier experience during later learning. In experience replay, the learner stores selected previous examples and mixes them with current data. Instead of optimizing only the latest distribution, it approximately optimizes a mixture of past and present distributions. This directly addresses the missing-data aspect of sequential training: old capabilities remain represented in the learning signal even after their original training phase ends.\n\nA typical objective combines current and remembered examples:\n\n\\[L(\\theta) = \\mathbb{E}_{(x,y)\\sim D_{\\mathrm{new}}} [\\ell(f_\\theta(x),y)] + \\beta\\, \\mathbb{E}_{(x,y)\\sim M} [\\ell(f_\\theta(x),y)],\\]\n\nwhere \\(M\\) is a memory buffer and \\(\\beta\\) controls its contribution. The balance matters. Too little replay allows interference; too much can slow acquisition of the new distribution or cause repeated fitting to a small collection of stored examples. Replay frequency, sampling weights, and the composition of training batches are therefore part of the method rather than incidental implementation choices.\n\nMemory selection becomes central under a fixed storage budget. Reservoir sampling can maintain a uniform sample of all examples seen in a stream without storing the complete history. Other strategies seek class balance, feature-space coverage, examples near decision boundaries, or samples expected to reveal interference. Each choice preserves different information. A uniformly sampled buffer represents historical frequency, whereas a class-balanced buffer intentionally changes that frequency to protect less common categories.\n\nThe 2017 iCaRL method combined a bounded exemplar memory, representation learning with distillation, and classification using class representatives in feature space. Its importance was not simply that it stored old images, but that it integrated memory management with a class-incremental prediction procedure. Other replay approaches store previous model outputs along with inputs, allowing training to preserve aspects of the earlier function as well as the original labels. Buffer-management strategies can themselves become costly when they require repeated feature extraction or evaluation of many candidate examples.\n\nGenerative replay replaces, or supplements, stored examples with samples produced by a learned generator. “Deep Generative Replay,” introduced in 2017 by Hanul Shin and colleagues, demonstrated a framework in which a generator supplies synthetic earlier experience. Potential generators include a variational autoencoder or a generative adversarial network. The resulting samples can be labeled by a previous solver or generated conditionally on a class.\n\nGenerative replay transfers rather than eliminates the retention problem: the generator must also preserve older distributions. Repeated approximation can omit rare modes or compound errors, and generating sufficiently realistic high-dimensional samples may be expensive. Neither synthetic replay nor an exemplar buffer automatically provides privacy protection. Both may retain information about training examples, so retention performance and privacy guarantees require separate analysis."
    },
    {
      "heading": "Parameter protection, distillation, and constrained updates",
      "text": "A second family of methods attempts to preserve earlier knowledge without necessarily storing raw past examples. Parameter-based regularization adds a penalty for moving weights that were important to previous tasks. The intended distinction is between parameters that can change relatively freely and parameters whose movement would strongly affect old predictions. This approach often has modest storage requirements relative to retaining large datasets, although importance estimates and parameter references still consume memory.\n\nElastic weight consolidation, introduced by James Kirkpatrick and colleagues in 2017, uses a quadratic penalty around previously learned parameters:\n\n\\[L_{\\mathrm{EWC}}(\\theta) = L_{\\mathrm{new}}(\\theta) + \\frac{\\lambda}{2} \\sum_k F_k(\\theta_k-\\theta_k^\\star)^2.\\]\n\nHere, \\(\\theta^\\star\\) is an earlier parameter vector, \\(F_k\\) estimates parameter importance, and \\(\\lambda\\) controls protection strength. The method uses a diagonal approximation associated with Fisher information to estimate how sensitive the earlier model’s predictive distribution is to parameter changes. It is commonly interpreted through an approximate Bayesian account of sequential updating.\n\nThe approximation imposes limitations. Diagonal importance estimates ignore dependencies between parameters, even though many neural-network functions can be preserved by coordinated changes across several weights. A weight that appears important in isolation may be replaceable through such a coordinated adjustment. Quadratic penalties are also local approximations around earlier solutions. If protection becomes too strong across many tasks, the system can retain old behavior at the expense of learning new tasks. Online formulations combine accumulated information to reduce the need for a separate penalty term for every previous task.\n\nRelated methods estimate importance differently. Synaptic Intelligence, introduced in 2017, accumulates information about parameter contributions along the training trajectory. Memory Aware Synapses, introduced in 2018, estimates importance from sensitivity of model outputs. Their differences matter because “importance” is not an intrinsic permanent property of a weight: it depends on the data, output function, and approximation used.\n\nFunctional regularization preserves outputs rather than directly protecting individual coordinates. Learning without Forgetting, introduced by Zhizhong Li and Derek Hoiem in 2016, uses knowledge distillation to encourage a new model to reproduce a previous model’s responses while learning a new task. Distillation can preserve relative scores that hard labels do not contain. However, if preservation is enforced only on new-task inputs, those inputs may inadequately represent the regions where old-task behavior matters.\n\nGradient constraints provide another route. Gradient Episodic Memory, introduced in 2017, uses stored examples to constrain updates so that they do not increase selected remembered-task losses under a local approximation. Its 2019 variant, Averaged GEM, reduces computational overhead through an averaged reference constraint. These approaches connect the geometry of interference to an explicit update rule, but their protection remains dependent on the representativeness of memory and the accuracy of local approximations."
    },
    {
      "heading": "Architectural separation and capacity allocation",
      "text": "Architectural methods reduce interference by controlling which parameters different tasks can modify. At one extreme, a separate network can be trained for each task. This largely prevents destructive updates between models, but storage and inference organization become more demanding as tasks accumulate. It also leaves the problem of selecting the appropriate model when task identity is unavailable. Complete separation protects knowledge but forgoes some opportunities for shared representation and positive transfer.\n\nProgressive Neural Networks, introduced by Andrei Rusu and colleagues in 2016, provide a structured form of expansion. A new network column is added for a new task while earlier columns are frozen. Lateral connections allow the new column to use features from previous columns without modifying them. Earlier parameters are therefore protected from direct overwriting. The trade-off is increasing model size and connectivity as more tasks arrive, which limits applicability when the number of tasks is large or unknown.\n\nParameter isolation can also operate within a fixed or partly fixed network. Masks specify which weights participate in each task, and already allocated weights may be frozen while unused weights remain trainable. PackNet, introduced in 2018 by Arun Mallya and Svetlana Lazebnik, uses iterative pruning and retraining to make room for successive tasks within a network. Its mechanism relies on the observation that a network trained for one task can often be compressed, leaving parameters available for additional tasks.\n\nFixed-capacity allocation eventually faces saturation. Once most useful parameters have been committed, a new task must share existing resources, displace previous allocations, or accept lower performance. Moreover, task-specific masks generally require a routing mechanism at inference time. If the correct mask is selected using an externally supplied task label, the system addresses task-incremental learning rather than unrestricted class-incremental recognition. Mask storage and routing costs are also part of the resource budget.\n\nModular systems attempt a middle course by learning reusable components and selecting combinations appropriate to an input. A mixture-of-experts architecture can support such specialization, but sparsity alone does not guarantee retention. The routing mechanism can change, an expert can still be overwritten, and newly trained experts can alter competition among outputs. Protection must therefore encompass both component behavior and the mechanism that decides which components to use.\n\nFreezing a feature extractor and learning new output layers is a simpler form of separation. It can work well when the existing representation already captures the distinctions required by later tasks, particularly with pretrained visual or language models. Its limitation is reduced representational adaptation: a new domain may require features absent from the frozen backbone. Architectural approaches consequently distribute the stability–plasticity trade-off across components rather than removing it."
    },
    {
      "heading": "Foundation models and sequential fine-tuning",
      "text": "Catastrophic forgetting is especially consequential when adapting a foundation model that already supports many tasks. A pretrained large language model may contain linguistic patterns, factual associations, programming abilities, and instruction-following behavior acquired from varied data. Subsequent fine-tuning on a narrow dataset changes parameters shared by these capabilities. Better performance on the new task can therefore coexist with deterioration elsewhere, even when the fine-tuning dataset is much smaller than the pretraining corpus.\n\nContinued pretraining, supervised instruction tuning, and preference optimization present different retention problems. Continued pretraining changes the distribution of text prediction examples, perhaps toward a specialized domain or language. Instruction tuning emphasizes particular input–response formats. Preference-based training changes the relative desirability of outputs under its training objective. These stages may produce useful specialization, but none guarantees preservation of every previously measured ability. Attributing a regression specifically to forgetting requires comparing checkpoints under consistent evaluation conditions.\n\nParameter-efficient methods change only a subset of the effective model parameters. Low-rank adaptation, for example, represents an update to a weight matrix through a low-rank product while keeping the original matrix frozen during training. This preserves access to the original weights and enables separate adaptations to be stored compactly. Nevertheless, the active adapted model can still perform worse on older tasks: freezing the base matrix does not imply that adding an update preserves its original input–output function.\n\nSeparate adapters or prompts can reduce direct interference when each is trained and selected independently. The remaining questions concern routing, composition, and shared components. If an application always knows which adapter is appropriate, retention is easier to manage than when a single deployed model must handle all requests without task labels. Combining adaptations can also introduce interactions absent when they are evaluated separately. Reports of forgetting therefore need to distinguish changes to the base model from changes in the behavior of a particular adapted configuration.\n\nReplay and distillation extend naturally to foundation models, but their scale changes the economics. Broad retention may require varied text, code, multilingual examples, and other modalities. Access to the original pretraining corpus may be restricted, and distilling a previous checkpoint incurs substantial computation. Evaluation must also avoid benchmark data contamination, since apparent retention can be inflated if evaluation material or close variants enter later training.\n\nExternal information access offers a complementary approach. Retrieval-augmented generation can introduce documents without updating model weights, reducing the need to encode every factual change parametrically. It does not, however, restore a lost reasoning procedure or guarantee that the model will correctly use retrieved evidence. Similarly, in-context learning changes behavior within a supplied context rather than necessarily producing durable parameter-level learning. These mechanisms address different parts of the adaptation problem and should not be conflated with permanent retention."
    },
    {
      "heading": "Deployment constraints and unresolved questions",
      "text": "Real deployments rarely reproduce a clean sequence of equally sized, clearly labeled tasks. A robot encounters recurring environments, a speech recognizer encounters changing accents and microphones, and a recommender system receives evolving user behavior. In reinforcement learning, the learner’s own policy changes which experiences it observes. New training can therefore alter both competence and the future data distribution, creating feedback loops that static supervised benchmarks do not capture.\n\nA central practical issue is deciding what should be retained. Some old patterns remain useful, while others become obsolete. A recommendation model should not indefinitely preserve outdated preferences at the expense of current ones. A system incorporating corrected information should not treat the earlier error as a capability to protect. Catastrophic forgetting research usually assumes that earlier tasks remain valid evaluation targets; deployment may require an explicit distinction between persistent knowledge, transient patterns, and information that should be removed.\n\nThis distinction also connects retention with data privacy. Replay buffers can conflict with data-minimization or deletion requirements, while a model may retain information even after its source examples are removed. Machine unlearning concerns intentionally reducing the influence of specified training data; catastrophic-forgetting mitigation ordinarily seeks to preserve influence that remains useful. These objectives are not simple opposites, because removing one source’s contribution while maintaining unrelated capabilities requires selective change rather than indiscriminate degradation.\n\nAnother unresolved question is how much apparent forgetting reflects inaccessible knowledge rather than complete loss. A model may perform poorly under one prompt, output head, or decision rule yet retain representations from which an earlier task can be relearned quickly. Behavioral tests, feature analyses, and relearning experiments measure different aspects of this distinction. None independently provides a complete account of what a high-dimensional network has retained, and claims about internal knowledge require care beyond reporting a drop in task accuracy.\n\nLong-horizon scalability remains difficult. Fixed memory must represent an expanding history; regularization can progressively restrict useful movement; expanding architectures consume growing resources; generators can accumulate approximation errors. Task relatedness can relieve these pressures through shared features and positive transfer, but unrelated or conflicting objectives expose finite-capacity limits. Computational budgets further determine whether repeated replay or full historical retraining is feasible.\n\nResearch therefore examines retention under explicit memory, computation, latency, and data-access constraints, rather than treating zero forgetting as a sufficient objective. Particularly demanding cases include unlabeled streams, gradual changes without task boundaries, rare-event retention, and combined visual–language experience in multimodal learning. In these settings, the learner must determine which changes represent useful novelty, which reflect noise, and which earlier distinctions remain relevant, while evaluation must separate those judgments from the consequences of limited storage and optimization."
    }
  ],
  "links": [
    "artificial-neural-network",
    "database",
    "deep-learning",
    "loss-function",
    "backpropagation",
    "learning-rate",
    "overfitting",
    "distribution-shift",
    "test-set",
    "variational-autoencoder",
    "generative-adversarial-network",
    "regularization",
    "fisher-information",
    "knowledge-distillation",
    "foundation-model",
    "large-language-model",
    "fine-tuning",
    "low-rank-adaptation",
    "benchmark-data-contamination",
    "retrieval-augmented-generation",
    "in-context-learning",
    "recommender-system",
    "reinforcement-learning",
    "data-privacy",
    "multimodal-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}