Continual Learning for LLMs

LLM: Large Language Model

Follow this topic Ask about this topic

The field investigates how large language models can keep learning through continual pretraining, instruction tuning, alignment, and personalization without losing earlier knowledge or skills. Common approaches include replaying old or generated examples, distillation and regularization, selective or orthogonal parameter updates, modular adapters, and external memories that leave the model weights unchanged. Evaluations use sequential task suites such as TRACE and SuperNI, domain streams such as biomedical NLP, and evolving knowledge or user-interaction settings, measuring both new-task performance and retention over time. No single approach works reliably across settings: results depend on task order, model capacity, and update type, while long-term evaluation must balance adaptation, forgetting, transfer, safety, and compute cost.

Latest papers

236 papers44 papers in the last four weeks, up 110% on the four weeks before. 0.3% of all new papers.

Mar 10, 2026

On Catastrophic Forgetting in Low-Rank Decomposition-Based Parameter-Efficient Fine-Tuning

Parameter-efficient fine-tuning (PEFT) based on low-rank decomposition, such as LoRA, has become a standard for adapting large pretrained models. However, its behavior in sequential learning -- specifically regarding catastrophic forgetting -- remains insufficiently understood. In this work, we present an empirical study showing that forgetting is strongly influenced by the geometry and parameterization of the update subspace. While methods that restrict updates to small, shared matrix subspaces often suffer from task interference, tensor-based decompositions (e.g., LoRETTA) mitigate forgetting by capturing richer structural information within ultra-compact budgets, and structurally aligned parameterizations (e.g., WeGeFT) preserve pretrained representations. Our findings highlight update subspace design as a key factor in continual learning and offer practical guidance for selecting efficient adaptation strategies in sequential settings.

Muhammad Ahmad, Jingjing Zheng, Yankai Cao

Department of Electrical and Computer Engineering, The University of British Columbia · Department of Mathematics, The University of British Columbia · Department of Chemical and Biological Engineering, The University of British Columbia

Parameter-Efficient Fine-Tuning Catastrophic Forgetting Low-Rank Adaptation Continual Learning for LLMs

Mar 8, 2026

Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams

LLMs operating in dynamic real-world contexts often encounter knowledge that evolves continuously or emerges incrementally. To remain accurate and effective, models must adapt to newly arriving information on the fly. We introduce Online Adaptation to Continual Knowledge Streams(OAKS) to evaluate this capability, establishing a benchmark for online adaptation over streaming, continually updating knowledge. Specifically, the benchmark is structured as a sequence of fine-grained context chunks where facts change dynamically across time intervals. OAKS comprises two datasets: OAKS-BABI and OAKS-Novel, where individual facts evolve multiple times across context chunks. These datasets include dense annotations to measure whether models track changes accurately. Evaluating 14 models with varied inference approaches, we observe significant limitations in current methodologies. Both state-of-the-art models and agentic memory systems fail to adapt robustly on OAKS, demonstrating delays in state-tracking and susceptibility to distraction within streaming environments.

Jiyeon Kim, Hyunji Lee, Dylan Zhou +6

KAIST AI · UNC Chapel Hill · Google +3

LLM Evaluation Continual Learning for LLMs Online Continual Learning Memory-Augmented Language Models

Mar 6, 2026

A Dynamic Self-Evolving Extraction System

The extraction of structured information from raw text is a fundamental component of many NLP applications, including document retrieval, ranking, and relevance estimation. High-quality extractions often require domain-specific accuracy, up-to-date understanding of specialized taxonomies, and the ability to incorporate emerging jargon and rare outliers. In many domains--such as medical, legal, and HR--the extraction model must also adapt to shifting terminology and benefit from explicit reasoning over structured knowledge. We propose DySECT, a Dynamic Self-Evolving Extraction and Curation Toolkit, which continually improves as it is used. The system incrementally populates a versatile, self-expanding knowledge base (KB) with triples extracted by the LLM. The KB further enriches itself through the integration of probabilistic knowledge and graph-based reasoning, gradually accumulating domain concepts and relationships. The enriched KB then feeds back into the LLM extractor via prompt tuning, sampling of relevant few-shot examples, or fine-tuning using KB-derived synthetic data. As a result, the system forms a symbiotic closed-loop cycle in which extraction continuously improves knowledge, and knowledge continuously improves extraction.

Moin Amin-Naseri, Hannah Kim, Estevam Hruschka

Megagon Labs

Document Information Extraction LLM Information Extraction Knowledge Augmentation for Language Models Language Model Self-Improvement Continual Learning for LLMs

Mar 2, 2026

Surgical Post-Training: Proximal On-Policy Distillation for Reasoning with Knowledge Retention

Injecting new reasoning knowledge into Large Language Models (LLMs) via post-training often induces catastrophic forgetting. Recent studies emphasize the importance of on-policy data but suggest that KL-divergence fails to mitigate forgetting. In contrast, we show, both analytically and empirically, that the KL-constrained reward formulation actually plays a critical role in retaining knowledge during post-training. This motivates our Surgical Post-Training (SPOT), a proximal on-policy distillation framework designed to optimize reasoning efficiently while preserving prior knowledge. SPOT consists of (1) a data rectification pipeline employing an Oracle to surgically correct erroneous steps via minimal edits, generating proximal on-policy data; and (2) a reward-based binary cross-entropy objective essential for enhancing reasoning and mitigating forgetting. Empirically, with only 4k rectified math pairs, SPOT improves Qwen3-8B's accuracy by 6.2% on average across in-domain and out-of-domain tasks, requiring merely 16-minute model training on 8x H800 GPUs. Moreover, SPOT provides a superior initialization for subsequent reinforcement learning, significantly elevating the performance ceiling. Code: https://github.com/Visual-AI/SPoT

Wenye Lin, Kai Han

The University of Hong Kong

LLM Post-Training LLM Reasoning On-Policy Distillation Knowledge Distillation Continual Learning for LLMs

Mar 1, 2026

Understanding LoRA as Knowledge Memory: An Empirical Analysis

Continuous knowledge updating for pre-trained large language models (LLMs) is increasingly necessary yet remains challenging. Although inference-time methods like In-Context Learning (ICL) and Retrieval-Augmented Generation (RAG) are popular, they face constraints in context budgets, costs, and retrieval fragmentation. Departing from these context-dependent paradigms, this work investigates a parametric approach using Low-Rank Adaptation (LoRA) as a modular knowledge memory. Although few recent works examine this concept, the fundamental mechanics governing its capacity and composability remain largely unexplored. We bridge this gap through the first systematic empirical study mapping the design space of LoRA-based memory, ranging from characterizing storage capacity and optimizing internalization to scaling multi-module systems and evaluating long-context reasoning. Rather than proposing a single architecture, we provide practical guidance on the operational boundaries of LoRA memory. Overall, our findings position LoRA as the complementary axis of memory alongside RAG and ICL, offering distinct advantages.

Seungju Back, Dongwoo Lee, Naun Kang +4

KAIST · Samsung SDS · New York University

Continual Learning for LLMs Memory-Augmented Language Models Knowledge Augmentation for Language Models Persistent Memory for Language Models Low-Rank Adaptation

Feb 27, 2026

Enhancing Continual Learning for Software Vulnerability Prediction: Addressing Catastrophic Forgetting via Hybrid-Confidence-Aware Selective Replay for Temporal LLM Fine-Tuning

Recent work applies Large Language Models (LLMs) to source-code vulnerability detection, but most evaluations still rely on random train-test splits that ignore time and overestimate real-world performance. In practice, detectors are deployed on evolving code bases and must recognise future vulnerabilities under temporal distribution shift. This paper investigates continual fine-tuning of a decoder-style language model (microsoft/phi-2 with LoRA) on a CVE-linked dataset spanning 2018-2024, organised into bi-monthly windows. We evaluate eight continual learning strategies, including window-only and cumulative training, replay-based baselines and regularisation-based variants. We propose Hybrid Class-Aware Selective Replay (Hybrid-CASR), a confidence-aware replay method for binary vulnerability classification that prioritises uncertain samples while maintaining a balanced ratio of VULNERABLE and FIXED functions in the replay buffer. On bi-monthly forward evaluation Hybrid-CASR achieves a Macro-F1 of 0.667, improving on the window-only baseline (0.651) by 0.016 with statistically significant gains (

p = 0.026

) and stronger backward retention (IBR@1 of 0.741). Hybrid-CASR also reduces training time per window by about 17 percent compared to the baseline, whereas cumulative training delivers only a minor F1 increase (0.661) at a 15.9-fold computational cost. Overall, the results show that selective replay with class balancing offers a practical accuracy-efficiency trade-off for LLM-based temporal vulnerability detection under continuous temporal drift.

Xuhui Dou, Hayretdin Bahsi, Alejandro Guerra-Manzanares

School of Computer Science, University of Nottingham, Nottingham, United Kingdom · School of Informatics and Computing, Northern Arizona University, Flagstaff, United States of America · School of Computer Science, University of Nottingham, Ningbo, China

Software Vulnerability Detection Continual Learning for LLMs Replay-Based Continual Learning LLMs for Cybersecurity

Feb 25, 2026

Efficient Continual Learning in Language Models via Thalamically Routed Cortical Columns

Large language models deployed in the wild must adapt to evolving data, user behavior, and task mixtures without erasing previously acquired capabilities. In practice, this remains difficult: sequential updates induce catastrophic forgetting, while many stabilization methods rely on external procedures that are costly, brittle, or difficult to scale. We present TRC

^{2}

(Thalamically Routed Cortical Columns), a decoder-only architecture that makes continual learning a property of the backbone itself. TRC

^{2}

combines stacked cortical columns with a thalamic modulatory pathway for selective inter-column communication and a hippocampal pathway for event selective retrieval, delayed surprise-based writing, and replay-driven consolidation. This design localizes fast plasticity while preserving a slower stable computation pathway. We further introduce a causal memory-update scheme and an online replay controller that adjusts consolidation strength from measured forgetting. Across a task-sequential language-modeling stream over C4, WikiText-103, and GSM8K, TRC

^{2}

consistently improves task-boundary modeling quality and substantially reduces cumulative forgetting relative to Transformer, Mamba, MoE, DeepSeek and continual learning baselines trained under the same pipeline. Ablations show that the thalamic and hippocampal components are central to the retention gains, while the full model remains competitive in throughput and training cost.

Afshin Khadangi

SnT, University of Luxembourg

Continual Learning for LLMs Decoder-Only Language Models Memory-Augmented Language Models Replay-Based Continual Learning Modular Neural Networks

Feb 20, 2026

Non-Interfering Weight Fields: Treating Model Parameters as a Continuously Extensible Function

Large language models store all learned knowledge in a single, fixed weight vector. Teaching a model new capabilities requires modifying those same weights, inevitably degrading previously acquired knowledge. This fundamental limitation, known as catastrophic forgetting, has resisted principled solutions for decades. Existing approaches treat weights as immutable artifacts that must be protected through techniques like regularization heuristics, replay buffers, or isolated adapter modules. The problem is none of these provide a structural guarantee against forgetting. In this work, we propose Non-Interfering Weight Fields (NIWF), a framework that replaces the fixed weight paradigm with a learned function that generates weight configurations on demand from a continuous capability coordinate space. After training on a task, we commit the occupied coordinate region by snapshotting the fields outputs on anchor points to enforce a functional lock during all future training. We validate NIWF on sequential instructionfollowing and code generation tasks using Mistral-7B, demonstrating zero forgetting on committed tasks with competitive perplexity on new tasks. The framework introduces the notion of software-like versioning for neural network intelligence, where capabilities can be committed, extended, composed, and rolled back without retraining.

Sarim Chaudhry

Purdue University

Continual Learning for LLMs Catastrophic Forgetting Hypernetworks

Feb 18, 2026

Aligning Language Models from User Interactions

Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.

Thomas Kleine Buening, Jonas Hübotter, Barna Pásztor +3

ETH Zurich · MIT · University of Zurich

LLM Alignment Personalized Language Model Alignment Self-Training for Language Models Continual Learning for LLMs Language Model Distillation

Feb 17, 2026

Updating Parametric Knowledge with Context Distillation Retains Post-Training Capabilities

Post-training endows pretrained LLMs with a variety of desirable skills, including instruction-following, reasoning, and others. However, these post-trained LLMs only encode knowledge up to a cut-off date, necessitating continual adaptation. Unfortunately, existing solutions cannot simultaneously learn new knowledge from an adaptation document corpora and mitigate the forgetting of earlier learned capabilities. To address this, we introduce Distillation via Split Contexts (DiSC), a simple context-distillation based approach for continual knowledge adaptation. \methodname~derives student and teacher distributions by conditioning on distinct segments of the training example and minimizes the KL divergence between the shared tokens. This allows us to efficiently apply context-distillation without requiring explicit generation steps during training. We run experiments on four post-trained models and two adaptation domains. Compared to prior finetuning and distillation methods for continual adaptation, DiSC consistently reports the best trade-off between learning new knowledge and mitigating forgetting of previously learned skills like instruction-following, reasoning, and factual knowledge.

Shankar Padmanabhan, Mustafa Omer Gul, Tanya Goyal

Department of Computer Science, Cornell University.

Continual Learning for LLMs Context Distillation Catastrophic Forgetting

Feb 16, 2026

Panini: Continual Learning in Token Space via Structured Memory

Language models are increasingly used to reason over content they were not trained on, such as new documents, evolving knowledge, and user-specific data. A common approach is retrieval-augmented generation (RAG), which stores verbatim documents externally (as chunks) and retrieves only a relevant subset at inference time for an LLM to reason over. However, this results in inefficient usage of test-time compute (LLM repeatedly reasons over the same documents); moreover, chunk retrieval can inject irrelevant context that increases unsupported generation. We propose a human-like non-parametric continual learning framework, where the base model remains fixed, and learning occurs by integrating each new experience into an external semantic memory state that accumulates and consolidates itself continually. We present Panini, which realizes this by representing documents as Generative Semantic Workspaces (GSW) -- an entity- and event-aware network of question-answer (QA) pairs, sufficient for an LLM to reconstruct the experienced situations and mine latent knowledge via reasoning-grounded inference chains on the network. Given a query, Panini only traverses the continually-updated GSW (not the verbatim documents or chunks), and retrieves the most likely inference chains. Across six QA benchmarks, Panini achieves the highest average performance, 5%-7% higher than other competitive baselines, while using 2-30x fewer answer-context tokens, supports fully open-source pipelines, and reduces unsupported answers on curated unanswerable queries. The results show that efficient and accurate structuring of experiences at write time -- as achieved by the GSW framework -- yields both efficiency and reliability gains at read time. Code is available at https://github.com/roychowdhuryresearch/gsw-memory.

Shreyas Rajesh, Pavan Holur, Mehmet Yigit Turali +2

Department of Electrical and Computer Engineering, University of California, Los Angeles, USA.

Continual Learning for LLMs Memory-Augmented Language Models Persistent Memory for Language Models Knowledge Augmentation for Language Models RAG Robustness LLM Reasoning with Graphs

Jan 24, 2026

Sparse Subspace-to-Expert Sharing for Task-Agnostic Continual Learning

Continual learning in Large Language Models (LLMs) is hindered by the plasticity-stability dilemma, where acquiring new capabilities often leads to catastrophic forgetting of previous knowledge. Existing methods typically treat parameters uniformly, failing to distinguish between specific task knowledge and shared capabilities. We introduce Mixture of Sparse Experts for Task-Agnostic Continual Learning (SETA), a framework that resolves the plasticity-stability conflict through adaptive sparse subspace decomposition into task-specific expert modules. Unlike standard updates, where tasks compete for the same parameters, SETA separates knowledge into unique experts, designed to isolate task-specific patterns, and shared experts, responsible for capturing common features. SETA shares an overlapping parameter block only when the new task's update passes a subspace-agreement test and a sign test against its prior expert; shared knowledge is protected by adaptive elastic anchoring, and independently gated routers retrieve the relevant experts at inference without task identifiers. Extensive experiments across diverse domain-specific benchmarks demonstrate that SETA achieves competitive or superior overall performance relative to state-of-the-art continual learning baselines, improving overall performance over the strongest baseline by 1.1 to 1.7 points on TRACE and by up to 5.5 points on the O-LoRA benchmark, with backward transfer between -2.2 and -1.3 on TRACE, against 4.6 to 6.3 points of forgetting for MoE-CL and 8 to 15 points for the other LoRA baselines.

Fatema Siddika, Md Anwar Hossen, Tanwi Mallick +1

Iowa State University, Ames, USA · Argonne National Laboratory, USA

Continual Learning for LLMs Continual Learning Task-Free Continual Learning Mixture-of-Experts Language Models

Jan 15, 2026

SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation

Personalizing Large Language Models typically relies on static retrieval or one-time adaptation, assuming user preferences remain invariant over time. However, real-world interactions are dynamic, where user interests continuously evolve, posing a challenge for models to adapt to preference drift without catastrophic forgetting. Standard continual learning approaches often struggle in this context, as they indiscriminately update on noisy interaction streams, failing to distinguish genuine preference shifts from transient contexts. To address this, we introduce SPRInG, a novel semi-parametric framework designed for effective continual personalization. During training, SPRInG employs drift-driven selective adaptation, which utilizes a likelihood-based scoring function to identify high-novelty interactions, selectively updating the user-specific adapter on drift signals while preserving hard-to-learn residuals in a replay buffer. During inference, we apply strict relevance gating and fuse parametric knowledge with retrieved history via probability interpolation. Experiments on the long-form personalized generation benchmark demonstrate that SPRInG significantly outperforms existing baselines, validating its robustness for real-world continual personalization.

Seoyeon Kim, Jaehyung Kim

Yonsei University

Continual Learning for LLMs LLM Personalization Memory-Augmented Language Models

Dec 22, 2025

When Less is More: 8-bit Quantization Improves Continual Learning in Large Language Models

Catastrophic forgetting poses a fundamental challenge in continual learning, particularly when models are quantized for deployment efficiency. We systematically investigate the interplay between quantization precision (FP16, INT8, INT4) and replay buffer strategies in large language models, revealing unexpected dynamics. While FP16 achieves superior initial task performance (74.44% on NLU), we observe a striking inversion on subsequent tasks: quantized models outperform FP16 by 8-15% on final task forward accuracy, with INT4 achieving nearly double FP16's performance on Code generation (40% vs 20%). Critically, even minimal replay buffers (0.1%) dramatically improve retention - increasing NLU retention after Math training from 45% to 65% across all precision levels - with INT8 consistently achieving the optimal balance between learning plasticity and knowledge retention. We hypothesize that quantization-induced noise acts as implicit regularization, preventing the overfitting to new task gradients that plagues high-precision models. These findings challenge the conventional wisdom that higher precision is always preferable, suggesting instead that INT8 quantization offers both computational efficiency and superior continual learning dynamics. Our results provide practical guidelines for deploying compressed models in continual learning scenarios: small replay buffers (1-2%) suffice for NLU tasks, while Math and Code benefit from moderate buffers (5-10%), with quantized models requiring less replay than FP16 to achieve comparable retention. Code is available at https://github.com/Festyve/LessIsMore.

Michael S. Zhang, Rishi A. Ruia, Arnav Kewalram +3

Algoverse

LLM Quantization Continual Learning for LLMs Replay-Based Continual Learning Catastrophic Forgetting

Nov 25, 2025

Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs

Large language models (LLMs) often produce incorrect or outdated content after being employed. Efficient and accurate knowledge updates without costly retraining are a major challenge. This problem is particularly challenging in lifelong settings, where complex, unstructured knowledge must coexist without interference. We introduce RILKE (Representation Intervention for Lifelong KnowledgE Control), a robust and scalable method that treats knowledge control as interventions within the model's representation space. Leveraging representation-space expressiveness, we identify two key properties enabling RILKE to achieve fine-grained control over complex, unstructured knowledge while maintaining general utility with frozen base weights. During training, RILKE learns paraphrase-robust and edit-localized modules that limit each update to a low-dimensional subspace to minimize cross-edit interference. At inference, a query-adaptive router selects the appropriate module to guide the model's generation. Across LLaMA and Qwen models, RILKE scales effectively to large-scale benchmarks, demonstrating high edit success and strong paraphrase generalization while preserving general utility with modest memory overhead. These results show RILKE is an effective and scalable solution for lifelong knowledge control in LLMs.

Xuyuan Liu, Shengyu Chen, Xinshuai Dong +6

Dartmouth College · NEC Laboratories America · Carnegie Mellon University

Knowledge Editing Continual Learning for LLMs

Nov 19, 2025

Parameter Importance-Driven Continual Learning for Foundation Models

Domain-specific post-training often causes catastrophic forgetting, making foundation models lose their general reasoning ability and limiting their adaptability to dynamic real-world environments. Preserving general capabilities while acquiring downstream domain knowledge is a central challenge for large language and multimodal models. Traditional continual learning methods, such as regularization, replay and architectural isolation, suffer from poor downstream performance, reliance on inaccessible historical data, or additional parameter overhead. While recent parameter-efficient tuning (PET) methods can alleviate forgetting, their effectiveness strongly depends on the choice of parameters and update strategies. In this paper, we introduce PIECE, a Parameter Importance Estimation-based Continual Enhancement method that preserves general ability while efficiently learning domain knowledge without accessing prior training data or increasing model parameters. PIECE selectively updates only 0.1% of core parameters most relevant to new tasks, guided by two importance estimators: PIECE-F based on Fisher Information, and PIECE-S based on a second-order normalization that combines gradient and curvature information. Experiments across three language models and two multimodal models show that PIECE maintains general capabilities and achieves state-of-the-art continual learning performance across diverse downstream tasks. Our results highlight a practical path to scalable, domain-adaptive foundation models without catastrophic forgetting.

Lingxiang Wang, Hainan Zhang, Zhiming Zheng

Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, Beihang University · School of Artificial Intelligence, Beihang University

Continual Learning for LLMs Parameter-Efficient Fine-Tuning

Nov 9, 2025

Mixtures of SubExperts for Large Language Continual Learning

Enabling lifelong learning in LLMs demands resolving the stability-plasticity dilemma (i.e., models must incorporate new knowledge without overwriting prior representations) while maintaining scalability under bounded parameter growth. Existing PEFT methods fail to satisfy this triad; shared-parameter approaches suffer from catastrophic interference, while task-isolated expansions preclude knowledge transfer and scale linearly. We propose Mixtures of SubExperts (MoSEs), a modular and sparse framework that factorizes model capacity into reusable, compositional primitives. MoSEs augment transformer layers with lightweight SubExperts and a learned sub-routing function that dynamically selects and composes a sparse subset of modules conditioned on task inputs. This induces a structured decomposition of the parameter space where knowledge is localized yet accessible, mitigating interference while preserving reuse. Specifically, MoSEs balance the dilemma via three pillars: (i) stability by isolating knowledge within sparsely activated modules, (ii) plasticity through routing-driven recombination and selective expansion, and (iii) scalability via sublinear growth in effective capacity. Notably, the routing mechanism enables compositional generalization, allowing new tasks to be represented as combinations of previously acquired sub-functions. We empirically validate MoSEs on TRACE and SuperNI, showing reduced forgetting, improved forward transfer, and better parameter efficiency over strong PEFT baselines. MoSEs establish a new Pareto frontier, achieving state-of-the-art performance while maintaining strict parameter budgets. Our results suggest that modular sparsity and compositional routing are key inductive biases for building foundation models that continually learn without saturation.

Haeyong Kang, Hee Suk Yoon, Dahua Feng +1

Korea Advanced Institute of Science and Technology (KAIST)

Continual Learning for LLMs Mixture-of-Experts Language Models Incremental Learning Parameter-Efficient Fine-Tuning Catastrophic Forgetting

Oct 21, 2025

Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting

Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SFT while achieving comparable or higher target task performance. To investigate the cause for this difference, we consider a simplified setting in which the LM is modeled as a mixture of two distributions, one corresponding to prior knowledge and the other to the target task. We identify that the mode-seeking nature of RL, which stems from its use of on-policy data, enables keeping prior knowledge intact when learning the target task. We then verify this insight by demonstrating that the use on-policy data underlies the robustness of RL to forgetting in practical settings, as opposed to other algorithmic choices such as the KL regularization or advantage estimation. Lastly, as a practical implication, our results highlight the potential of mitigating forgetting using approximately on-policy data, which can be substantially more efficient to obtain than fully on-policy data.

Howard Chen, Noam Razin, Karthik Narasimhan +1

Princeton Language and Intelligence, Princeton University.

Continual Learning for LLMs RL Post-Training Catastrophic Forgetting

Oct 15, 2025

K-Merge: Online Continual Merging of Adapters for On-device Large Language Models

On-device deployment of Large Language Models (LLMs) frequently leverages Low-Rank Adapters (LoRAs) to support diverse downstream tasks under tight resource constraints. To address the limited storage capacity of mobile devices, recent works have explored model merging techniques to fuse multiple LoRAs into a single one. In practice, however, LoRAs are often delivered incrementally, as users request support for new tasks (e.g., novel problem types or languages). This scenario introduces a new challenge: on-device online continual merging, where the objective is to incorporate new LoRAs while preserving the performance on previously supported tasks. In this paper, we propose a data-free and computationally efficient strategy for selecting and merging LoRAs when a new one becomes available, assuming the device can store only a limited number of adapters. Extensive experiments across real-world tasks demonstrate the superiority of our approach compared to alternative strategies while adhering to the storage budget and compute limitations of on-device settings. The project page is available at: https://donaldssh.github.io/K-Merge.

Donald Shenaj, Ondrej Bohdal, Taha Ceritli +3

Samsung R&D Institute UK · University of Pisa · University of Padova

Continual Learning for LLMs On-Device Language Model Inference Low-Rank Adaptation Language Model Merging

Oct 15, 2025

LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X

We propose and test the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in large language models (LLMs). To unveil junk effects, we designed a novel controlled experiment on real Twitter/X corpora, by constructing junk and reverse-controlled datasets via two orthogonal operationalizations: M1 (engagement degree) and M2 (semantic quality), with matched token scale and training operations across conditions. Compared to the control group, continual pre-training of 4 LLMs on the junk dataset causes non-trivial declines (Hedges' g>0.3) on reasoning, long-context understanding, safety, and inflating "dark traits" (e.g., psychopathy, narcissism). The gradual mixtures of junk and control datasets also yield dose-response cognition decay: for example, under M1, ARC-Challenge with Chain-of-Thought drops 72.1 -> 57.2 and RULER-CWE 83.7 -> 52.3 as junk ratio rises from 0% to 100%. Error forensics reveal several key insights. First, we identify thought-skipping as the primary lesion in reasoning: models increasingly truncate or skip chains. Second, partial but incomplete healing is observed: scaling instruction tuning and clean continual pre-training improve the declined cognition, yet cannot restore baseline capability, suggesting persistent representational drift rather than format mismatch. Finally, we discover that the popularity, a non-semantic metric, of a tweet is a better indicator of the Brain Rot effect than the length in M1. Together, the results provide significant, multi-perspective evidence that social effects of data could be a causal driver of LLM capability decay in continual pre-training, thereby motivating routine "cognitive health checks" for deployed and evolving LLMs.

Shuo Xing, Junyuan Hong, Yifan Wang +5

Texas A&M University · University of Texas at Austin · Purdue University

LLM Training LLM Reliability Continual Learning for LLMs

Jul 7, 2025

Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training

Continual post-training (CPT) is a popular and effective technique for adapting foundation models like multimodal large language models to ever-evolving downstream tasks. While existing research primarily focuses on methods like data replay, model expansion, or parameter regularization, the fundamental role of the learning paradigm remains largely unexplored. This paper presents a comparative analysis of two core post-training paradigms: supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), investigating their respective impacts on knowledge retention during CPT. Our experiments are conducted across multiple multimodal tasks, utilizing Qwen2.5-VL-7B-Instruct as the base model. The investigation yields two significant findings: (1) When continuously learning on downstream tasks, SFT leads to catastrophic forgetting of previously learned tasks. In contrast, RFT inherently preserves prior knowledge and achieves performance comparable to multi-task training. (2) RFT successfully protects and even enhances the model's general knowledge on standard benchmarks, while SFT degrades general model capabilities severely. Further analysis reveals that this stability is not primarily due to explicit mechanisms like KL penalty or chain-of-thought reasoning. We investigate RFT's learning dynamics and find that its selective update mechanism inherently prevents interference with established knowledge. Based on this insight, we propose a rollout-based instance filtering algorithm (RIF-RFT) that enhances the training efficiency of RFT by focusing on learnable samples. Our comprehensive study demonstrates the superiority of RFT as a robust paradigm for continual post-training.

Song Lai, Haohan Zhao, Rong Feng +9

Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong, China · City University of Hong Kong, Hong Kong, China · School of Artificial Intelligence, Jilin University, Changchun, China +3

Continual Learning for LLMs LLM Post-Training RL Fine-Tuning Catastrophic Forgetting

Jun 21, 2025

Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models

Large language models can answer questions correctly and still hallucinate their citations. Most citation systems address this by retrieving documents at inference time. Retrieval adds latency and depends on external infrastructure. Retrieval noise can also affect the answer. We study whether LLMs can cite documents from their (continual) pretraining without test-time retrieval. We introduce CitePretrainBench to evaluate this setting. It includes Wikipedia, Common Crawl, and arXiv as real-world corpora and also includes novel, unseen documents. The benchmark includes short-form citation tasks for a single fact and long-form tasks for multiple facts. Continual-pretraining first binds factual knowledge to a persistent identifier for each document. Instruction tuning then teaches the model to put these identifiers in its answers as citations. Passive Indexing appends an identifier at the end of each document. Synthetic data used in Active Indexing restate facts in diverse, compositional forms. We train two mappings together. These mappings are source

\to

fact and fact

\to

source. The first direction involves the model generating content given a cited source. In the second, it learns to attribute its own answers when facts are paraphrased or combined. On Qwen-2.5-7B and 3B, Active Indexing improves citation precision over Passive Indexing on every evaluated citation task, with gains of up to 30.2 percentage points. In the scaling ablation, citation performance continues to improve as we add augmented data, including at 16 times the original token count. Combining internal and external citations also improves robustness to retrieval noise.

Yukun Huang, Sanxing Chen, Jian Pei +2

Duke University · Meta

Continual Learning for LLMs

Jun 12, 2025

TreeLoRA: Efficient Continual Learning via Layer-Wise LoRAs Guided by a Hierarchical Gradient-Similarity Tree

Many real-world applications collect data in a streaming environment, where learning tasks are encountered sequentially. This necessitates continual learning (CL) to update models online, enabling adaptation to new tasks while preserving past knowledge to prevent catastrophic forgetting. Nowadays, with the flourish of large pre-trained models (LPMs), efficiency has become increasingly critical for CL, due to their substantial computational demands and growing parameter sizes. In this paper, we introduce TreeLoRA (K-D Tree of Low-Rank Adapters), a novel approach that constructs layer-wise adapters by leveraging hierarchical gradient similarity to enable efficient CL, particularly for LPMs. To reduce the computational burden of task similarity estimation, we employ bandit techniques to develop an algorithm based on lower confidence bounds to efficiently explore the task structure. Furthermore, we use sparse gradient updates to facilitate parameter optimization, making the approach better suited for LPMs. Theoretical analysis is provided to justify the rationale behind our approach, and experiments on both vision transformers (ViTs) and large language models (LLMs) demonstrate the effectiveness and efficiency of our approach across various domains, including vision and natural language processing tasks.

Yu-Yang Qian, Yuan-Ze Xu, Zhen-Yu Zhang +2

National Key Laboratory for Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, China · RIKEN AIP, Tokyo, Japan

Continual Learning for LLMs Low-Rank Adaptation Online Continual Learning Incremental Learning Parameter-Efficient Fine-Tuning

Jan 9, 2025

Forget About It: Mitigating Forgetting Without Replay or Auxiliary Objectives

Have we made mitigating catastrophic forgetting more complicated than it needs to be? Continual adaptation often degrades performance on earlier tasks, motivating rehearsal buffers, auxiliary objectives, and adaptive merging procedures. We investigate how far a simpler alternative-periodic parameter interpolation during fine-tuning-can go on its own. We introduce Sequential Fine-tuning with Averaging (SFA) and evaluate it alongside replay-free Lookahead, originally designed for optimization stability to find compelling results. In Qwen3-14B adapter continual learning, fixed configurations of both methods outperform the tested replay and adaptive-merging configurations in mean final accuracy using up to 36.3% less task training time. Furthermore, across language and vision tasks, our results compete with more complex continual-learning methods. Together, these findings suggest that simple in-training interpolation deserves consideration before introducing more elaborate forgetting-mitigation mechanisms.

Anat Kleiman, Gintare Karolina Dziugaite, Jonathan Frankle +2

Harvard University, Kempner Institute · Google DeepMind · Databricks

Catastrophic Forgetting Continual Learning for LLMs

Nov 11, 2024

Continual Memorization of Factoids in Language Models

As new knowledge rapidly accumulates, language models (LMs) with pretrained knowledge quickly become obsolete. A common approach to updating LMs is fine-tuning them directly on new knowledge. However, recent studies have shown that fine-tuning for memorization may be ineffective in storing knowledge or may exacerbate hallucinations. In this work, we introduce a setting we call continual memorization, where a model must memorize and retain a set of factoids through multiple stages of fine-tuning on subsequent datasets. We characterized the forgetting patterns through extensive experiments and show that LMs widely suffer from forgetting, especially when needing to memorize factoids in the second stage. We posit that forgetting can be alleviated by modifying training dynamics: (1) protecting the memorization process when learning factoids or (2) reducing interference from subsequent training stages. Intriguingly, we find that mixing randomly generated word sequences or generic data sampled from pretraining corpora at different training stages effectively mitigates forgetting REMIX: Random and Generic Data Mixing). REMIX can recover performance from severe forgetting, outperforming replay methods and other continual learning baselines. We analyze how REMIX influences the learning process and find that robust memorization follows a distinct pattern: the model stores factoids in earlier layers than usual and diversifies the layers that retain them, which results in easier recall and manipulate of the learned factoids.

Howard Chen, Jiayi Geng, Adithya Bhaskar +2

Princeton Language and Intelligence (PLI), Princeton University

Continual Learning for LLMs Memorization in Language Models Continual Learning Neural Network Memorization Persistent Memory for Language Models Memorization in Generative Models+2

Oct 30, 2024

EF-LLM: A Continual Learning-Based Framework for AI-Assisted Automation, Enhanced Sparse Prediction, and Hallucination Detection

Accurate prediction helps achieve supply-demand balance in energy systems,thereby supporting operational decision-making and scheduling.Traditional models,which lack AI-assisted automation,often rely heavily on expert knowledge,incur high labor costs,and perform poorly under sparse-data conditions.To address these challenges,this work proposes an Energy Forecasting Large Language Model (EF-LLM),which integrates domain knowledge and temporal data for time-series forecasting while supporting both pre-forecast operations and post-forecast decision-making.EF-LLM's human-artificial intelligence (AI) interaction capabilities further lower the entry barrier for forecasting tasks and reduce the need for additional expert involvement.To support these functions, a continual learning framework with updatable Low-Rank Adaptation (LoRA) and a multi-channel architecture is developed to align heterogeneous multimodal data, enabling EF-LLM to continuously acquire and adapt multimodal knowledge.In addition,EF-LLM improves prediction accuracy under sparse-data conditions by leveraging its ability to process and fuse multimodal information.A Fusion Parameter-Efficient Fine-Tuning (F-PEFT) method is further introduced to effectively exploit both time-series data and textual information.EF-LLM is also the first energy-specific LLM to detect hallucinations and quantify their occurrence rate, achieved via multi-task learning and semantic similarity analysis.This work has achieved success in energy prediction scenarios for load, PV, and wind power forecast.

Zihang Qiu, Jayashri Ravishankar, Anam Malik +2

LLM-Based Time Series Forecasting Time Series Forecasting LLM Hallucination Detection Continual Learning for LLMs