Forgetting

Representation learning on dynamic graphs requires capturing complex dependencies that evolve across both time and structure. Existing approaches typically adopt fixed temporal decay schemes or predetermined structural propagation depths, limiting their ability to generalize across graphs with diverse interaction frequencies and topological characteristics. We propose Dual-Scale Retentive Dynamics (DSRD), a unified framework that maintains a retentive representation state encoding both temporal memory and structural context. DSRD introduces two key components: (i) a retentive state with dual-scale adaptation that jointly models temporal dynamics and structural propagation within a single recurrent formulation, and (ii) adaptive decay kernels with learnable time-sensitivity parameters that automatically balance short-term responsiveness and long-term retention based on the underlying interaction patterns. We provide theoretical analysis establishing the equivalence between event-wise parallel aggregation and efficient recurrent state updates, as well as stability and boundedness guarantees for the learned dynamics. Extensive experiments on 14 real-world benchmarks demonstrate that DSRD consistently achieves state-of-the-art performance on both link prediction and node classification tasks, with strong generalization across transductive and inductive settings.

Randomly initialized neural networks induce a prior over functions, but the predictor used in practice is produced only after training. We ask how much of this initial bias survives the training pipeline. To make the question measurable, we introduce initialization memory: the dependence of the validation-selected predictor on the scale of the random initialization. We perform controlled CIFAR-10 experiments on ResNets where initialization memory already sharply separates training regimes. Low-learning-rate SGD can interpolate while still remembering its initialization: on ResNet-9 with batch size $b=128$, test accuracy varies by $26.5$ percentage points across initialization scales despite $\ge99.5\%$ training accuracy. This is not undertraining: extending the same low-learning-rate regime to $5{,}000$ epochs leaves the spread essentially unchanged. In contrast, Adam-family methods largely erase the dependence. SGD can also be made to forget when larger learning rates are paired with explicit $L_2$ norm control. We interpret these findings in terms of the time scale of forgetting: gradient-flow-like dynamics can preserve initialization memory, whereas stochastic finite-step effects, explicit norm decay, and adaptive preconditioning erase it on scales governed by the size of explicit or implicit regularization. The practical inductive bias of a trained network is therefore not the architectural prior alone, but the architectural prior after being filtered by the forgetting dynamics of the training pipeline; and the same regularizers that improve generalization are precisely those that erase memory of initialization.

In-context learning (ICL) derives its power from enabling Large Language Models to adapt to new tasks via prompt-based reasoning alone, entirely bypassing the need for parameter updates. Existing theories primarily study ICL in single-task settings, while real-world prompts often contain sequences of heterogeneous tasks, leaving a gap in understanding whether Large Language Models implicitly perform continual learning during inference. To bridge this gap, we propose the first theoretical framework for in-context continual learning, modeling how a pretrained Transformer processes multiple sequential tasks within a single prompt through shared attention mechanisms. Focusing on linear and masked linear self-attention, we derive error expressions for model predictions under sequential task prompts and analyze their generalization and forgetting behavior. Our results reveal that standard attention mechanisms inevitably induce intertask interference by uniformly or causally aggregating historical contexts, leading to systematic bias. We further provide a bias-variance-interference decomposition of prediction error, characterizing when historical in-context information yields positive transfer or provable negative transfer. This analysis exposes fundamental limits of attention-based continual inference and offers theoretical explanations for order sensitivity and performance degradation in long prompts.

Vision-language-action (VLA) models provide a promising foundation for general-purpose robotics. However, their successful deployment in real-world scenarios requires the ability to continually acquire new skills while retaining previously learned behaviors. While pioneering research has studied the continual learning of VLA models in narrowly simulated environments, this challenge remains largely unexplored under realistic conditions. To address this limitation, we construct a real-world continual learning dataset comprising four sequential manipulation tasks, spanning rigid-object pick-and-place, contact-rich pressing, and deformable-object folding. Using this dataset, we conduct comprehensive experiments and find that VLA models suffer significant catastrophic forgetting when continually learning from heterogeneous real-world demonstrations. We then systematically evaluate experience replay and uncover key implementation factors that govern its success. In summary, this work provides the first empirical study of real-world continual VLA learning and offers practical guidance for deploying long-lived robot policies.

Major data protection regulations all mention the "right to be forgotten," and that's what pushed federated unlearning (FU) techniques forward. But one stubborn issue remains: catastrophic forgetting--you erase the target knowledge, yet somehow you also end up throwing out essential retained knowledge, which then hurts the model's global generalization. To get a better balance between unlearning effectiveness and generalization ability, we propose something called Image Feature Fusion-based Federated Client Unlearning (IFF-FCU). The idea is to bring in a linear Image Feature Fusion mechanism (Mixup) that dynamically creates mixed samples, bridging the gap between forget-distribution and retain-distribution. What this strategy does isn't just deleting a few discrete data points--it theoretically widens and regularizes the forgetting boundary. We ran extensive experiments on medical imaging benchmarks (RSNA-ICH and ISIC2018), and the results show that our approach achieves reasonably good unlearning. For instance, on the ICH dataset, IFF-FCU achieves a highly competitive Error deviation from the retrained gold standard, demonstrating robust improvements over existing baselines.

Models trained on a new task typically degrade on prior tasks, a phenomenon known as forgetting. Traditionally, mitigating forgetting has required replaying stored exemplars from prior tasks, which is often impractical. By contrast, language models can sample from their own training distribution, and we show that these self-generated samples serve as effective replay data, nearly eliminating forgetting. We find that forgetting nonetheless persists when the model has little remaining capacity: models pretrained close to saturation cannot absorb new information without overwriting prior knowledge. When capacity is not the limiting factor, low learning rates reduce forgetting but require substantially more training steps. Replay breaks this tradeoff, enabling fast, high-learning-rate finetuning without forgetting.

Dementia detection from spontaneous speech offers a scalable approach to cognitive screening, yet NLP systems remain predominantly English-centric. This limitation is especially acute in the Philippines, where Filipino-English code-switching is pervasive and no prior work has addressed NLP-based dementia detection. We present the first systematic evaluation of transformer-based dementia detection in Filipino speech and the first assessment of NeoBERT in a clinical NLP setting. To separate language from domain effects, we construct a parallel bilingual dataset of 4,000 DementiaBank-derived transcripts, with Filipino translations produced manually to preserve discourse-level markers of cognitive decline. We evaluate five model families, TF-IDF + LogReg, BERT, NeoBERT, XLM-R, and RoBERTa-Tagalog, under monolingual, zero-shot cross-lingual, and bilingual fine-tuning settings. We find that in-domain performance does not transfer across languages, with English-trained BERT dropping to Macro-F1 = 0.455 on Filipino, and that architectural modernization alone does not improve robustness. Bilingual fine-tuning, however, eliminates cross-lingual degradation across all transformer models, converging to Macro-F1 = 0.969-0.973. These results suggest that multilingual clinical NLP performance is driven primarily by linguistic coverage during training rather than model scale or architecture.

Legal NLP benchmarks evaluate models on randomly split data, implicitly assuming that legal language is stationary. We test this assumption by fine-tuning four transformer encoders -- XLM-RoBERTa (base and large) and their legal-domain variants -- on Ukrainian court decisions from three temporal epochs defined by geopolitical disruptions: pre-war (2008-2013), hybrid war (2014-2021), and full-scale invasion (2022-2026). Each model is trained on one epoch and evaluated on all three, producing a 3x3 cross-temporal generalization matrix. Four findings emerge. (1) Forward degradation is severe: models trained on pre-war data lose up to 27.2 percentage points of macro-F1 when applied to full-scale invasion era decisions. (2) The degradation is asymmetric: backward transfer (full-scale to pre-war) is substantially more robust than forward transfer, consistent with the hypothesis that legal language is additive. (3) Legal-domain pretraining (Legal-XLM-R) does not improve absolute performance but reduces forward degradation magnitude and asymmetry. (4) Chronological continual learning eliminates catastrophic forgetting for general XLM-R: pre-war knowledge is fully retained (+1.8 to +6.2 pp) while full-scale performance gains +16.5 to +19.0 pp; reverse-chronological training causes severe forgetting. Cross-jurisdictional pretraining on Swiss Judgment Prediction data improves absolute performance but does not reduce temporal degradation magnitude, confirming that temporal drift is an intrinsic property of legal language evolution. The dataset (428K decisions across three epochs) is publicly available as a LEXTREME contribution.

Fine-tuning large language models on new data improves task performance but degrades capabilities learned during pretraining, a phenomenon known as catastrophic forgetting. Existing methods mitigate this by modifying the fine-tuning objective to suppress high-loss tokens or sequences, but these tokens are essential for learning new tasks, especially those with poor pretraining coverage. In such settings, hard tokens should still contribute to learning, so forgetting must be controlled without suppressing them. We identify a simple mechanism for doing so: per-step forgetting is bounded by the product of the learning rate and the square root of the current training loss. This suggests that high-loss batches are especially prone to inducing forgetting. Motivated by this observation, we introduce FINCH, a loss-adaptive learning-rate schedule that reduces the learning rate on high-loss batches and increases it as the model converges, while leaving the fine-tuning objective unchanged. Across knowledge acquisition, science, and low-resource language adaptation benchmarks, FINCH reduces forgetting by 93% on average while matching the task performance of standard fine-tuning. On Qwen3-4B knowledge acquisition, FINCH cuts TruthfulQA degradation by 5x and reverses HaluEval degradation, while better preserving confidence calibration. Overall, our results show that learning-rate schedules are an effective tool to shape model behavior during fine-tuning, beyond just target-task optimization.

Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.

Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.

Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite every deployed language model being quantized first. Recent work has shown that 4-bit post-training quantization can reverse machine unlearning; we show this is not a tuning artefact but a systematic dual failure: gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures trace to the same root cause: across all baselines, per-parameter updates lie 47-828x below the NF4 quantization bin width; updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a sparsity-permanence tradeoff. We present MANSU (Mechanistic-Aligned Null-Space Unlearning), which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction. We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression, a distinction existing metrics cannot make. Across multiple model families and hazard benchmarks, MANSU is the first method to jointly satisfy all four properties with margin on each (meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure), while gradient-based baselines recover up to +0.05 accuracy under compression.

Neural networks suffer from catastrophic forgetting in class-incremental learning (CIL) settings. Rehearsal$\unicode{x2013}$replaying a subset of past samples$\unicode{x2013}$is a well-established mitigation strategy. However, recent results suggest that, despite balanced rehearsal allocation, some classes are forgotten substantially more than others. Despite its relevance, this imbalanced forgetting phenomenon remains underexplored. This work shows that imbalanced forgetting arises systematically and severely in rehearsal-based CIL and investigates it extensively. Specifically, we construct, from a principled analysis, three last-layer coefficients that capture different gradient-level sources of interference affecting each past class during an incremental step. We then demonstrate that, together, they reliably predict how past classes will rank in terms of forgetting at the end of that step. While predictive performance alone does not establish causality, these results support the interpretation of the coefficients as a plausible mechanistic account linking last-layer gradient-level interactions during training to class-level forgetting outcomes. Notably, one coefficient$\unicode{x2013}$capturing self-induced interference$\unicode{x2013}$emerges as the strongest predictor, with controlled experiments providing evidence consistent with this coefficient being influenced by the new-class interference coefficient. Overall, our findings provide valuable insights and suggest promising directions for mitigating imbalanced forgetting by reducing class-wise disparities in the identified sources of interference.

Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance level, making it difficult to precisely remove target knowledge without affecting unrelated semantics. This issue is especially pronounced since a single image often contains multiple entangled concepts, including both target concepts to be forgotten and contextual information that should be preserved. In this paper, we propose an interpretable concept-level unlearning framework for VLMs, which constructs a compact task-specific concept vocabulary from the forgetting set using a multimodal large language model. In addition to modality alignment, visual representations are decomposed into sparse, nonnegative combinations of semantic concepts, providing an explicit interface for fine-grained knowledge manipulation. Based on this decomposition, our method formulates unlearning as concept-level optimization, where target concepts are selectively suppressed while intra-instance non-target semantics and global cross-modal knowledge are preserved. Extensive experiments across both in-domain and out-of-domain forgetting settings demonstrate that our method enables more comprehensive target forgetting, better preserves non-target knowledge within the same image, and maintains competitive model utility compared with existing VLM unlearning methods.

How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.

Continual learning studies how models can adapt to new tasks while retaining previously acquired knowledge. Although a broad spectrum of methods has been proposed to mitigate catastrophic forgetting, the field remains predominantly performance-driven, with limited insight into what forgetting actually corresponds to within the vision model's representation space. Prior work has primarily analyzed forgetting through task-level performance or coarse measures of representational drift, without disentangling output-level accessibility from changes in finer-grained internal structure. To this end, we propose a diagnostic framework that leverages Sparse Autoencoders (SAEs) to define a task-anchored latent feature space, enabling analysis of how task-specific information evolves at a finer granularity, where individual SAE latents are treated as concept proxies for recurring and relatively disentangled visual patterns in the model's internal computations. Within this framework, we decompose forgetting into apparent concept deletion, recoverability, and decodability. We show that a large portion of seemingly lost concept-level information can often be recovered under linearity assumption, with concept decodability degrading as more tasks are introduced. Overall, our findings suggest that a significant part of concept-level forgetting can be attributed to changes in the representational accessibility rather than complete information erasure.

Large language models confidently produce outdated answers, and no existing method can detect them. We show this is not an engineering failure but a structural one: temporal drift, whether a stored fact has changed since training, is encoded as a direction in the residual stream geometrically orthogonal to both correctness and uncertainty. Any method operating on correctness or uncertainty signals is therefore blind to drift by construction. We verify this across six instruction-tuned models. A linear probe trained directly on drift labels achieves AUROC $0.83$--$0.95$; methods based on token entropy, semantic entropy, CCS, and SAPLMA all remain near chance ($0.49$--$0.57$). Five tests confirm the geometric orthogonality: weight cosines ($|\cos| \leq 0.14$), score correlations ($|r| \leq 0.20$), bidirectional null-space projection ($|Δ| \leq 0.008$), iterative null-space projection with $k{=}10$, and difference-of-means dissociation. Mechanistically, the MLP retrieval circuit produces identical dynamics for stale recall and confabulation ($r > 0.81$, six models), explaining why output confidence cannot separate them. A cross-cutoff experiment holds inputs constant and varies only the model: the probe fires on the model whose training predates the fact's transition and stays silent otherwise ($P(A{>}B) = 0.975$--$0.998$, twelve model pairs), confirming it reads model-internal knowledge state rather than input properties. Our code and datasets will be publicly released.

Unlearning in large language models (LLMs) aims to remove harmful training data while preserving overall utility. However, we find that existing methods often hallucinate, generate abnormal token sequences, or behave inconsistently, raising safety and trust concerns. According to prior literature on LLM honesty, such behaviors are often associated with dishonesty. This motivates us to investigate the notion of honesty in the context of model unlearning. We propose a formal definition of unlearning honesty, which includes: (1) preserving both utility and honesty on retained knowledge, and (2) ensuring effective forgetting while encouraging the model to acknowledge its limitations and respond consistently to questions related to forgotten knowledge. To systematically evaluate the honesty of unlearning, we introduce a suite of metrics that cover utility, honesty on the retained set, effectiveness of forgetting, rejection rate and refusal stability in Q&A and MCQ settings. Evaluating 9 methods across 3 mainstream families shows that all current methods fail to meet these standards. After experimental and theoretical analyses, we present ReVa, a representation-alignment procedure that fine-tunes feature-randomized unlearned models to better acknowledge forgotten knowledge. On Q&A tasks from the forget set, ReVa achieves the highest rejection rate after two rounds of interaction, nearly doubling the performance of the second-best method. Remarkably, It also improves honesty on the retained set. We release our data and code at https://github.com/renjiegu.

Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.

Continual learning (CL) is essential for deploying medical image segmentation models in clinical environments where imaging domains, anatomical targets, and diagnostic tasks evolve over time. However, continual segmentation still faces three main challenges. First, the scenarios for this task remain insufficiently standardized for real-world clinical settings. Second, existing research has been primarily focused on mitigating forgetting, overlooking the other essential properties such as plasticity. Third, a benchmark work with comprehensive evaluation on existing methods is stll desirable. To address these gaps, we present such benchmark study of continual medical image segmentation. We first define three clinically motivated scenarios, namely Domain-CL, Class-CL, and Organ-CL, to respectively capture the cross-center domain shift, the incremental anatomical structure segmentation, and the cross-organ segmentation. We then introduce an evaluation framework that measures not only general performance and forgetting, but also plasticity, forward generalizability, parameter efficiency, and replay burden. The results, from extensive experiments with representative CL methods, showed that it was still challenging to develop a model that could satisfy all the requirements simultaneously. Nevertheless, these studies also suggested that the replay-based methods achieve the best overall balance between stability and plasticity, the parameter-isolation methods should be effective at reducing forgetting, though at the cost of increased model size, and the forward generalizability remain a significantly understudied aspect of this research field. Finally, we discuss related learning paradigms and outline future directions for continual medical image segmentation.

Neural operators perform well on structured domains, yet their behaviour on irregular geometries remains poorly understood. We show that this limitation is not merely an encoding issue, but a depth-wise failure mode inherent to deep operator architectures. We formalise the Geometric Forgetting Hypothesis: due to the Markovian structure of operator layers and their reliance on global mixing mechanisms, neural operators progressively lose access to domain geometry as depth increases. Using layer-wise geometric probing, we demonstrate that both spectral and attention-based operators systematically lose geometric fidelity. We show that this geometric forgetting degrades accuracy, stability, and generalisation. To counteract it, we introduce a lightweight geometry memory injection mechanism that restores geometric constraints at intermediate depths with minimal architectural overhead. This simple intervention consistently mitigates forgetting and exposes a geometric shortcut instability in transformer-based operators, revealing that geometric retention is a structural requirement rather than a design choice.

Generating realistic 3D Human-Object Interactions (HOI) is a fundamental task for applications ranging from embodied AI to virtual content creation, which requires harmonizing high-level semantic intent with strict low-level physical constraints. Existing methods excel at semantic alignment, however, they struggle to maintain precise object contact. We reveal a key finding termed \textit{Geometric Forgetting}: as diffusion model depth increases, semantic feature tend to overshadow object geometry feature, causing the model to lose its perception to object geometry. To address this, we propose MaMi-HOI, a hierarchical framework reconciling \textbf{Ma}cro-level kinematic fluidity with \textbf{Mi}cro-level spatial precision. First, to counteract geometric forgetting, we introduce the Geometry-Aware Proximity Adapter (GAPA), which explicitly re-injects dense object details to perform residual snapping corrections for precise contact. Nevertheless, such aggressive local enforcement can disrupt global dynamics, leading to robotic stiffness. In response, we introduce the Kinematic Harmony Adapter (KHA), which proactively aligns whole-body posture with spatial objectives, ensuring the skeleton actively accommodates constraints without compromising naturalness. Extensive experiments validate that MaMi-HOI simultaneously achieves natural motion and precise contact. Crucially, it extends generation capabilities to long-term tasks with complex trajectories, effectively bridging the gap between global navigation and high-fidelity manipulation in 3D scenes. Code is available at https://github.com/DON738110198/MaMi-HOI.git

Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.

While Large Vision-Language Models (LVLMs) offer powerful capabilities, they pose privacy risks by unintentionally memorizing sensitive personal information. Current unlearning benchmarks attempt to mitigate this using fictitious identities but overlook a critical stage 1 failure: models fail to effectively memorize target information initially, rendering subsequent unlearning evaluations unreliable. Diagnosing under-memorization and the multi-hop curse as root causes, we introduce ReMem, a Reliable Multi-hop and Multi-image Memorization Benchmark. ReMem ensures robust foundational learning through principled data scaling, reasoning-aware QA pairs, and diverse visual contexts. Additionally, we propose a novel Exposure metric to quantify the depth of information erasure from the model's internal probability distribution. Extensive experiments demonstrate that ReMem provides a rigorous and trustworthy framework for diagnosing both learning and unlearning behaviors in LVLMs.

Large Vision-Language Models (LVLMs), trained on web-scale data, risk memorizing and regenerating copyrighted visual content such as characters and logos, creating significant challenges. Machine unlearning offers a path to mitigate these risks by removing specific content post-training, but evaluating its effectiveness, especially in the complex multimodal setting of LVLMs, remains an open problem. Current evaluation methods often lack robustness or fail to capture the nuances of cross-modal concept erasure. To address this critical gap, we introduce the CoVUBench benchmark, the first framework specifically designed for evaluating copyright content unlearning in LVLMs. CoVUBench utilizes procedurally generated, legally safe synthetic data coupled with systematic visual variations spanning compositional changes and diverse domain manifestations to ensure realistic and robust evaluation of unlearning generalization. Our comprehensive multimodal evaluation protocol assesses both forgetting efficacy from the copyright holder perspective and the preservation of general model utility from the deployer viewpoint. By rigorously measuring this crucial trade-off, CoVUBench provides a standardized tool to advance the development of responsible and effective unlearning methods for LVLMs.

Adapting a pretrained language model to a new task often hurts the general capabilities it already had, a problem known as catastrophic forgetting. Sparse Memory Finetuning (SMF) tries to avoid this by adding key-value memory layers to the model and, on each training step, updating only the small set of memory rows that the current batch reads most heavily. We re-implement SMF on Qwen-2.5-0.5B-Instruct and compare it with LoRA and full finetuning on MedMCQA, a 4-choice medical exam task, using WikiText perplexity and TriviaQA accuracy as forgetting probes. SMF improves MedMCQA by 2.5 percentage points while keeping both forgetting probes within roughly 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift on both. We also compare two row-selection rules (KL-divergence and TF-IDF), which balance the two forgetting metrics differently.

Pretraining optimizers are tuned to produce the strongest possible base model, on the assumption that a stronger starting point yields a stronger model after subsequent changes like post-training and quantization. This overlooks the geometry of the base model which controls how much of the base model's capabilities survive subsequent parameter updates. We study three pretraining optimization approaches that bias optimization toward flatter minima: Sharpness-Aware Minimization (SAM), large learning rates, and shortened learning rate annealing periods. Across model sizes ranging from 20M to 150M parameters, we find that these interventions consistently improve downstream performance after post-training on five common datasets with up to 80% less forgetting. These principles hold at scale: a short SAM mid-training phase applied to an existing OLMo-2-1B checkpoint reduces forgetting by 31% after MetaMath post-training and by 40% after 4-bit quantization.

Differentiable physical networks provide a simple setting in which learning can be studied through the interaction between trainable parameters and physical equilibrium constraints. We investigate sequential learning in differentiable resistor networks governed by Kirchhoff's laws. Although individual input--output mappings can be learned by gradient-based adjustment of edge conductances, sequential training on conflicting tasks produces catastrophic forgetting. We show that forgetting is controlled by task conflict and by the degree of adaptation to the new task. Uniform anchoring and normalised gradient-weighted anchoring reduce forgetting only by increasing the final loss on the new task, giving a clear forgetting--adaptation trade-off. We also show that forgetting is associated with localised conductance changes on high-current edges, giving a physical interpretation as reconfiguration of dominant transport pathways. Broader random-task ensembles show that the strongest forgetting occurs when the second task reverses the output ordering imposed by the first task. Finally, comparisons across Erdős--Rényi, small-world, scale-free, and random-geometric graph ensembles show that topology changes the forgetting--adaptation balance. These results position differentiable resistor networks as compact, physically interpretable testbeds for studying continual learning in tunable matter.

Federated Multimodal Learning (FML) trains multimodal models across decentralized clients while keeping their image-text pairs private. However, joint embedding training entangles forgotten knowledge across both modalities and client gradient subspaces, hindering federated unlearning. Previous federated unlearning approaches neither sever the cross-modal reconstruction channel mediated by bilinear coupling nor separate forget-exclusive update directions from those shared with retained clients. We identify an Anchor Principle for federated multimodal contrastive unlearning: forgotten alignments persist through three residual anchors arising from bilinear cross-modal coupling, principal-angle subspace entanglement, and continued federated updates. At the modality level, we show that bilateral displacement of both visual and language branches closes the cross-modal reconstruction channel. Correspondingly, our method addresses subspace entanglement through Cosine--Sine decomposition of client-update subspaces, isolating forget-exclusive directions from retain support. Moreover, we propose a direction-selective Forget Lock that bounds residual drift across rounds. Combining these strategies, we present EASE, an Entanglement-Aware Subspace Excision framework that closes all three anchor channels under a unified design. EASE demonstrates consistent superiority across multiple datasets and unlearning scenarios, for instance, matching the retrain reference to within 0.2 and 4.2 R@1 points on the forget and retain sides under client unlearning on Flickr30K with CLIP-B/32.

Continual learning agents with finite capacity must balance acquiring new knowledge with retaining the old. This requires controlled forgetting of knowledge that is no longer needed, freeing up capacity to learn. Weight decay, viewed as a mechanism for forgetting, can serve this role by gradually discarding information stored in the weights. However, a fixed scalar weight decay drives this forgetting uniformly over time and uniformly across all parameters, even when some encode stable knowledge while others track rapidly changing targets. We introduce Forgetting through Adaptive Decay (FADE), which adapts per-parameter weight decay rates online via approximate meta-gradient descent. We derive FADE for the online linear setting and apply it to the final layer of neural networks. Our empirical analysis shows that FADE automatically discovers distinct decay rates for different parameters, complements step-size adaptation, and consistently improves over fixed weight decay across online tracking and streaming classification problems.

Large language models (LLMs) often memorize private information during training, raising serious privacy concerns. While machine unlearning has emerged as a promising solution, its true effectiveness against privacy attacks remains unclear. To address this, we propose PrivUn, a new evaluation framework that systematically assesses unlearning robustness through three-tier attack scenarios: direct retrieval, in-context learning recovery, and fine-tuning restoration; combined with quantitative analysis using forgetting scores, association metrics, and forgetting depth assessment. Our study exposes significant weaknesses in current unlearning methods, revealing two key findings: 1) unlearning exhibits gradient-driven ripple effects: unlike traditional forgetting which follows semantic relations (e.g., knowledge graphs), privacy unlearning propagates across latent gradient-based associations; and 2) most methods suffer from shallow forgetting, failing to remove private information distributed across multiple deep model layers. To validate these insights, we explore two strategies: association-aware core-set selection that leverages gradient similarity, and multi-layer deep intervention through representational constraints. These strategies represent a paradigm shift from shallow forgetting to deep forgetting.

Streaming Continual Learning (CL) typically converts a continuous stream into a sequence of discrete tasks through temporal partitioning. We argue that this temporal taskification step is not a neutral preprocessing choice, but a structural component of evaluation: different valid splits of the same stream can induce different CL regimes and therefore different benchmark conclusions. To study this effect, we introduce a taskification-level framework based on plasticity and stability profiles, a profile distance between taskifications, and Boundary-Profile Sensitivity (BPS), which diagnoses how strongly small boundary perturbations alter the induced regime before any CL model is trained. We evaluate continual finetuning, Experience Replay, Elastic Weight Consolidation, and Learning without Forgetting on network traffic forecasting with CESNET-Timeseries24, keeping the stream, model, and training budget fixed while varying only the temporal taskification. Across 9-, 30-, and 44-day splits, we observe substantial changes in forecasting error, forgetting, and backward transfer, showing that taskification alone can materially affect CL evaluation. We further find that shorter taskifications induce noisier distribution-level patterns, larger structural distances, and higher BPS, indicating greater sensitivity to boundary perturbations. These results show that benchmark conclusions in streaming CL depend not only on the learner and the data stream, but also on how that stream is taskified, motivating temporal taskification as a first-class evaluation variable.

Continual learning (CL) studies how models acquire tasks sequentially while retaining previously learned knowledge. Despite substantial progress in benchmarking CL methods, comparative evaluations typically keep the fine-tuning regime fixed. In this paper, we argue that the fine-tuning regime, defined by the trainable parameter subspace, is itself a key evaluation variable. We formalize adaptation regimes as projected optimization over fixed trainable subspaces, showing that changing the trainable depth alters the effective update signal through which both current task fitting and knowledge preservation operate. This analysis motivates the hypothesis that method comparisons need not be invariant across regimes. We test this hypothesis in task incremental CL, five trainable depth regimes, and four standard methods: online EWC, LwF, SI, and GEM. Across five benchmark datasets, namely MNIST, Fashion MNIST, KMNIST, QMNIST, and CIFAR-100, and across 11 task orders per dataset, we find that the relative ranking of methods is not consistently preserved across regimes. We further show that deeper adaptation regimes are associated with larger update magnitudes, higher forgetting, and a stronger relationship between the two. These results show that comparative conclusions in CL can depend strongly on the chosen fine-tuning regime, motivating regime-aware evaluation protocols that treat trainable depth as an explicit experimental factor.

Personalized agents that interact with users over long periods must maintain persistent memory across sessions and update it as circumstances change. However, existing benchmarks predominantly frame long-term memory evaluation as fact retrieval from past conversations, providing limited insight into agents' ability to consolidate memory over time or handle frequent knowledge updates. We introduce Memora, a long-term memory benchmark spanning weeks to months long user conversations. The benchmark evaluates three memory-grounded tasks: remembering, reasoning, and recommending. To ensure data quality, we employ automated memory-grounding checks and human evaluation. We further introduce Forgetting-Aware Memory Accuracy (FAMA), a metric that penalizes reliance on obsolete or invalidated memory when evaluating long-term memory. Evaluations of four LLMs and six memory agents reveal frequent reuse of invalid memories and failures to reconcile evolving memories. Memory agents offer marginal improvements, exposing shortcomings in long-term memory for personalized agents.

Knowledge Graph Embeddings (KGEs) support a wide range of downstream tasks over Knowledge Graphs (KGs). In practice, KGs evolve as new entities and facts are added, motivating Continual Knowledge Graph Embedding (CKGE) methods that update embeddings over time. Current CKGE approaches address catastrophic forgetting (i.e., the performance degradation on previously learned tasks) primarily by limiting changes to existing embeddings. However, we show that this view is incomplete. When new entities are introduced, their embeddings can interfere with previously learned ones, causing the model to predict them in place of previously correct answers. This phenomenon, which we call entity interference, has been largely overlooked and is not accounted for in current CKGE evaluation protocols. As a result, the assessment of catastrophic forgetting becomes misleading, and CKGE methods performance is systematically overestimated. To address this issue, we introduce a corrected CKGE evaluation protocol that accounts for entity interference. Through experiments on multiple benchmarks, we show that ignoring this effect can lead to performance overestimation of up to 25%, particularly in scenarios with significant entity growth. We further analyze how different CKGE methods and KGE models are affected by the different sources of forgetting, and introduce a catastrophic forgetting metric tailored to CKGE.

As a means to balance the growth of the AI industry with the need for privacy protection, machine unlearning plays a crucial role in realizing the ``right to be forgotten'' in artificial intelligence. This technique enables AI systems to remove the influence of specific data while preserving the rest of the learned knowledge. Although it has been actively studied, most existing unlearning methods assume that unlearning is performed only once. In this work, we evaluate existing unlearning algorithms in a more realistic scenario where unlearning is conducted repeatedly, and in this setting, we identify two critical phenomena: (1) Knowledge Erosion, where the accuracy on retain data progressively degrades over unlearning phases, and (2) Forgetting Reversal, where previously forgotten samples become recognizable again in later phases. To address these challenges, we propose SAFER (StAbility-preserving Forgetting with Effective Regularization), a continual unlearning framework that maintains representation stability for retain data while enforcing negative logit margins for forget data. Extensive experiments show that SAFER mitigates not only knowledge erosion but also forgetting reversal, achieving stable performance across multiple unlearning phases.

In continual learning, the primary challenge is to learn new information without forgetting old knowledge. A common solution addresses this trade-off through regularization, penalizing changes to parameters critical for previous tasks. In most cases, this regularization term is directly added to the training loss and optimized with standard gradient descent, which blends learning and retention signals into a single update and does not explicitly separate essential parameters from redundant ones. As task sequences grow, this coupling can over-constrain the model, limiting forward transfer and leading to inefficient use of capacity. We propose a different approach that separates task learning from stability enforcement via operator splitting. The learning step focuses on minimizing the current task loss, while a proximal stability step applies a sparse regularizer to prune unnecessary parameters and preserve task-relevant ones. This turns the stability-plasticity into a negotiated update between two complementary operators, rather than a conflicting gradient. We provide theoretical justification for the splitting method on the continual-learning objective, and demonstrate that our proposed solver achieves state-of-the-art results on standard benchmarks, improving both stability and adaptability without the need for replay buffers, Bayesian sampling, or meta-learning components.

The inability to filter out in advance all potentially problematic data from the pre-training of large language models has given rise to the need for methods for unlearning specific pieces of knowledge after training. Existing techniques overlook the need for continuous and immediate action, causing them to suffer from degraded utility as updates accumulate and protracted exposure of sensitive information. To address these issues, we propose Continual Unlearning in Real Time with Ensured Preservation of LLM Knowledge (CURaTE). Our method begins by training a sentence embedding model on a dataset designed to enable the formation of sharp decision boundaries for determining whether a given input prompt corresponds to any stored forget requests. The similarity of a given input to the forget requests is then used to determine whether to answer or return a refusal response. We show that even with such a simple approach, not only does CURaTE achieve more effective forgetting than existing methods, but by avoiding modification of the language model parameters, it also maintains near perfect knowledge preservation over any number of updates and is the only method capable of continual unlearning in real-time.

Audio-visual continual test-time adaptation involves continually adapting a source audio-visual model at test-time, to unlabeled non-stationary domains, where either or both modalities can be distributionally shifted, which hampers online cross-modal learning and eventually leads to poor accuracy. While previous works have tackled this problem, we find that SOTA methods suffer from catastrophic forgetting where the model's performance drops well below even the source model due to continual parameter updates at test-time. In this work, we first show that adapting only the modality fusion layer to a target domain not only improves performance on that domain but can also enhance performance on subsequent domains. Based on this strong cross-task transferability of the fusion layer's parameters, we propose a method, $\texttt{AVReCAP}$, that improves test-time performance of the models without access to any source data. Our approach works by using a selective parameter retrieval mechanism that dynamically retrieves the best fusion layer parameters from a buffer using only a small batch of test data. These parameters are then integrated into the model, adapted to the current test distribution, and saved back for future use. Extensive experiments on benchmark datasets involving unimodal and bimodal corruptions show our proposed $\texttt{AVReCAP}$ significantly outperforms existing methods while minimizing catastrophic forgetting.

Parameter-efficient continual learning aims to adapt pre-trained models to sequential tasks without forgetting previously acquired knowledge. Most existing approaches treat continual learning as avoiding interference with past updates, rather than considering what properties make the current task-specific update naturally preserve previously acquired knowledge. From a knowledge-decomposition perspective, we observe that low-rank adaptations exhibit highly imbalanced singular value spectra: a few dominant components absorb most of the adaptation energy, thereby (i) more likely to disrupt previously acquired knowledge and (ii) making the update more vulnerable to interference from subsequent tasks. To enable explicit balance among components, we decouple the magnitude of the task update from its directional structure and formulate it as a constrained optimization problem on a restricted Stiefel manifold. We address this problem using a projected first-order method compatible with standard deep-learning optimizers used in vision-language models. Our method mitigates both backward and forward forgetting, consistently outperforming continual learning baselines. The implementation code is available at https://github.com/haodotgu/EBLoRA.

Multimodal large language models (MLLMs) can inadvertently memorize privacy-sensitive information during training. While existing unlearning methods can remove such content, they often severely degrade the model's foundational capabilities, such as general image understanding. This critical shortfall motivates our investigation into benign memory forgetting, the precise removal of targeted, privacy-sensitive knowledge while rigorously preserving unrelated capabilities. To pioneer and evaluate progress toward this objective, we introduce S-MLLMUn Bench, the first benchmark designed to jointly and quantitatively assess an unlearning method's efficacy in knowledge erasure and the preservation of image understanding. Furthermore, we propose the Sculpted Memory Forgetting Adapter (SMFA), a new framework that enables benign memory forgetting. SMFA confines forgetting to designated memory regions, maintaining overall model performance. By initially fine-tuning the model to replace sensitive outputs with refusals, SMFA generates a memory forgetting adapter, followed by a retaining anchor-guided masking mechanism that safeguards unrelated knowledge. Extensive experiments on S-MLLMUn Bench demonstrate that existing methods fail to achieve benign forgetting, whereas our proposed SMFA serves as an effective baseline, successfully achieving targeted knowledge erasure without compromising the model's foundational visual capabilities. Code and data are available at https://github.com/zeng-zhen/S-MLLMUn.

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.

Data unlearning aims to remove the influence of specific training samples from a trained model. In fine-tuning methods, data unlearning relies primarily on loss maximization over forget samples, which often leads to quality degradation or incomplete forgetting. Existing methods perform unlearning uniformly across diffusion stages, ignoring diffusion dynamics from noise to data. Our systematic study of diffusion phases shows that forgetting in diffusion models is uneven across time and frequency, with theoretical justification of distributive distortion and forgetting-utility trade-off. By selectively forgetting time and frequency in diffusion models, we achieve both higher unlearning success rates and improved generation quality across diverse settings, including both conditional and unconditional scenarios. We also introduce an improved SSCD metric that measures dissimilarity using a normalized perturbation distance. Together, we provide practical insights for understanding and improving data unlearning in diffusion models.

As Large Language Models (LLMs) demonstrate remarkable capabilities learned from vast corpora, concerns regarding data privacy and safety are receiving increasing attention. LLM unlearning, which aims to remove the influence of specific data while preserving overall model utility, is becoming an important research area. One of the mainstream unlearning classes is optimization-based methods, which achieve forgetting directly through fine-tuning, exemplified by Negative Preference Optimization (NPO). However, NPO's effectiveness is limited by its inherent lack of explicit positive preference signals. Attempts to introduce such signals by constructing preferred responses often necessitate domain-specific knowledge or well-designed prompts, fundamentally restricting their generalizability. In this paper, we shift the focus to the distribution-level, directly targeting the next-token probability distribution instead of entire responses, and derive a novel unlearning algorithm termed \textbf{Di}stribution \textbf{P}reference \textbf{O}ptimization (DiPO). We show that the requisite preference distribution pairs for DiPO, which are distributions over the model's output tokens, can be constructed by selectively amplifying or suppressing the model's high-confidence output logits, thereby effectively overcoming NPO's limitations. We theoretically prove the consistency of DiPO's loss function with the desired unlearning direction. Extensive experiments demonstrate that DiPO achieves a strong trade-off between model utility and forget quality. Notably, DiPO attains the highest forget quality on the TOFU benchmark, and maintains leading scalability and sustainability in utility preservation on the MUSE benchmark.

Modern LLM-based agents operate through a harness of tools, reusable skills, and specialist agents that shapes what they observe and what they can do. In practice, this harness continually evolves as new capabilities are added. We introduce EVOHARNESSBENCH, a benchmark for evaluating agents under controlled harness evolution across three axes (tools, skills, and agents). Unlike existing continual-learning benchmarks for agents, which typically place non-stationarity (i.e., what changes over time) in the task stream while keeping the harness fixed, EVOHARNESSBENCH places non-stationarity in the externally supplied harness itself. It contains 17 multi-stage harness streams constructed deterministically from verifier-based benchmarks, comprising 802 tasks, 520 tools, 42 skills, and 62 agents. We evaluate two complementary settings corresponding to the central challenges of harness evolution: deployment evaluation, which isolates retention of previously accessible competence as the harness expands, and self-evolving adaptation evaluation, which tests whether accumulated experience remains useful as new capabilities are introduced. Our results reveal three persistent gaps. First, harness expansion alone can degrade performance on previously solved tasks, producing harness-induced forgetting. Second, gains from self-evolving adaptation remain inconsistent across stages of harness evolution, capability axes, and environments. Third, retention and adaptation can pull in different directions: preserving earlier competence does not necessarily improve adaptation to newly introduced capabilities, and vice versa. These results establish harness evolution as a distinct challenge for building agents that can keep pace with an evolving harness while preserving previously effective behavior.