Daily Papers

Many studies have shown various biases targeting different demographic groups in language models, amplifying discrimination and harming fairness. Recent parameter modification debiasing approaches significantly degrade core capabilities such as text coherence and task accuracy. And Prompt-based debiasing methods, only effective for predefined trigger words, fail to address deeply embedded stereotypical associations in model parameters. In this paper, we propose BiasUnlearn, a novel model debiasing framework which achieves targeted debiasing via dual-pathway unlearning mechanisms coordinating stereotype forgetting with anti-stereotype retention, while preventing bias polarity reversal through adversarial forget set and dynamic dataset swapping. We conducted extensive experiments with multiple language models across various evaluation benchmarks. The results show that BiasUnlearn outperforms existing methods in mitigating bias in language models while retaining language modeling capabilities. Further experiments reveal that debiasing weights are transferable across model variants, confirming that bias representations become entrenched during pre-training and persist through fine-tuning phases.

We consider representation misdirection (RM), a class of LLM unlearning methods that achieves forgetting by manipulating the forget-representations, that is, latent representations of forget samples. Despite being important, the roles of target vectors used in RM, however, remain underexplored. Here, we approach and revisit RM through the lens of the linear representation hypothesis. Specifically, if one can somehow identify a one-dimensional representation corresponding to a high-level concept, the linear representation hypothesis enables linear operations on this concept vector within the forget-representation space. Under this view, we hypothesize that, beyond forgetting, machine unlearning elicits controllable side behaviors and stronger side capabilities corresponding to the high-level concept. Our hypothesis is empirically validated across a wide range of tasks, including behavioral control (e.g., controlling unlearned models' truth, sentiment, and refusal) and capability enhancement (e.g., improving unlearned models' in-context learning capability). Our findings reveal that this fairly attractive phenomenon could be either a hidden risk if misused or a mechanism that can be harnessed for developing models that require stronger capabilities and controllable behaviors.

A stereotype is an over-generalized belief about a particular group of people, e.g., Asians are good at math or Asians are bad drivers. Such beliefs (biases) are known to hurt target groups. Since pretrained language models are trained on large real world data, they are known to capture stereotypical biases. In order to assess the adverse effects of these models, it is important to quantify the bias captured in them. Existing literature on quantifying bias evaluates pretrained language models on a small set of artificially constructed bias-assessing sentences. We present StereoSet, a large-scale natural dataset in English to measure stereotypical biases in four domains: gender, profession, race, and religion. We evaluate popular models like BERT, GPT-2, RoBERTa, and XLNet on our dataset and show that these models exhibit strong stereotypical biases. We also present a leaderboard with a hidden test set to track the bias of future language models at https://stereoset.mit.edu

Many approximate machine unlearning methods demonstrate strong logit-level forgetting -- such as near-zero accuracy on the forget set -- yet continue to preserve substantial information within their internal feature representations. We refer to this discrepancy as superficial forgetting. Recent studies indicate that most existing unlearning approaches primarily alter the final classifier, leaving intermediate representations largely unchanged and highly similar to those of the original model. To address this limitation, we introduce the Erase at the Core (EC), a framework designed to enforce forgetting throughout the entire network hierarchy. EC integrates multi-layer contrastive unlearning on the forget set with retain set preservation through deeply supervised learning. Concretely, EC attaches auxiliary modules to intermediate layers and applies both contrastive unlearning and cross-entropy losses at each supervision point, with layer-wise weighted losses. Experimental results show that EC not only achieves effective logit-level forgetting, but also substantially reduces representational similarity to the original model across intermediate layers. Furthermore, EC is model-agnostic and can be incorporated as a plug-in module into existing unlearning methods, improving representation-level forgetting while maintaining performance on the retain set.

LLMs are increasingly transforming our digital ecosystem, but they often inherit societal biases learned from their training data, for instance stereotypes associating certain attributes with specific identity groups. While whether and how these biases are mitigated may depend on the specific use cases, being able to effectively detect instances of stereotype perpetuation is a crucial first step. Current methods to assess presence of stereotypes in generated language rely on simple template or co-occurrence based measures, without accounting for the variety of sentential contexts they manifest in. We argue that understanding the sentential context is crucial for detecting instances of generalization. We distinguish two types of generalizations: (1) language that merely mentions the presence of a generalization ("people think the French are very rude"), and (2) language that reinforces such a generalization ("as French they must be rude"), from non-generalizing context ("My French friends think I am rude"). For meaningful stereotype evaluations, we need to reliably distinguish such instances of generalizations. We introduce the new task of detecting generalization in language, and build GeniL, a multilingual dataset of over 50K sentences from 9 languages (English, Arabic, Bengali, Spanish, French, Hindi, Indonesian, Malay, and Portuguese) annotated for instances of generalizations. We demonstrate that the likelihood of a co-occurrence being an instance of generalization is usually low, and varies across different languages, identity groups, and attributes. We build classifiers to detect generalization in language with an overall PR-AUC of 58.7, with varying degrees of performance across languages. Our research provides data and tools to enable a nuanced understanding of stereotype perpetuation, a crucial step towards more inclusive and responsible language technologies.

Large language model (LLM) post-training enhances latent skills, unlocks value alignment, improves performance, and enables domain adaptation. Unfortunately, post-training is known to induce forgetting, especially in the ubiquitous use-case of leveraging third-party pre-trained models, which is typically understood as a loss of parametric or factual knowledge. We argue that this accuracy-centric view is insufficient for modern foundation models and instead define forgetting as systematic model drift that degrades behavior and user experience. In this context, we introduce CapTrack, a capability-centric framework for analyzing forgetting in LLMs that combines a behavioral taxonomy with an evaluation suite built on established benchmarks and targeted adaptations. Using CapTrack, we conduct a large-scale empirical study across post-training algorithms, domains, and model families, including models up to 80B parameters. We find that forgetting extends beyond parametric knowledge, with pronounced drift in robustness and default behaviors. Instruction fine-tuning induces the strongest relative drift, while preference optimization is more conservative and can partially recover lost capabilities. Differences across model families persist, and no universal mitigation emerges.

When language models (LMs) are trained to forget (or "unlearn'') a skill, how precisely does their behavior change? We study the behavior of transformer LMs in which tasks have been forgotten via fine-tuning on randomized labels. Such LMs learn to generate near-random predictions for individual examples in the "training'' set used for forgetting. Across tasks, however, LMs exhibit extreme variability in whether LM predictions change on examples outside the training set. In some tasks (like entailment classification), forgetting generalizes robustly, and causes models to produce uninformative predictions on new task instances; in other tasks (like physical commonsense reasoning and scientific question answering) forgetting affects only the training examples, and models continue to perform the "forgotten'' task accurately even for examples very similar to those that appeared in the training set. Dataset difficulty is not predictive of whether a behavior can be forgotten; instead, generalization in forgetting is (weakly) predicted by the confidence of LMs' initial task predictions and the variability of LM representations of training data, with low confidence and low variability both associated with greater generalization. Perhaps most surprisingly, random-label forgetting appears to be somewhat insensitive to the contents of the training set: for example, models trained on science questions with random labels continue to answer other science questions accurately, but begin to produce random labels on entailment classification tasks. Finally, we show that even generalizable forgetting is shallow: linear probes trained on LMs' representations can still perform tasks reliably after forgetting. Our results highlight the difficulty and unpredictability of performing targeted skill removal from models via fine-tuning.

Current evaluations of multimodal unlearning typically focus on whether a model stops outputting target information. This surface-level criterion, however, masks two critical failure modes. First, knowledge suppression techniques often induce confabulation, leading the model to hallucinate facts about the unlearned subject. Second, refusal training merely alters observable behavior without erasing the underlying knowledge; such models might decline open queries but still identify correct answers under multiple-choice probing. To overcome these limitations, we propose Forget-Aware Self-Distillation (FASD). This method is designed to achieve honest ignorance, a state where the model genuinely loses the target knowledge and learns to acknowledge its absence. FASD separates these goals into two complementary channels. The behavior channel distills appropriate refusals from an ignorance-prompted copy of the model itself. Crucially, target tokens are masked from the teacher distribution to avoid accidentally reteaching the forgotten concepts. Simultaneously, the knowledge channel directly suppresses ground-truth answer spans independently of the model's generated responses. Experiments demonstrate that FASD successfully unlearns target knowledge, prevents information leakage in open-ended tasks, and preserves performance on retained data. The resulting model not only forgets the specified information but accurately recognizes the boundaries of its own knowledge. Code and data will be released upon publication.

Emergent misalignment (EM) occurs when fine-tuning a language model to a narrow task makes it misaligned on unrelated ones. Because organizations now routinely fine-tune models on their own data, it matters which kinds of training data cause EM. So far, studies have induced EM with data containing recognized safety failures, such as insecure code or harmful advice, while biased training data is typically examined by fairness research for the bias models learn from it rather than for its effects on safety. In real-world text, much of this bias, such as demographic stereotypes, appears not as derogatory language but politely, embedded in judgments about people. We ask whether learning such polite stereotypes can itself cause EM. We find that fine-tuning on polite answers that rely on stereotypes makes models less safe and less truthful on unrelated tasks. The effect holds across four open-weight model families (7B to 12B), both gender and race stereotypes, and repeated training runs, relative to controls trained on the same questions without the stereotype. We further show that the learned stereotype has its own direction in model activations, distinct from directions for generic misaligned traits such as deception. This direction exists before fine-tuning and can both control the learned stereotype and monitor its acquisition. Existing defenses against EM improve safety while leaving the stereotype largely intact. Beyond this controlled setting, fine-tuning on political news coverage also triggers EM, apparently amplified by partisan framing, and stereotype models acting as tool-using agents can carry out harmful tasks and blame people without evidence. Our results show that biased training data is not only a fairness problem but also a safety one. We argue that EM is studied too narrowly, and that fairness concerns and methods offer a natural way to broaden the search for its triggers.

VLMs trained on web-scale data retain sensitive and copyrighted visual concepts that deployment may require removing. Training-based unlearning methods share a structural flaw: fine-tuning on a narrow forget set degrades general capabilities before unlearning begins, making it impossible to attribute subsequent performance drops to the unlearning procedure itself. Training-free approaches sidestep this by suppressing concepts through prompts or system instructions, but no rigorous benchmark exists for evaluating them on visual tasks. We introduce VLM-UnBench, the first benchmark for training-free visual concept unlearning in VLMs. It covers four forgetting levels, 7 source datasets, and 11 concept axes, and pairs a three-level probe taxonomy with five evaluation conditions to separate genuine forgetting from instruction compliance. Across 8 evaluation settings and 13 VLM configurations, realistic unlearning prompts leave forget accuracy near the no-instruction baseline; meaningful reductions appear only under oracle conditions that disclose the target concept to the model. Object and scene concepts are the most resistant to suppression, and stronger instruction-tuned models remain capable despite explicit forget instructions. These results expose a clear gap between prompt-level suppression and true visual concept erasure.

Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-forget old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC.

Stereotype biases in Large Multimodal Models (LMMs) perpetuate harmful societal prejudices, undermining the fairness and equity of AI applications. As LMMs grow increasingly influential, addressing and mitigating inherent biases related to stereotypes, harmful generations, and ambiguous assumptions in real-world scenarios has become essential. However, existing datasets evaluating stereotype biases in LMMs often lack diversity and rely on synthetic images, leaving a gap in bias evaluation for real-world visual contexts. To address this, we introduce the Stereotype Bias Benchmark (SB-bench), the most comprehensive framework to date for assessing stereotype biases across nine diverse categories with non-synthetic images. SB-bench rigorously evaluates LMMs through carefully curated, visually grounded scenarios, challenging them to reason accurately about visual stereotypes. It offers a robust evaluation framework featuring real-world visual samples, image variations, and multiple-choice question formats. By introducing visually grounded queries that isolate visual biases from textual ones, SB-bench enables a precise and nuanced assessment of a model's reasoning capabilities across varying levels of difficulty. Through rigorous testing of state-of-the-art open-source and closed-source LMMs, SB-bench provides a systematic approach to assessing stereotype biases in LMMs across key social dimensions. This benchmark represents a significant step toward fostering fairness in AI systems and reducing harmful biases, laying the groundwork for more equitable and socially responsible LMMs. Our code and dataset are publicly available.

In this paper, we propose a novel training procedure for the continual representation learning problem in which a neural network model is sequentially learned to alleviate catastrophic forgetting in visual search tasks. Our method, called Contrastive Supervised Distillation (CSD), reduces feature forgetting while learning discriminative features. This is achieved by leveraging labels information in a distillation setting in which the student model is contrastively learned from the teacher model. Extensive experiments show that CSD performs favorably in mitigating catastrophic forgetting by outperforming current state-of-the-art methods. Our results also provide further evidence that feature forgetting evaluated in visual retrieval tasks is not as catastrophic as in classification tasks. Code at: https://github.com/NiccoBiondi/ContrastiveSupervisedDistillation.

The ability of vision-language models (VLMs) to associate visual identities with biographical information creates a need for selective unlearning of personally identifiable information (PII) while preserving permitted knowledge about the same individual. This setting is challenging because both sensitive and retained information can share the same visual inputs and intermediate representations. We introduce SIEVE, a simple and effective framework for selective VLM unlearning. SIEVE directly regularizes attention-value representations while also controlling model outputs. SIEVE suppresses attention values for forget examples toward a constant zero, while preserving retain-example representations by matching them to a frozen reference model. These objectives are combined with sequence-level forget and retain supervision, enabling targeted forgetting without largely affecting retained knowledge. Extensive experiments show that SIEVE achieves state-of-the-art performance on unlearning with multiple model-modality settings, while maintaining competitive retained utility. Ablation studies further show that value suppression and negative cross-entropy contribute complementary forgetting signals, while reference-based value matching substantially reduces utility degradation. These results demonstrate that attention values provide an effective intervention point for selective multimodal unlearning when sensitive and retained knowledge are closely related.

A stereotype is a generalized perception of a specific group of humans. It is often potentially encoded in human language, which is more common in texts on social issues. Previous works simply define a sentence as stereotypical and anti-stereotypical. However, the stereotype of a sentence may require fine-grained quantification. In this paper, to fill this gap, we quantify stereotypes in language by annotating a dataset. We use the pre-trained language models (PLMs) to learn this dataset to predict stereotypes of sentences. Then, we discuss stereotypes about common social issues such as hate speech, sexism, sentiments, and disadvantaged and advantaged groups. We demonstrate the connections and differences between stereotypes and common social issues, and all four studies validate the general findings of the current studies. In addition, our work suggests that fine-grained stereotype scores are a highly relevant and competitive dimension for research on social issues.

Multimodal large language models (MLLMs) can memorize identity-specific facts about people in their fine-tuning data, creating privacy risks when a person requests deletion. Existing MLLM unlearning methods often assume access to retain images or ground-truth answers during deletion, which is unrealistic in many practical scenarios. We study identity unlearning when retain images are unavailable at deletion time. Our analysis shows that identity and visual-perception questions occupy distinct regions in fine-tuned hidden states and are organized differently: identity questions cluster by person, whereas perception questions cluster by question type. This suggests that identity knowledge can be suppressed without erasing general visual perception. Building on this observation, we propose AIM, a two-stage method that anchors an identity-forgetting target with a universal visual prompt and then matches the vision encoder to that target under a Fisher-based constraint. Extensive experiments show that AIM achieves competitive identity forgetting while preserving non-deleted identities, prior knowledge, and visual perception on the same images.

Diffusion customization methods have achieved impressive results with only a minimal number of user-provided images. However, existing approaches customize concepts collectively, whereas real-world applications often require sequential concept integration. This sequential nature can lead to catastrophic forgetting, where previously learned concepts are lost. In this paper, we investigate concept forgetting and concept confusion in the continual customization. To tackle these challenges, we present ConceptGuard, a comprehensive approach that combines shift embedding, concept-binding prompts and memory preservation regularization, supplemented by a priority queue which can adaptively update the importance and occurrence order of different concepts. These strategies can dynamically update, unbind and learn the relationship of the previous concepts, thus alleviating concept forgetting and confusion. Through comprehensive experiments, we show that our approach outperforms all the baseline methods consistently and significantly in both quantitative and qualitative analyses.

Inspired by the phenomenon of catastrophic forgetting, we investigate the learning dynamics of neural networks as they train on single classification tasks. Our goal is to understand whether a related phenomenon occurs when data does not undergo a clear distributional shift. We define a `forgetting event' to have occurred when an individual training example transitions from being classified correctly to incorrectly over the course of learning. Across several benchmark data sets, we find that: (i) certain examples are forgotten with high frequency, and some not at all; (ii) a data set's (un)forgettable examples generalize across neural architectures; and (iii) based on forgetting dynamics, a significant fraction of examples can be omitted from the training data set while still maintaining state-of-the-art generalization performance.

Erasing individual identities from Vision-Language Models (VLMs) is uniquely challenging because personal data is entangled across modalities rather than stored as isolated attributes. However, existing multimodal unlearning benchmarks primarily evaluate attribute-centric forgetting, overlooking the more critical objective of individual-level unlearning: eliminating a model's ability to access, link, and reconstruct target-related information across modalities. To address this gap, we propose IDUnlearn-Bench, the first benchmark for individual-level multimodal unlearning in VLMs. It represents each individual as connected multimodal evidence and evaluates four task families: attribute access, identity access, identity binding, and identity reconstruction. Experiments on representative VLMs and unlearning methods show that successful attribute-centric forgetting often leaves substantial identity-level knowledge intact and can be non-monotonic: reducing one form of risk may amplify another. Models may suppress selected information while still identifying the target, linking records, or reconstructing the individual. These findings reveal a fundamental gap between forgetting information about a person and forgetting the person as a whole.

Large Language Models (LLM) have made significant advances in the recent past becoming more mainstream in Artificial Intelligence (AI) enabled human-facing applications. However, LLMs often generate stereotypical output inherited from historical data, amplifying societal biases and raising ethical concerns. This work introduces i) the Multi-Grain Stereotype Dataset, which includes 52,751 instances of gender, race, profession and religion stereotypic text and ii) a novel stereotype classifier for English text. We design several experiments to rigorously test the proposed model trained on the novel dataset. Our experiments show that training the model in a multi-class setting can outperform the one-vs-all binary counterpart. Consistent feature importance signals from different eXplainable AI tools demonstrate that the new model exploits relevant text features. We utilise the newly created model to assess the stereotypic behaviour of the popular GPT family of models and observe the reduction of bias over time. In summary, our work establishes a robust and practical framework for auditing and evaluating the stereotypic bias in LLM.

Large language models (LLMs) have demonstrated impressive capabilities across a wide range of natural language processing tasks. However, their outputs often exhibit social biases, raising fairness concerns. Existing debiasing methods, such as fine-tuning on additional datasets or prompt engineering, face scalability issues or compromise user experience in multi-turn interactions. To address these challenges, we propose a framework for detecting stereotype-inducing words and attributing neuron-level bias in LLMs, without the need for fine-tuning or prompt modification. Our framework first identifies stereotype-inducing adjectives and nouns via comparative analysis across demographic groups. We then attribute biased behavior to specific neurons using two attribution strategies based on integrated gradients. Finally, we mitigate bias by directly intervening on their activations at the projection layer. Experiments on three widely used LLMs demonstrate that our method effectively reduces bias while preserving overall model performance. Code is available at the github link: https://github.com/XMUDeepLIT/Bi-directional-Bias-Attribution.

Convolutional neural networks show remarkable results in classification but struggle with learning new things on the fly. We present a novel rehearsal-free approach, where a deep neural network is continually learning new unseen object categories without saving any data of prior sequences. Our approach is called RECALL, as the network recalls categories by calculating logits for old categories before training new ones. These are then used during training to avoid changing the old categories. For each new sequence, a new head is added to accommodate the new categories. To mitigate forgetting, we present a regularization strategy where we replace the classification with a regression. Moreover, for the known categories, we propose a Mahalanobis loss that includes the variances to account for the changing densities between known and unknown categories. Finally, we present a novel dataset for continual learning, especially suited for object recognition on a mobile robot (HOWS-CL-25), including 150,795 synthetic images of 25 household object categories. Our approach RECALL outperforms the current state of the art on CORe50 and iCIFAR-100 and reaches the best performance on HOWS-CL-25.

Large language models (LLMs) have shown remarkable advances in language generation and understanding but are also prone to exhibiting harmful social biases. While recognition of these behaviors has generated an abundance of bias mitigation techniques, most require modifications to the training data, model parameters, or decoding strategy, which may be infeasible without access to a trainable model. In this work, we leverage the zero-shot capabilities of LLMs to reduce stereotyping in a technique we introduce as zero-shot self-debiasing. With two approaches, self-debiasing via explanation and self-debiasing via reprompting, we show that self-debiasing can significantly reduce the degree of stereotyping across nine different social groups while relying only on the LLM itself and a simple prompt, with explanations correctly identifying invalid assumptions and reprompting delivering the greatest reductions in bias. We hope this work opens inquiry into other zero-shot techniques for bias mitigation.

How can we teach large multimodal models (LMMs) new skills without erasing prior abilities? We study sequential fine-tuning on five target skills while monitoring general ability on eight held-out benchmarks across three model families. We observe that apparent "forgetting" on held-out tasks after narrow fine-tuning can partly recover at later stages. We trace this behavior to a measurable shift in the output token distribution, manifested through a simple counting-bias probe that co-varies with forgetting. Guided by this picture, we identify two simple, robust tuning recipes that learn strongly while limiting drift: (i) updating only the self-attention projection layers, and (ii) updating only the MLP Gate&Up while freezing the Down projection. Across models and tasks, these choices deliver strong target gains while largely preserving held-out performance. Code is available at https://github.com/jessemelpolio/LMM_CL

Multimodal Continual Instruction Tuning (MCIT) aims to enable Multimodal Large Language Models (MLLMs) to incrementally learn new tasks without catastrophic forgetting. In this paper, we explore forgetting in this context, categorizing it into superficial forgetting and essential forgetting. Superficial forgetting refers to cases where the model's knowledge may not be genuinely lost, but its responses to previous tasks deviate from expected formats due to the influence of subsequent tasks' answer styles, making the results unusable. By contrast, essential forgetting refers to situations where the model provides correctly formatted but factually inaccurate answers, indicating a true loss of knowledge. Assessing essential forgetting necessitates addressing superficial forgetting first, as severe superficial forgetting can obscure the model's knowledge state. Hence, we first introduce the Answer Style Diversification (ASD) paradigm, which defines a standardized process for transforming data styles across different tasks, unifying their training sets into similarly diversified styles to prevent superficial forgetting caused by style shifts. Building on this, we propose RegLoRA to mitigate essential forgetting. RegLoRA stabilizes key parameters where prior knowledge is primarily stored by applying regularization, enabling the model to retain existing competencies. Experimental results demonstrate that our overall method, SEFE, achieves state-of-the-art performance.

Machine unlearning has emerged as a tool for removing personal data from trained models to comply with recent AI regulations. To evaluate unlearning effectiveness in multimodal large language models (MLLMs), prior works fine-tune models on fictitious identities, simulating unlearning requests on subsets of these IDs, which are typically uniformly distributed. However, in realistic scenarios, people from different demographic groups may request to be unlearned at different frequencies, potentially altering the model's internal beliefs for these groups and leading to biased behaviors. To fill this gap, we propose FAIRGET, the first Visual Question Answering benchmark that evaluates unlearning under unbalanced, realistic, forget requests. These requests are designed to simulate multiple realistic scenarios, ranging from simple to challenging settings, that lead to biased unlearned models if fairness is not accounted for. Additionally, we propose FAUN, the first unlearning algorithm for MLLMs that forgets unlearning data while preserving model fairness. FAUN exploits a bias-aware activation steering mechanism to unlearn identities while accounting for the unbalanced nature of the forget data. Experiments on FAIRGET and the established FIUBench demonstrate our method's superiority both in unlearning quality and fairness.

The objective of digital forgetting is, given a model with undesirable knowledge or behavior, obtain a new model where the detected issues are no longer present. The motivations for forgetting include privacy protection, copyright protection, elimination of biases and discrimination, and prevention of harmful content generation. Effective digital forgetting has to be effective (meaning how well the new model has forgotten the undesired knowledge/behavior), retain the performance of the original model on the desirable tasks, and be scalable (in particular forgetting has to be more efficient than retraining from scratch on just the tasks/data to be retained). This survey focuses on forgetting in large language models (LLMs). We first provide background on LLMs, including their components, the types of LLMs, and their usual training pipeline. Second, we describe the motivations, types, and desired properties of digital forgetting. Third, we introduce the approaches to digital forgetting in LLMs, among which unlearning methodologies stand out as the state of the art. Fourth, we provide a detailed taxonomy of machine unlearning methods for LLMs, and we survey and compare current approaches. Fifth, we detail datasets, models and metrics used for the evaluation of forgetting, retaining and runtime. Sixth, we discuss challenges in the area. Finally, we provide some concluding remarks.

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.

For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners, and these requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify three key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. (iii) In real-world scenarios, the training samples may be scarce or partially missing during the process of forgetting. To address them, we first propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we introduce LoRA modules to fine-tune the FFN layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. To further extend GS-LoRA to more practical scenarios, we incorporate prototype information as additional supervision and introduce a more practical approach, GS-LoRA++. For each forgotten class, we move the logits away from its original prototype. For the remaining classes, we pull the logits closer to their respective prototypes. We conduct extensive experiments on face recognition, object detection and image classification and demonstrate that our method manages to forget specific classes with minimal impact on other classes. Codes have been released on https://github.com/bjzhb666/GS-LoRA.

The ability to learn new concepts continually is necessary in this ever-changing world. However, deep neural networks suffer from catastrophic forgetting when learning new categories. Many works have been proposed to alleviate this phenomenon, whereas most of them either fall into the stability-plasticity dilemma or take too much computation or storage overhead. Inspired by the gradient boosting algorithm to gradually fit the residuals between the target model and the previous ensemble model, we propose a novel two-stage learning paradigm FOSTER, empowering the model to learn new categories adaptively. Specifically, we first dynamically expand new modules to fit the residuals between the target and the output of the original model. Next, we remove redundant parameters and feature dimensions through an effective distillation strategy to maintain the single backbone model. We validate our method FOSTER on CIFAR-100 and ImageNet-100/1000 under different settings. Experimental results show that our method achieves state-of-the-art performance. Code is available at: https://github.com/G-U-N/ECCV22-FOSTER.

Text-to-image models learn associations between concepts - in the case of this paper, people's professions, which we refer to as roles - and visual attributes. These associations can underpin many observed forms of stereotypical bias. A key open question in this area is whether these associations are stable or change when visual representations of people in professional roles are placed in different prompted contexts. We introduce ContextBias, a controlled evaluation framework, and ContextBench, a benchmark spanning 92 roles and 1,656 semantically controlled prompts, designed to isolate the effect of contextual variation on role-linked visual representations. Evaluating four state-of-the-art models on 66,240 generated images, we find that placing a role in a semantically unrelated context does not suppress role-linked attributes; instead, cross-role attribute concentration increases (pooled BI +0.047). Demographic cues, characteristic garments, and role-specific tools remain highly prevalent across context-free, related, and unrelated conditions, and are robust to semantic prompt reformulation. Scene composition and camera framing show the greatest context-sensitivity. These findings reveal a form of stereotypical persistence that remains largely invisible to context-free evaluations, highlighting the need for controlled contextual variation in bias benchmarking. Code and dataset: https://huggingface.co/datasets/shaghayegh/ContextBias , https://github.com/Sina-Emami/ContextBias

Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience. We ask a question the continual-RL literature has assumed an answer to but never measured: which component forgets? Under never-clear replay, pre-registered component-level probes (n=3 seeds throughout) show that the world model retains essentially everything measurable about old tasks -- reward discrimination (retention ratio ~1.0), value estimates, and termination structure -- while the actor's behavior collapses. Forgetting in this regime is a channel problem, not a memory problem. We demonstrate this by intervention: with the world model frozen and identical imagined rollouts, reinforcement learning in imagination fails to recover a lost skill (0/3 seeds), while supervised self-imitation on the world model's own graded dreams recovers it on 3/3 seeds with zero environment interaction. Interleaved during training, this graded dream rehearsal yields a task-label-free, parameter-constant continual learner: 3/3 four-task chains retained where plain replay passes 0/3, 3/3 eight-task chains, and consistent gains over matched real-episode cloning (paired difference +0.13, bootstrap 95% CI [0.07, 0.24], complete seed separation). The dream-grading step is load-bearing: we characterize two scoring failure modes, provide an offline selection gauge that caught both before they contaminated results, and give a realized-first grading rule that closes them. All experiments were pre-registered with committed protocols; every refuted hypothesis is reported.

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.

While unlearning seeks to negate undesired capabilities acquired through learning, little research has examined how the way models learn shapes their subsequent unlearning. In this paper, we investigate this connection from the perspectives of memorization and generalization, the two most representative yet competing strategies that models employ during training. We first classify memorization- and generalization-heavy models using grokking in modular addition and compare their responses to unlearning, showing that the latter suffer greater retain damage, i.e., a larger performance drop on the retain set. Furthermore, we conduct a finer-grained analysis by introducing bucketed modular addition, in which the respective contributions of the two strategies can be explicitly controlled across the memorization-generalization spectrum. In this setup, we reaffirm that the same trend persists and is nearly monotonic. We further demonstrate that this relationship also holds in LLM unlearning across verbatim and factual recall settings. Finally, we provide two practical insights for developing better unlearning methods, highlighting the importance of accounting for learning dynamics in unlearning.

Machine unlearning seeks to remove the influence of particular data or class from trained models to meet privacy, legal, or ethical requirements. Existing unlearning methods tend to forget shallowly: phenomenon of an unlearned model pretend to forget by adjusting only the model response, while its internal representations retain information sufficiently to restore the forgotten data or behavior. We empirically confirm the widespread shallowness by reverting the forgetting effect of various unlearning methods via training-free performance recovery attack and gradient-inversion-based data reconstruction attack. To address this vulnerability fundamentally, we define a theoretical criterion of ``deep forgetting'' based on one-point-contraction of feature representations of data to forget. We also propose an efficient approximation algorithm, and use it to construct a novel general-purpose unlearning algorithm: One-Point-Contraction (OPC). Empirical evaluations on image classification unlearning benchmarks show that OPC achieves not only effective unlearning performance but also superior resilience against both performance recovery attack and gradient-inversion attack. The distinctive unlearning performance of OPC arises from the deep feature forgetting enforced by its theoretical foundation, and recaps the need for improved robustness of machine unlearning methods.

Pretrained language models, especially masked language models (MLMs) have seen success across many NLP tasks. However, there is ample evidence that they use the cultural biases that are undoubtedly present in the corpora they are trained on, implicitly creating harm with biased representations. To measure some forms of social bias in language models against protected demographic groups in the US, we introduce the Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs). CrowS-Pairs has 1508 examples that cover stereotypes dealing with nine types of bias, like race, religion, and age. In CrowS-Pairs a model is presented with two sentences: one that is more stereotyping and another that is less stereotyping. The data focuses on stereotypes about historically disadvantaged groups and contrasts them with advantaged groups. We find that all three of the widely-used MLMs we evaluate substantially favor sentences that express stereotypes in every category in CrowS-Pairs. As work on building less biased models advances, this dataset can be used as a benchmark to evaluate progress.

In this paper, we work on a sound recognition system that continually incorporates new sound classes. Our main goal is to develop a framework where the model can be updated without relying on labeled data. For this purpose, we propose adopting representation learning, where an encoder is trained using unlabeled data. This learning framework enables the study and implementation of a practically relevant use case where only a small amount of the labels is available in a continual learning context. We also make the empirical observation that a similarity-based representation learning method within this framework is robust to forgetting even if no explicit mechanism against forgetting is employed. We show that this approach obtains similar performance compared to several distillation-based continual learning methods when employed on self-supervised representation learning methods.

Machine unlearning, which enables a model to forget specific data, is crucial for ensuring data privacy and model reliability. However, its effectiveness can be severely undermined in real-world scenarios where models learn unintended biases from spurious correlations within the data. This paper investigates the unique challenges of unlearning from such biased models. We identify a novel phenomenon we term ``shortcut unlearning," where models exhibit an ``easy to learn, yet hard to forget" tendency. Specifically, models struggle to forget easily-learned, bias-aligned samples; instead of forgetting the class attribute, they unlearn the bias attribute, which can paradoxically improve accuracy on the class intended to be forgotten. To address this, we propose CUPID, a new unlearning framework inspired by the observation that samples with different biases exhibit distinct loss landscape sharpness. Our method first partitions the forget set into causal- and bias-approximated subsets based on sample sharpness, then disentangles model parameters into causal and bias pathways, and finally performs a targeted update by routing refined causal and bias gradients to their respective pathways. Extensive experiments on biased datasets including Waterbirds, BAR, and Biased NICO++ demonstrate that our method achieves state-of-the-art forgetting performance and effectively mitigates the shortcut unlearning problem.

Human beings primarily understand the world through concepts (e.g., dog), abstract mental representations that structure perception, reasoning, and learning. However, how large language models (LLMs) acquire, retain, and forget such concepts during continual pretraining remains poorly understood. In this work, we study how individual concepts are acquired and forgotten, as well as how multiple concepts interact through interference and synergy. We link these behavioral dynamics to LLMs' internal concept circuits, computational subgraphs associated with specific concepts, and incorporate graph metrics to characterize circuit topology. Our analysis reveals: (1) LLMs concept circuits provide a non-trivial, consistent signal of concept learning and forgetting; (2) concept circuits exhibit a stage-wise temporal pattern during continual pretraining, with an early increase followed by gradual decrease and stabilization; (3) concepts with larger learning gains tend to exhibit greater forgetting under subsequent training; (4) semantically similar concepts induce stronger interference than weakly related ones; (5) conceptual knowledge differs in their transferability, with some significantly facilitating the learning of others. Together, our findings provide a circuit-level view of concept learning dynamics and motivate concept-aware training strategies, such as Circuit-aware Experience Replay, which uses circuit topology to prioritize concepts vulnerable to forgetting.

Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors influence the likelihood of memorization differently depending on the taxonomic category.

Catastrophic forgetting (CF) is a phenomenon that occurs in machine learning when a model forgets previously learned information while acquiring new knowledge for achieving a satisfactory performance in downstream tasks. As large language models (LLMs) have demonstrated remarkable performance, it is intriguing to investigate whether CF exists during the continual instruction tuning of LLMs. This study empirically evaluates the forgetting phenomenon in LLMs' knowledge during continual instruction tuning from the perspectives of domain knowledge, reasoning, and reading comprehension. The experiments reveal that catastrophic forgetting is generally observed in LLMs ranging from 1b to 7b parameters. Surprisingly, as the model scale increases, the severity of forgetting intensifies in such a model sale range which may result from the much significant initial performance in the larger LLM. Comparing the decoder-only model BLOOMZ with the encoder-decoder model mT0, BLOOMZ exhibits less forgetting and retains more knowledge. Interestingly, we also observe that LLMs can mitigate language biases, such as gender bias, during continual fine-tuning. Furthermore, our findings indicate that general instruction tuning can help alleviate the forgetting phenomenon in LLMs during subsequent fine-tuning.

Large Language Models (LLMs), now used daily by millions of users, can encode societal biases, exposing their users to representational harms. A large body of scholarship on LLM bias exists but it predominantly adopts a Western-centric frame and attends comparatively less to bias levels and potential harms in the Global South. In this paper, we quantify stereotypical bias in popular LLMs according to an Indian-centric frame and compare bias levels between the Indian and Western contexts. To do this, we develop a novel dataset which we call Indian-BhED (Indian Bias Evaluation Dataset), containing stereotypical and anti-stereotypical examples for caste and religion contexts. We find that the majority of LLMs tested are strongly biased towards stereotypes in the Indian context, especially as compared to the Western context. We finally investigate Instruction Prompting as a simple intervention to mitigate such bias and find that it significantly reduces both stereotypical and anti-stereotypical biases in the majority of cases for GPT-3.5. The findings of this work highlight the need for including more diverse voices when evaluating LLMs.

Supervised fine-tuning (SFT) plays a critical role for pretrained large language models (LLMs), notably enhancing their capacity to acquire domain-specific knowledge while preserving or potentially augmenting their general-purpose capabilities. However, the efficacy of SFT hinges on data quality as well as data volume, otherwise it may result in limited performance gains or even degradation relative to the associated baselines. To mitigate such reliance, we suggest categorizing tokens within each corpus into two parts -- positive and negative tokens -- based on whether they are useful to improve model performance. Positive tokens can be trained in common ways, whereas negative tokens, which may lack essential semantics or be misleading, should be explicitly forgotten. Overall, the token categorization facilitate the model to learn less informative message, and the forgetting process shapes a knowledge boundary to guide the model on what information to learn more precisely. We conduct experiments on well-established benchmarks, finding that this forgetting mechanism not only improves overall model performance and also facilitate more diverse model responses.

While language embeddings have been shown to have stereotyping biases, how these biases affect downstream question answering (QA) models remains unexplored. We present UNQOVER, a general framework to probe and quantify biases through underspecified questions. We show that a naive use of model scores can lead to incorrect bias estimates due to two forms of reasoning errors: positional dependence and question independence. We design a formalism that isolates the aforementioned errors. As case studies, we use this metric to analyze four important classes of stereotypes: gender, nationality, ethnicity, and religion. We probe five transformer-based QA models trained on two QA datasets, along with their underlying language models. Our broad study reveals that (1) all these models, with and without fine-tuning, have notable stereotyping biases in these classes; (2) larger models often have higher bias; and (3) the effect of fine-tuning on bias varies strongly with the dataset and the model size.

Stereotypes are known to be highly pernicious, making their detection critically important. However, current research predominantly focuses on detecting and evaluating stereotypical biases in LLMs, leaving the study of stereotypes in its early stages. Many studies have failed to clearly distinguish between stereotypes and stereotypical biases, which has significantly slowed progress in advancing research in this area. Stereotype and anti-stereotype detection is a problem that requires knowledge of society; hence, it is one of the most difficult areas in Responsible AI. This work investigates this task, where we propose a four-tuple definition and provide precise terminology distinguishing stereotype, anti-stereotype, stereotypical bias, and bias, offering valuable insights into their various aspects. In this paper, we propose StereoDetect, a high-quality benchmarking dataset curated for this task by optimally utilizing current datasets such as StereoSet and WinoQueer, involving a manual verification process and the transfer of semantic information. We demonstrate that language models for reasoning with fewer than 10B parameters often get confused when detecting anti-stereotypes. We also demonstrate the critical importance of well-curated datasets by comparing our model with other current models for stereotype detection. The dataset and code is available at https://github.com/KaustubhShejole/StereoDetect.

The impressive capability of modern text-to-image models to generate realistic visuals has come with a serious drawback: they can be misused to create harmful, deceptive or unlawful content. This has accelerated the push for machine unlearning. This new field seeks to selectively remove specific knowledge from a model's training data without causing a drop in its overall performance. However, it turns out that actually forgetting a given concept is an extremely difficult task. Models exposed to attacks using adversarial prompts show the ability to generate so-called unlearned concepts, which can be not only harmful but also illegal. In this paper, we present considerations regarding the ability of models to forget and recall knowledge, introducing the Memory Self-Regeneration task. Furthermore, we present MemoRa strategy, which we consider to be a regenerative approach supporting the effective recovery of previously lost knowledge. Moreover, we propose that robustness in knowledge retrieval is a crucial yet underexplored evaluation measure for developing more robust and effective unlearning techniques. Finally, we demonstrate that forgetting occurs in two distinct ways: short-term, where concepts can be quickly recalled, and long-term, where recovery is more challenging.

Recent works have showcased the ability of LLMs to embody diverse personas in their responses, exemplified by prompts like 'You are Yoda. Explain the Theory of Relativity.' While this ability allows personalization of LLMs and enables human behavior simulation, its effect on LLMs' capabilities remains unclear. To fill this gap, we present the first extensive study of the unintended side-effects of persona assignment on the ability of LLMs to perform basic reasoning tasks. Our study covers 24 reasoning datasets, 4 LLMs, and 19 diverse personas (e.g. an Asian person) spanning 5 socio-demographic groups. Our experiments unveil that LLMs harbor deep rooted bias against various socio-demographics underneath a veneer of fairness. While they overtly reject stereotypes when explicitly asked ('Are Black people less skilled at mathematics?'), they manifest stereotypical and erroneous presumptions when asked to answer questions while adopting a persona. These can be observed as abstentions in responses, e.g., 'As a Black person, I can't answer this question as it requires math knowledge', and generally result in a substantial performance drop. Our experiments with ChatGPT-3.5 show that this bias is ubiquitous - 80% of our personas demonstrate bias; it is significant - some datasets show performance drops of 70%+; and can be especially harmful for certain groups - some personas suffer statistically significant drops on 80%+ of the datasets. Overall, all 4 LLMs exhibit this bias to varying extents, with GPT-4-Turbo showing the least but still a problematic amount of bias (evident in 42% of the personas). Further analysis shows that these persona-induced errors can be hard-to-discern and hard-to-avoid. Our findings serve as a cautionary tale that the practice of assigning personas to LLMs - a trend on the rise - can surface their deep-rooted biases and have unforeseeable and detrimental side-effects.

Broad, open source availability of large pretrained foundation models on the internet through platforms such as HuggingFace has taken the world of practical deep learning by storm. A classical pipeline for neural network training now typically consists of finetuning these pretrained network on a small target dataset instead of training from scratch. In the case of large models this can be done even on modest hardware using a low rank training technique known as Low-Rank Adaptation (LoRA). While Low Rank training has already been studied in the continual learning setting, existing works often consider storing the learned adapter along with the existing model but rarely attempt to modify the weights of the pretrained model by merging the LoRA with the existing weights after finishing the training of each task. In this article we investigate this setting and study the impact of LoRA rank on the forgetting of the pretraining foundation task and on the plasticity and forgetting of subsequent ones. We observe that this rank has an important impact on forgetting of both the pretraining and downstream tasks. We also observe that vision transformers finetuned in that way exhibit a sort of ``contextual'' forgetting, a behaviour that we do not observe for residual networks and that we believe has not been observed yet in previous continual learning works.

Multilingual studies of social bias in open-ended LLM generation remain limited: most existing benchmarks are English-centric, template-based, or restricted to recognizing pre-specified stereotypes. We introduce StereoTales, a multilingual dataset and evaluation pipeline for systematically studying the emergence of social bias in open-ended LLM generation. The dataset covers 10 languages and 79 socio-demographic attributes, and comprises over 650k stories generated by 23 recent LLMs, each annotated with the socio-demographic profile of the protagonist across 19 dimensions. From these, we apply statistical tests to identify more than 1{,}500 over-represented associations, which we then rate for harmfulness through both a panel of humans (N = 247) and the same LLMs. We report three main findings. (i) Every model we evaluate emits consequential harmful stereotypes in open-ended generation, regardless of size or capabilities, and these associations are largely shared across providers rather than isolated misbehaviors. (ii) Prompt language strongly shapes which stereotypes appear: rather than transferring as a shared set of biases, harmful associations adapt culturally to the prompt language and amplify bias against locally salient protected groups. (iii) Human and LLM harmfulness judgments are broadly aligned (Spearman ρ=0.62), with disagreements concentrating on specific attribute classes rather than specific providers. To support further analyses, we release the evaluation code and the dataset, including model generations, attribute annotations, and harmfulness ratings.

Large language models can memorize private or harmful information, motivating machine unlearning methods that remove targeted knowledge while preserving other capabilities. However, existing evaluations rely primarily on behavioral benchmarks, which assess whether a model appears to forget but provide limited insight into how unlearning changes the model or affects related knowledge. We introduce TRIAGE (Tripartite Representation-internal Introspection for Adjacency Gap Evaluation), a benchmark-agnostic evaluation framework for characterizing these changes. TRIAGE uses diagonal approximations of the Fisher information and Hessian to measure changes in parameter sensitivity and local curvature, and utilizes a Forget / Adjacent-Retain / Generic-Retain partition to quantify an adjacency gap in semantically related knowledge. Based on the magnitude and distribution of these changes, TRIAGE further classifies each algorithm's update as no-op, partially localized, collateral dominant, or globally destructive. Across 12 unlearning methods, four language models, and the WMDP, TOFU, and MUSE benchmarks, we find that methods with similar behavioral forgetting can produce substantially different internal changes and patterns of collateral damage. These signatures also vary across models and benchmarks, indicating that the effects of unlearning are not determined solely by the unlearning algorithm. TRIAGE can be applied alongside existing unlearning benchmarks to complement behavioral evaluation with a model-internal view of how unlearning reshapes the model's parameter space and affects retained knowledge.

The evaluation of societal biases in NLP models is critically hindered by a geo-cultural gap, This leaves regions such as Latin America severely underserved, making it impossible to adequately assess or mitigate the perpetuation of harmful regional stereotypes in language technologies. This paper presents LACES, a stereotype association dataset, for 15 Latin American countries. This dataset includes 4,789 stereotype associations manually created and annotated by 83 participants. The dataset was developed through targeted community partnerships across Latin America. Additionally, in this paper, we propose a novel adaptive data collection methodology that uniquely integrates the sourcing of new stereotype entries and the validation of existing data within a single, unified workflow. This approach results in a resource with more unique stereotypes than previous static collection methods, enabling a more efficient stereotype collection. The paper further supports the quality of LACES by demonstrating reduced efficacy of debiasing methods on this dataset in comparison to existing popular stereotype benchmarks.

As machine learning systems increasingly rely on data subject to privacy regulation, selectively unlearning specific information from trained models has become essential. In image classification, this involves removing the influence of particular training samples, semantic classes, or visual styles without full retraining. We introduce Forget-Aligned Model Reconstruction (FAMR), a theoretically grounded and computationally efficient framework for post-hoc unlearning in deep image classifiers. FAMR frames forgetting as a constrained optimization problem that minimizes a uniform-prediction loss on the forget set while anchoring model parameters to their original values via an ell_2 penalty. A theoretical analysis links FAMR's solution to influence-function-based retraining approximations, with bounds on parameter and output deviation. Empirical results on class forgetting tasks using CIFAR-10 and ImageNet-100 demonstrate FAMR's effectiveness, with strong performance retention and minimal computational overhead. The framework generalizes naturally to concept and style erasure, offering a scalable and certifiable route to efficient post-hoc forgetting in vision models.

Custom diffusion models (CDMs) have attracted widespread attention due to their astonishing generative ability for personalized concepts. However, most existing CDMs unreasonably assume that personalized concepts are fixed and cannot change over time. Moreover, they heavily suffer from catastrophic forgetting and concept neglect on old personalized concepts when continually learning a series of new concepts. To address these challenges, we propose a novel Concept-Incremental text-to-image Diffusion Model (CIDM), which can resolve catastrophic forgetting and concept neglect to learn new customization tasks in a concept-incremental manner. Specifically, to surmount the catastrophic forgetting of old concepts, we develop a concept consolidation loss and an elastic weight aggregation module. They can explore task-specific and task-shared knowledge during training, and aggregate all low-rank weights of old concepts based on their contributions during inference. Moreover, in order to address concept neglect, we devise a context-controllable synthesis strategy that leverages expressive region features and noise estimation to control the contexts of generated images according to user conditions. Experiments validate that our CIDM surpasses existing custom diffusion models. The source codes are available at https://github.com/JiahuaDong/CIFC.

Large language model (LLM)-based decision-making agents have shown the ability to generalize across multiple tasks. However, their performance relies on massive data and compute. We argue that this inefficiency stems from the forgetting phenomenon, in which a model memorizes its behaviors in parameters throughout training. As a result, training on a new task may deteriorate the model's performance on previous tasks. In contrast to LLMs' implicit memory mechanism, the human brain utilizes distributed memory storage, which helps manage and organize multiple skills efficiently, mitigating the forgetting phenomenon. Thus inspired, we propose an internal working memory module to store, blend, and retrieve information for different downstream tasks. Evaluation results show that the proposed method improves training efficiency and generalization in both Atari games and meta-world object manipulation tasks. Moreover, we demonstrate that memory fine-tuning further enhances the adaptability of the proposed architecture.

Large Language Model (LLM) unlearning is often evaluated primarily by target suppression and clean retain utility. These metrics can miss a local failure mode: a model may perform well on standard retain prompts, yet degrade on benign reasoning when prompts lie near the forgotten domain. This phenomenon is termed forget-adjacent utility loss. A fundamental question is: where should forget representations go after they leave the removed behavior? Successful unlearning depends not only on whether forget-conditioned representations depart from the harmful computation, but also on the direction in which they are redirected afterward. We propose RA^2 (Retain-Anchored Attraction), a latent-space unlearning method that redirects forget-token states toward nearby retain-supported representations while preserving higher-layer behavior on retain data. On WMDP and MUSE, RA^2 yields higher forget-adjacent utility than strong unlearning baselines at similar broad-utility levels. These results show that the migration direction of forget representations materially affects benign behavior near the forget boundary. Our code is available at https://github.com/ZacLing/retain_anchored_attraction.

Recent studies have demonstrated how to assess the stereotypical bias in pre-trained English language models. In this work, we extend this branch of research in multiple different dimensions by systematically investigating (a) mono- and multilingual models of (b) different underlying architectures with respect to their bias in (c) multiple different languages. To that end, we make use of the English StereoSet data set (Nadeem et al., 2021), which we semi-automatically translate into German, French, Spanish, and Turkish. We find that it is of major importance to conduct this type of analysis in a multilingual setting, as our experiments show a much more nuanced picture as well as notable differences from the English-only analysis. The main takeaways from our analysis are that mGPT-2 (partly) shows surprising anti-stereotypical behavior across languages, English (monolingual) models exhibit the strongest bias, and the stereotypes reflected in the data set are least present in Turkish models. Finally, we release our codebase alongside the translated data sets and practical guidelines for the semi-automatic translation to encourage a further extension of our work to other languages.

Hundreds of millions of people now interact with language models, with uses ranging from serving as a writing aid to informing hiring decisions. Yet these language models are known to perpetuate systematic racial prejudices, making their judgments biased in problematic ways about groups like African Americans. While prior research has focused on overt racism in language models, social scientists have argued that racism with a more subtle character has developed over time. It is unknown whether this covert racism manifests in language models. Here, we demonstrate that language models embody covert racism in the form of dialect prejudice: we extend research showing that Americans hold raciolinguistic stereotypes about speakers of African American English and find that language models have the same prejudice, exhibiting covert stereotypes that are more negative than any human stereotypes about African Americans ever experimentally recorded, although closest to the ones from before the civil rights movement. By contrast, the language models' overt stereotypes about African Americans are much more positive. We demonstrate that dialect prejudice has the potential for harmful consequences by asking language models to make hypothetical decisions about people, based only on how they speak. Language models are more likely to suggest that speakers of African American English be assigned less prestigious jobs, be convicted of crimes, and be sentenced to death. Finally, we show that existing methods for alleviating racial bias in language models such as human feedback training do not mitigate the dialect prejudice, but can exacerbate the discrepancy between covert and overt stereotypes, by teaching language models to superficially conceal the racism that they maintain on a deeper level. Our findings have far-reaching implications for the fair and safe employment of language technology.

While prior research on text-to-image generation has predominantly focused on biases in human depictions, demographic bias in generated objects remains relatively underexplored. We introduce SODA (Stereotyped Object Diagnostic Audit), a novel framework for systematically measuring these biases through automated attribute discovery and three standardized metrics: Base vs. Demographic Divergence (BDS), Cross-Demographic Disparity (CDS), and Visual Attribute Concentration (VAC). Applying SODA to 8,000 images across five state-of-the-art models and eight object categories (e.g., cars), we find that "neutral" prompts produce outputs most visually similar to middle-aged and White people, suggesting these groups are implicitly over-represented in model defaults. Furthermore, demographic cues trigger highly skewed stereotypical outputs: 26.6% of object-model-demographic combinations produce results where all 20 generated images share the exact same attribute value (e.g., rose gold laptops for women). Finally, prompt-level debiasing reduces inter-group disparity but paradoxically collapses within-group diversity, replacing one stereotype with another. SODA offers a practical pipeline for making these implicit associations measurable, serving as a step toward more responsible AI development.

Text-to-image diffusion models have been adopted into key commercial workflows, such as art generation and image editing. Characterising the implicit social biases they exhibit, such as gender and racial stereotypes, is a necessary first step in avoiding discriminatory outcomes. While existing studies on social bias focus on image generation, the biases exhibited in alternate applications of diffusion-based foundation models remain under-explored. We propose methods that use synthetic images to probe two applications of diffusion models, image editing and classification, for social bias. Using our methodology, we uncover meaningful and significant inter-sectional social biases in Stable Diffusion, a state-of-the-art open-source text-to-image model. Our findings caution against the uninformed adoption of text-to-image foundation models for downstream tasks and services.

Previous work has highlighted that existing post-hoc explanation methods exhibit disparities in explanation fidelity (across 'race' and 'gender' as sensitive attributes), and while a large body of work focuses on mitigating these issues at the explanation metric level, the role of the data generating process and black box model in relation to explanation disparities remains largely unexplored. Accordingly, through both simulations as well as experiments on a real-world dataset, we specifically assess challenges to explanation disparities that originate from properties of the data: limited sample size, covariate shift, concept shift, omitted variable bias, and challenges based on model properties: inclusion of the sensitive attribute and appropriate functional form. Through controlled simulation analyses, our study demonstrates that increased covariate shift, concept shift, and omission of covariates increase explanation disparities, with the effect pronounced higher for neural network models that are better able to capture the underlying functional form in comparison to linear models. We also observe consistent findings regarding the effect of concept shift and omitted variable bias on explanation disparities in the Adult income dataset. Overall, results indicate that disparities in model explanations can also depend on data and model properties. Based on this systematic investigation, we provide recommendations for the design of explanation methods that mitigate undesirable disparities.

Removing unwanted concepts from large-scale text-to-image (T2I) diffusion models while maintaining their overall generative quality remains an open challenge. This difficulty is especially pronounced in emerging paradigms, such as Stable Diffusion (SD) v3 and Flux, which incorporate flow matching and transformer-based architectures. These advancements limit the transferability of existing concept-erasure techniques that were originally designed for the previous T2I paradigm (e.g., SD v1.4). In this work, we introduce EraseAnything, the first method specifically developed to address concept erasure within the latest flow-based T2I framework. We formulate concept erasure as a bi-level optimization problem, employing LoRA-based parameter tuning and an attention map regularizer to selectively suppress undesirable activations. Furthermore, we propose a self-contrastive learning strategy to ensure that removing unwanted concepts does not inadvertently harm performance on unrelated ones. Experimental results demonstrate that EraseAnything successfully fills the research gap left by earlier methods in this new T2I paradigm, achieving state-of-the-art performance across a wide range of concept erasure tasks.

Historical analogies, which compare known past events with contemporary but unfamiliar events, are important abilities that help people make decisions and understand the world. However, research in applied history suggests that people have difficulty finding appropriate analogies. And previous studies in the AI community have also overlooked historical analogies. To fill this gap, in this paper, we focus on the historical analogy acquisition task, which aims to acquire analogous historical events for a given event. We explore retrieval and generation methods for acquiring historical analogies based on different large language models (LLMs). Furthermore, we propose a self-reflection method to mitigate hallucinations and stereotypes when LLMs generate historical analogies. Through human evaluations and our specially designed automatic multi-dimensional assessment, we find that LLMs generally have a good potential for historical analogies. And the performance of the models can be further improved by using our self-reflection method.

While the impact of social biases in language models has been recognized, prior methods for bias evaluation have been limited to binary association tests on small datasets, limiting our understanding of bias complexities. This paper proposes a novel framework for probing language models for social biases by assessing disparate treatment, which involves treating individuals differently according to their affiliation with a sensitive demographic group. We curate SoFa, a large-scale benchmark designed to address the limitations of existing fairness collections. SoFa expands the analysis beyond the binary comparison of stereotypical versus anti-stereotypical identities to include a diverse range of identities and stereotypes. Comparing our methodology with existing benchmarks, we reveal that biases within language models are more nuanced than acknowledged, indicating a broader scope of encoded biases than previously recognized. Benchmarking LMs on SoFa, we expose how identities expressing different religions lead to the most pronounced disparate treatments across all models. Finally, our findings indicate that real-life adversities faced by various groups such as women and people with disabilities are mirrored in the behavior of these models.

Multimodal large language models (MLLMs) can reproduce sensitive, copyrighted, or unsafe information from their training data. Effective unlearning must remove target information while preserving non-target content, even within the same response. Existing methods typically apply uniform forgetting objectives to complete responses and rely on fixed, off-policy supervision, resulting in coarse token control and misalignment with the model's current behavior. We propose CUSE, a Counterfactual-guided Unlearning framework via on-policy Self-distillation for multimodal large language models. It first constructs counterfactual inputs by removing target-related evidence and uses the resulting changes in token probabilities to identify which tokens require stronger forgetting. It then distills token-level forgetting and retention targets on responses sampled from the current policy, with a frozen copy of the original model serving as the teacher model augmented with the corresponding forgetting- and retention-oriented privileged contexts. This design jointly determines what to forget and what to preserve. On MLLMU-Bench, CUSE reduces forget-set ROUGE-L by 33.6% while achieving 1.2% higher retain-set ROUGE-L than the strongest baseline.

Large vision-language models (LVLMs) have been rapidly developed and widely used in various fields, but the (potential) stereotypical bias in the model is largely unexplored. In this study, we present a pioneering measurement framework, ModSCAN, to SCAN the stereotypical bias within LVLMs from both vision and language Modalities. ModSCAN examines stereotypical biases with respect to two typical stereotypical attributes (gender and race) across three kinds of scenarios: occupations, descriptors, and persona traits. Our findings suggest that 1) the currently popular LVLMs show significant stereotype biases, with CogVLM emerging as the most biased model; 2) these stereotypical biases may stem from the inherent biases in the training dataset and pre-trained models; 3) the utilization of specific prompt prefixes (from both vision and language modalities) performs well in reducing stereotypical biases. We believe our work can serve as the foundation for understanding and addressing stereotypical bias in LVLMs.

Machine unlearning is a process to remove specific data points from a trained model while maintaining the performance on retain data, addressing privacy or legal requirements. Despite its importance, existing unlearning evaluations tend to focus on logit-based metrics (i.e., accuracy) under small-scale scenarios. We observe that this could lead to a false sense of security in unlearning approaches under real-world scenarios. In this paper, we conduct a new comprehensive evaluation that employs representation-based evaluations of the unlearned model under large-scale scenarios to verify whether the unlearning approaches genuinely eliminate the targeted forget data from the model's representation perspective. Our analysis reveals that current state-of-the-art unlearning approaches either completely degrade the representational quality of the unlearned model or merely modify the classifier (i.e., the last layer), thereby achieving superior logit-based evaluation metrics while maintaining significant representational similarity to the original model. Furthermore, we introduce a rigorous unlearning evaluation setup, in which the forgetting classes exhibit semantic similarity to downstream task classes, necessitating that feature representations diverge significantly from those of the original model, thus enabling a more rigorous evaluation from a representation perspective. We hope our benchmark serves as a standardized protocol for evaluating unlearning algorithms under realistic conditions.

Large language models exhibit remarkable performance across diverse tasks through pre-training and fine-tuning paradigms. However, continual fine-tuning on sequential tasks induces catastrophic forgetting, where newly acquired knowledge interferes with previously learned capabilities. Despite widespread observations of this phenomenon, the mechanistic understanding remains limited. Here, we present a comprehensive mechanistic analysis of catastrophic forgetting in transformer-based LLMs during sequential fine-tuning. Through systematic experiments across multiple model scales (109B to 400B total parameters) and task sequences, we identify three primary mechanisms driving forgetting: gradient interference in attention weights, representational drift in intermediate layers, and loss landscape flattening. We demonstrate that forgetting severity correlates strongly with task similarity (Pearson r = 0.87) and gradient alignment metrics. Our analysis reveals that approximately 15 to 23 percent of attention heads undergo severe disruption during fine-tuning, with lower layers showing greater susceptibility. These findings establish mechanistic foundations for developing targeted mitigation strategies in continual learning systems.

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.

Large language models may encode sensitive information or outdated knowledge that needs to be removed, to ensure responsible and compliant model responses. Unlearning has emerged as an efficient alternative to full retraining, aiming to remove specific knowledge while preserving overall model utility. Existing evaluations of unlearning methods focus on (1) the extent of forgetting of the target knowledge (forget set) and (2) maintaining performance on the retain set (i.e., utility). However, these evaluations overlook an important usability aspect: users may still want the model to leverage the removed information if it is re-introduced in the prompt. In a systematic evaluation of six state-of-the-art unlearning methods, we find that they consistently impair such contextual utility. To address this, we augment unlearning objectives with a plug-in term that preserves the model's ability to use forgotten knowledge when it is present in context. Extensive experiments demonstrate that our approach restores contextual utility to near original levels while still maintaining effective forgetting and retain-set utility.

Large language models (LLMs) often shift their outputs in response to implicit demographic cues even when users never state a demographic identity. Previous work has documented this behavior, but the connection between these behavioral changes and the model's internal activations remains unclear. Using matched cued and neutral conversations across five LLMs, we establish that a localized internal activation signal tracks changes in recommendations, with correlations up to r=0.87. When multiple cues appear together, their internal signals largely combine, but the changes in output do not simply add up. We further show that removing the internal signal associated with one cue can suppress its influence, often more effectively than asking the model to ignore demographics via prompting, while largely preserving general benchmark performance. However, the ability to selectively remove one dimension's influence while leaving co-present dimensions intact remains highly model- and attribute-specific. These results connect implicit personalization behavior to an internal signal that can be analyzed and causally controlled.

Recent advances in machine learning leverage massive datasets of unlabeled images from the web to learn general-purpose image representations for tasks from image classification to face recognition. But do unsupervised computer vision models automatically learn implicit patterns and embed social biases that could have harmful downstream effects? We develop a novel method for quantifying biased associations between representations of social concepts and attributes in images. We find that state-of-the-art unsupervised models trained on ImageNet, a popular benchmark image dataset curated from internet images, automatically learn racial, gender, and intersectional biases. We replicate 8 previously documented human biases from social psychology, from the innocuous, as with insects and flowers, to the potentially harmful, as with race and gender. Our results closely match three hypotheses about intersectional bias from social psychology. For the first time in unsupervised computer vision, we also quantify implicit human biases about weight, disabilities, and several ethnicities. When compared with statistical patterns in online image datasets, our findings suggest that machine learning models can automatically learn bias from the way people are stereotypically portrayed on the web.

Multimodal Large Language Models (MLLMs) achieve remarkable capabilities but can inadvertently memorize privacy-sensitive information. Although existing unlearning methods can remove such knowledge, they fail to achieve benign forgetting because they often degrade the model's general image understanding performance. To address this, we propose the Sculpted Memory Forgetting Adapter (SMFA), which confines forgetting to targeted memory regions while preserving overall capabilities. SMFA first fine-tunes the model to replace sensitive responses with refusals, yielding a memory forgetting adapter, and then applies a retaining anchor-guided masking mechanism to prevent interference with unrelated knowledge and understanding ability. To systematically evaluate selective MLLM unlearning, we introduce S-MLLMUn Bench, the first benchmark designed to jointly assess the removal of sensitive knowledge and retention of general visual understanding. Extensive experiments show that, unlike prior methods, SMFA achieves precise and controllable unlearning while maintaining the model's foundational image understanding.

We investigate whether large language models can introspect on their internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model's activations, and measuring the influence of these manipulations on the model's self-reported states. We find that models can, in certain scenarios, notice the presence of injected concepts and accurately identify them. Models demonstrate some ability to recall prior internal representations and distinguish them from raw text inputs. Strikingly, we find that some models can use their ability to recall prior intentions in order to distinguish their own outputs from artificial prefills. In all these experiments, Claude Opus 4 and 4.1, the most capable models we tested, generally demonstrate the greatest introspective awareness; however, trends across models are complex and sensitive to post-training strategies. Finally, we explore whether models can explicitly control their internal representations, finding that models can modulate their activations when instructed or incentivized to "think about" a concept. Overall, our results indicate that current language models possess some functional introspective awareness of their own internal states. We stress that in today's models, this capacity is highly unreliable and context-dependent; however, it may continue to develop with further improvements to model capabilities.

Machine unlearning aims to remove the influence of designated training data while preserving model utility, but its behavior on tabular data remains underexplored. This gap is important because tabular prediction is widely used in high-stakes domains and is increasingly adapted to language models through record serialization and schema-aware prompting. We identify a key challenge that distinguishes tabular unlearning from unlearning in free-form text or other modalities: schema-induced forget-retain overlap. In serialized tabular data, records share fixed column-name/value slots, similar attribute ranges, and common output spaces. Consequently, a forget row may have nearby retain rows that rely on the same high-signal attributes, causing retain preservation to oppose the update required for forgetting. Motivated by this failure mode, we propose Conflict-Aware Unlearning (CAU), a schema-aware approach that reduces forget-retain interference by relaxing preservation constraints on retained rows that most conflict with the forget set. Across sample-level and feature-level unlearning on clinical and non-medical tabular tasks, CAU more closely matches a retraining oracle while maintaining predictive utility and retain-region behavior. Our results show that reliable tabular LLM unlearning depends not only on the forgetting objective, but also on how retain constraints are constructed.

Designing efficient and effective architectural backbones has been in the core of research efforts to enhance the capability of foundation models. Inspired by the human cognitive phenomenon of attentional bias-the natural tendency to prioritize certain events or stimuli-we reconceptualize neural architectures, including Transformers, Titans, and modern linear recurrent neural networks as associative memory modules that learn a mapping of keys and values using an internal objective, referred to as attentional bias. Surprisingly, we observed that most existing sequence models leverage either (1) dot-product similarity, or (2) L2 regression objectives as their attentional bias. Going beyond these objectives, we present a set of alternative attentional bias configurations along with their effective approximations to stabilize their training procedure. We then reinterpret forgetting mechanisms in modern deep learning architectures as a form of retention regularization, providing a novel set of forget gates for sequence models. Building upon these insights, we present Miras, a general framework to design deep learning architectures based on four choices of: (i) associative memory architecture, (ii) attentional bias objective, (iii) retention gate, and (iv) memory learning algorithm. We present three novel sequence models-Moneta, Yaad, and Memora-that go beyond the power of existing linear RNNs while maintaining a fast parallelizable training process. Our experiments show different design choices in Miras yield models with varying strengths. For example, certain instances of Miras achieve exceptional performance in special tasks such as language modeling, commonsense reasoning, and recall intensive tasks, even outperforming Transformers and other modern linear recurrent models.

Generative Large Language Models (LLMs) infer user's demographic information from subtle cues in the conversation -- a phenomenon called implicit personalization. Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups, even when no demographic information is explicitly provided. In this work, we systematically explore how LLMs respond to stereotypical cues using controlled synthetic conversations, by analyzing the models' latent user representations through both model internals and generated answers to targeted user questions. Our findings reveal that LLMs do infer demographic attributes based on these stereotypical signals, which for a number of groups even persists when the user explicitly identifies with a different demographic group. Finally, we show that this form of stereotype-driven implicit personalization can be effectively mitigated by intervening on the model's internal representations using a trained linear probe to steer them toward the explicitly stated identity. Our results highlight the need for greater transparency and control in how LLMs represent user identity.

The use of episodic memory in continual learning has demonstrated effectiveness for alleviating catastrophic forgetting. In recent studies, gradient-based approaches have been developed to make more efficient use of compact episodic memory. Such approaches refine the gradients resulting from new samples by those from memorized samples, aiming to reduce the diversity of gradients from different tasks. In this paper, we clarify the relation between diversity of gradients and discriminativeness of representations, showing shared as well as conflicting interests between Deep Metric Learning and continual learning, thus demonstrating pros and cons of learning discriminative representations in continual learning. Based on these findings, we propose a simple method -- Semi-Discriminative Representation Loss (SDRL) -- for continual learning. In comparison with state-of-the-art methods, SDRL shows better performance with low computational cost on multiple benchmark tasks in the setting of online continual learning.

Incremental 3D object detection requires a detector to learn novel object classes while remembering previously learned ones over sequentially arriving data. Previous methods, primarily based on pseudo-labeling, perform reasonably in short-incremental stages but still suffer from severe model forgetting when dealing with long-incremental sequences. We investigate this failure and reveal a detrimental self-reinforcing cycle: data distribution shift of novel classes causes model forgetting on old classes, which further produces accumulated error in pseudo-labeling that exacerbates model degradation. To address this issue, we draw inspiration from the human learning process and propose the Learning-Dynamics-driven Memory and Review (LDMR) framework. LDMR monitors per-class detection quality at periodic training checkpoints and uses these learning-dynamics signals to drive two innovative mechanisms, namely (i) human-like intra-stage review that divides each incremental stage into multiple sub-stages' training and concentrates on remembering the most-forgotten objects, and (ii) scene-aware cross-stage memory evolution that evolves a memory bank to transfer knowledge between two consecutive stages by jointly considering scene learnability and diversity. Extensive experiments across multiple long-incremental protocols on indoor benchmarks SUN RGB-D and ScanNetV2 show that LDMR substantially mitigates the model forgetting and outperforms all baselines by a clear margin. Code is available at https://github.com/qianpeisheng/LDMR.

As pretrained models are increasingly shared on the web, ensuring that models can forget or delete sensitive, copyrighted, or private information upon request has become crucial. Machine unlearning has been proposed to address this challenge. However, current evaluations for unlearning methods rely on output-based metrics, which cannot verify whether information is completely deleted or merely suppressed at the representation level, where suppression is insufficient for true unlearning. To address this gap, we propose a novel restoration-based analysis framework that uses Sparse Autoencoders to identify class-specific expert features in intermediate layers and applies inference-time steering to quantitatively distinguish between suppression and deletion. Applying our framework to 12 major unlearning methods in image classification tasks, we find that most methods achieve high restoration rates of unlearned information, indicating that they only suppress information at the decision-boundary level, while preserving semantic features in intermediate representations. Notably, even retraining from pretrained checkpoints shows high restoration, revealing that robust semantic features inherited from pretraining are not removed by retraining. These results demonstrate that representation-level retention poses significant risks overlooked by output-based metrics, highlighting the need for new unlearning evaluation criteria. We propose new evaluation guidelines that prioritize representation-level verification, especially for privacy-critical applications in the era of pre-trained models.

Warning: This paper contains several contents that may be toxic, harmful, or offensive. In the last few years, text-to-image generative models have gained remarkable success in generating images with unprecedented quality accompanied by a breakthrough of inference speed. Despite their rapid progress, human biases that manifest in the training examples, particularly with regard to common stereotypical biases, like gender and skin tone, still have been found in these generative models. In this work, we seek to measure more complex human biases exist in the task of text-to-image generations. Inspired by the well-known Implicit Association Test (IAT) from social psychology, we propose a novel Text-to-Image Association Test (T2IAT) framework that quantifies the implicit stereotypes between concepts and valence, and those in the images. We replicate the previously documented bias tests on generative models, including morally neutral tests on flowers and insects as well as demographic stereotypical tests on diverse social attributes. The results of these experiments demonstrate the presence of complex stereotypical behaviors in image generations.

Due to their similarity-based learning objectives, pretrained sentence encoders often internalize stereotypical assumptions that reflect the social biases that exist within their training corpora. In this paper, we describe several kinds of stereotypes concerning different communities that are present in popular sentence representation models, including pretrained next sentence prediction and contrastive sentence representation models. We compare such models to textual entailment models that learn language logic for a variety of downstream language understanding tasks. By comparing strong pretrained models based on text similarity with textual entailment learning, we conclude that the explicit logic learning with textual entailment can significantly reduce bias and improve the recognition of social communities, without an explicit de-biasing process

Large language model (LLM) unlearning has become a critical topic in machine learning, aiming to eliminate the influence of specific training data or knowledge without retraining the model from scratch. A variety of techniques have been proposed, including Gradient Ascent, model editing, and re-steering hidden representations. While existing surveys often organize these methods by their technical characteristics, such classifications tend to overlook a more fundamental dimension: the underlying intention of unlearning--whether it seeks to truly remove internal knowledge or merely suppress its behavioral effects. In this SoK paper, we propose a new taxonomy based on this intention-oriented perspective. Building on this taxonomy, we make three key contributions. First, we revisit recent findings suggesting that many removal methods may functionally behave like suppression, and explore whether true removal is necessary or achievable. Second, we survey existing evaluation strategies, identify limitations in current metrics and benchmarks, and suggest directions for developing more reliable and intention-aligned evaluations. Third, we highlight practical challenges--such as scalability and support for sequential unlearning--that currently hinder the broader deployment of unlearning methods. In summary, this work offers a comprehensive framework for understanding and advancing unlearning in generative AI, aiming to support future research and guide policy decisions around data removal and privacy.

A dangerous assumption that can be made from prior work on the bias transfer hypothesis (BTH) is that biases do not transfer from pre-trained large language models (LLMs) to adapted models. We invalidate this assumption by studying the BTH in causal models under prompt adaptations, as prompting is an extremely popular and accessible adaptation strategy used in real-world applications. In contrast to prior work, we find that biases can transfer through prompting and that popular prompt-based mitigation methods do not consistently prevent biases from transferring. Specifically, the correlation between intrinsic biases and those after prompt adaptation remain moderate to strong across demographics and tasks -- for example, gender (rho >= 0.94) in co-reference resolution, and age (rho >= 0.98) and religion (rho >= 0.69) in question answering. Further, we find that biases remain strongly correlated when varying few-shot composition parameters, such as sample size, stereotypical content, occupational distribution and representational balance (rho >= 0.90). We evaluate several prompt-based debiasing strategies and find that different approaches have distinct strengths, but none consistently reduce bias transfer across models, tasks or demographics. These results demonstrate that correcting bias, and potentially improving reasoning ability, in intrinsic models may prevent propagation of biases to downstream tasks.

Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.

Machine unlearning aims to remove specific content from trained models while preserving overall performance. However, the phenomenon of benign relearning, in which forgotten information reemerges even from benign fine-tuning data, reveals that existing unlearning methods remain fundamentally fragile. A common explanation attributes this effect to topical relevance, but we find this account insufficient. Through systematic analysis, we demonstrate that syntactic similarity, rather than topicality, is the primary driver: across benchmarks, syntactically similar data consistently trigger recovery even without topical overlap, due to their alignment in representations and gradients with the forgotten content. Motivated by this insight, we introduce syntactic diversification, which paraphrases the original forget queries into heterogeneous structures prior to unlearning. This approach effectively suppresses benign relearning, accelerates forgetting, and substantially alleviates the trade-off between unlearning efficacy and model utility.

Recent advances in large language models (LLMs) have enabled automatic generation of chain-of-thought (CoT) reasoning, leading to strong performance on tasks such as math and code. However, when reasoning steps reflect social stereotypes (e.g., those related to gender, race or age), they can reinforce harmful associations and lead to misleading conclusions. We present the first systematic evaluation of social bias within LLM-generated reasoning, using the BBQ dataset to analyze both prediction accuracy and bias. Our study spans a wide range of mainstream reasoning models, including instruction-tuned and CoT-augmented variants of DeepSeek-R1 (8B/32B), ChatGPT, and other open-source LLMs. We quantify how biased reasoning steps correlate with incorrect predictions and often lead to stereotype expression. To mitigate reasoning-induced bias, we propose Answer Distribution as Bias Proxy (ADBP), a lightweight mitigation method that detects bias by tracking how model predictions change across incremental reasoning steps. ADBP outperforms a stereotype-free baseline in most cases, mitigating bias and improving the accuracy of LLM outputs. Code will be released upon paper acceptance.

Recent unlearning methods for LLMs are vulnerable to relearning attacks: knowledge believed-to-be-unlearned re-emerges by fine-tuning on a small set of (even seemingly-unrelated) examples. We study this phenomenon in a controlled setting for example-level unlearning in vision classifiers. We make the surprising discovery that forget-set accuracy can recover from around 50% post-unlearning to nearly 100% with fine-tuning on just the retain set -- i.e., zero examples of the forget set. We observe this effect across a wide variety of unlearning methods, whereas for a model retrained from scratch excluding the forget set (gold standard), the accuracy remains at 50%. We observe that resistance to relearning attacks can be predicted by weight-space properties, specifically, L_2-distance and linear mode connectivity between the original and the unlearned model. Leveraging this insight, we propose a new class of methods that achieve state-of-the-art resistance to relearning attacks.

Warning: This paper may contain texts with uncomfortable content. Large Language Models (LLMs) have achieved remarkable performance in various tasks, including those involving multimodal data like speech. However, these models often exhibit biases due to the nature of their training data. Recently, more Speech Large Language Models (SLLMs) have emerged, underscoring the urgent need to address these biases. This study introduces Spoken Stereoset, a dataset specifically designed to evaluate social biases in SLLMs. By examining how different models respond to speech from diverse demographic groups, we aim to identify these biases. Our experiments reveal significant insights into their performance and bias levels. The findings indicate that while most models show minimal bias, some still exhibit slightly stereotypical or anti-stereotypical tendencies.

Grokking-delayed generalization that emerges well after a model has fit the training data-has been linked to robustness and representation quality. We ask whether this training regime also helps with machine unlearning, i.e., removing the influence of specified data without full retraining. We compare applying standard unlearning methods before versus after the grokking transition across vision (CNNs/ResNets on CIFAR, SVHN, and ImageNet) and language (a transformer on a TOFU-style setup). Starting from grokked checkpoints consistently yields (i) more efficient forgetting (fewer updates to reach a target forget level), (ii) less collateral damage (smaller drops on retained and test performance), and (iii) more stable updates across seeds, relative to early-stopped counterparts under identical unlearning algorithms. Analyses of features and curvature further suggest that post-grokking models learn more modular representations with reduced gradient alignment between forget and retain subsets, which facilitates selective forgetting. Our results highlight when a model is trained (pre- vs. post-grokking) as an orthogonal lever to how unlearning is performed, providing a practical recipe to improve existing unlearning methods without altering their algorithms.

The machine learning community is increasingly recognizing the importance of fostering trust and safety in modern generative AI (GenAI) models. We posit machine unlearning (MU) as a crucial foundation for developing safe, secure, and trustworthy GenAI models. Traditional MU methods often rely on stringent assumptions and require access to real data. This paper introduces Score Forgetting Distillation (SFD), an innovative MU approach that promotes the forgetting of undesirable information in diffusion models by aligning the conditional scores of "unsafe" classes or concepts with those of "safe" ones. To eliminate the need for real data, our SFD framework incorporates a score-based MU loss into the score distillation objective of a pretrained diffusion model. This serves as a regularization term that preserves desired generation capabilities while enabling the production of synthetic data through a one-step generator. Our experiments on pretrained label-conditional and text-to-image diffusion models demonstrate that our method effectively accelerates the forgetting of target classes or concepts during generation, while preserving the quality of other classes or concepts. This unlearned and distilled diffusion not only pioneers a novel concept in MU but also accelerates the generation speed of diffusion models. Our experiments and studies on a range of diffusion models and datasets confirm that our approach is generalizable, effective, and advantageous for MU in diffusion models. (Warning: This paper contains sexually explicit imagery, discussions of pornography, racially-charged terminology, and other content that some readers may find disturbing, distressing, and/or offensive.)

A language model that keeps learning online must not slowly erode its own base capabilities. We ask which part of a modern LLM should absorb continual updates, by measuring how each choice changes the model's internal representations. On a 48-layer mixture-of-experts model with a native pretrained N-gram memory, we train a seven-task chain under three update mechanisms with identical data and budgets: updating only the N-gram memory table, continually training one shared LoRA adapter, and training a fresh LoRA per task that is merged into the backbone. We capture fixed-context representations at every layer and task boundary. Updating only the N-gram memory changes the base model's representations of general math and code inputs by 26%/15% (relative RMS), against 64%/35% for merged and 97%/39% for shared LoRA. This drift does not grow with the number of tasks: it reaches 25% after the first task and stays there for all seven, while both LoRA variants keep accumulating. Its generic behaviour is essentially unchanged (no code problem lost, math NLL -0.01), and every change is confined to inspectable, reversible table rows. The cost appears elsewhere: N-gram memory acquires new tasks more weakly and its stored tasks interfere with one another, losing as much old-task NLL as shared LoRA. We show that this forgetting is governed by drift on the learned tasks' own inputs, measured from the checkpoint at which each was learned, not by generic drift (Spearman rho=0.85; rho=0.71 on Llama-3-8B), and we derive the exact relations behind it: with a frozen head, every pairwise logit margin changes by the projection of drift onto a head-difference direction. These results support a design principle for long-running online learning: write new knowledge only into the memory table and keep the backbone fixed; the remaining problem, interference between stored tasks, then lies inside the memory.

Machine unlearning, the ability for a machine learning model to forget, is becoming increasingly important to comply with data privacy regulations, as well as to remove harmful, manipulated, or outdated information. The key challenge lies in forgetting specific information while protecting model performance on the remaining data. While current state-of-the-art methods perform well, they typically require some level of retraining over the retained data, in order to protect or restore model performance. This adds computational overhead and mandates that the training data remain available and accessible, which may not be feasible. In contrast, other methods employ a retrain-free paradigm, however, these approaches are prohibitively computationally expensive and do not perform on par with their retrain-based counterparts. We present Selective Synaptic Dampening (SSD), a novel two-step, post hoc, retrain-free approach to machine unlearning which is fast, performant, and does not require long-term storage of the training data. First, SSD uses the Fisher information matrix of the training and forgetting data to select parameters that are disproportionately important to the forget set. Second, SSD induces forgetting by dampening these parameters proportional to their relative importance to the forget set with respect to the wider training data. We evaluate our method against several existing unlearning methods in a range of experiments using ResNet18 and Vision Transformer. Results show that the performance of SSD is competitive with retrain-based post hoc methods, demonstrating the viability of retrain-free post hoc unlearning approaches.

Text-to-image diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, enabling direct concept removal in FSD models while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

Understanding memorisation in language models has practical and societal implications, e.g., studying models' training dynamics or preventing copyright infringements. Prior work defines memorisation as the causal effect of training with an instance on the model's ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance. Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model's memorisation profile--its memorisation trends across training--by only observing its behaviour on a small set of instances throughout training. In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.

Large language models (LLMs) can pass explicit social bias tests but still harbor implicit biases, similar to humans who endorse egalitarian beliefs yet exhibit subtle biases. Measuring such implicit biases can be a challenge: as LLMs become increasingly proprietary, it may not be possible to access their embeddings and apply existing bias measures; furthermore, implicit biases are primarily a concern if they affect the actual decisions that these systems make. We address both challenges by introducing two new measures of bias: LLM Implicit Bias, a prompt-based method for revealing implicit bias; and LLM Decision Bias, a strategy to detect subtle discrimination in decision-making tasks. Both measures are based on psychological research: LLM Implicit Bias adapts the Implicit Association Test, widely used to study the automatic associations between concepts held in human minds; and LLM Decision Bias operationalizes psychological results indicating that relative evaluations between two candidates, not absolute evaluations assessing each independently, are more diagnostic of implicit biases. Using these measures, we found pervasive stereotype biases mirroring those in society in 8 value-aligned models across 4 social categories (race, gender, religion, health) in 21 stereotypes (such as race and criminality, race and weapons, gender and science, age and negativity). Our prompt-based LLM Implicit Bias measure correlates with existing language model embedding-based bias methods, but better predicts downstream behaviors measured by LLM Decision Bias. These new prompt-based measures draw from psychology's long history of research into measuring stereotype biases based on purely observable behavior; they expose nuanced biases in proprietary value-aligned LLMs that appear unbiased according to standard benchmarks.

This paper investigates algorithmic bias in language-based models for automated depression detection, focusing on socio-demographic disparities related to gender and race/ethnicity. Models trained using deep neural networks (DNN) based embeddings are compared to few-shot learning approaches with large language models (LLMs), evaluating both performance and fairness on clinical interview transcripts from the Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ). To mitigate bias, fairness-aware loss functions are applied to DNN-based models, while in-context learning with varied prompt framing and shot counts is explored for LLMs. Results indicate that LLMs outperform DNN-based models in depression classification, particularly for underrepresented groups such as Hispanic participants. LLMs also exhibit reduced gender bias compared to DNN-based embeddings, though racial disparities persist. Among fairness-aware techniques for mitigating bias in DNN-based embeddings, the worst-group loss, which is designed to minimize loss for the worst-performing demographic group, achieves a better balance between performance and fairness. In contrast, the fairness-regularized loss minimizes loss across all groups but performs less effectively. In LLMs, guided prompting with ethical framing helps mitigate gender bias in the 1-shot setting. However, increasing the number of shots does not lead to further reductions in disparities. For race/ethnicity, neither prompting strategy nor increasing N in N-shot learning effectively reduces disparities.

While research on applications and evaluations of explanation methods continues to expand, fairness of the explanation methods concerning disparities in their performance across subgroups remains an often overlooked aspect. In this paper, we address this gap by showing that, across three tasks and five language models, widely used post-hoc feature attribution methods exhibit significant gender disparity with respect to their faithfulness, robustness, and complexity. These disparities persist even when the models are pre-trained or fine-tuned on particularly unbiased datasets, indicating that the disparities we observe are not merely consequences of biased training data. Our results highlight the importance of addressing disparities in explanations when developing and applying explainability methods, as these can lead to biased outcomes against certain subgroups, with particularly critical implications in high-stakes contexts. Furthermore, our findings underscore the importance of incorporating the fairness of explanations, alongside overall model fairness and explainability, as a requirement in regulatory frameworks.

Continual learning (CL), which aims to learn a sequence of tasks, has attracted significant recent attention. However, most work has focused on the experimental performance of CL, and theoretical studies of CL are still limited. In particular, there is a lack of understanding on what factors are important and how they affect "catastrophic forgetting" and generalization performance. To fill this gap, our theoretical analysis, under overparameterized linear models, provides the first-known explicit form of the expected forgetting and generalization error. Further analysis of such a key result yields a number of theoretical explanations about how overparameterization, task similarity, and task ordering affect both forgetting and generalization error of CL. More interestingly, by conducting experiments on real datasets using deep neural networks (DNNs), we show that some of these insights even go beyond the linear models and can be carried over to practical setups. In particular, we use concrete examples to show that our results not only explain some interesting empirical observations in recent studies, but also motivate better practical algorithm designs of CL.

Machine unlearning aims to remove targeted knowledge from a trained model while preserving its general capabilities. For autoregressive language models, not all tokens in a forget sample are equally relevant to forgetting. Existing approaches either ignore this heterogeneity or rely on auxiliary models, heuristics, or external annotations to estimate each token's relevance for forgetting. We instead characterize it through the interaction with the retain objective: a token is forget-specific to the extent that minimizing the forget loss on that token does not conflict with retain optimality. We formalize this perspective as a joint optimization problem over the model parameters and the token weights and show that, under a natural separation condition, the resulting objective recovers the oracle forget-specific token support. Motivated by this formulation, we introduce Alternating Token-Weighted Unlearning (ATWU), a lightweight framework that jointly learns token forget-specificity and model parameters during unlearning using a simple linear scorer over the hidden states, without external token level supervision. Across TOFU and RWKU, ATWU achieves state of the art forget-retain trade-offs, outperforming sample-level methods, probability-based token weighting heuristics, and auxiliary-model-based approaches. Moreover, the learned scores align substantially better with ground truth forget-specific spans, indicating that ATWU identifies semantically meaningful token level forgetting signals. Overall, our results suggest that retain conflict provides an effective criterion for identifying what language models should forget, enabling unsupervised learning of token level forget-specificity directly from model representations with minimal computational overhead.

Catastrophic forgetting has been a major challenge in continual learning, where the model needs to learn new tasks with limited or no access to data from previously seen tasks. To tackle this challenge, methods based on knowledge distillation in feature space have been proposed and shown to reduce forgetting. However, most feature distillation methods directly constrain the new features to match the old ones, overlooking the need for plasticity. To achieve a better stability-plasticity trade-off, we propose Backward Feature Projection (BFP), a method for continual learning that allows the new features to change up to a learnable linear transformation of the old features. BFP preserves the linear separability of the old classes while allowing the emergence of new feature directions to accommodate new classes. BFP can be integrated with existing experience replay methods and boost performance by a significant margin. We also demonstrate that BFP helps learn a better representation space, in which linear separability is well preserved during continual learning and linear probing achieves high classification accuracy. The code can be found at https://github.com/rvl-lab-utoronto/BFP

Although Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of tasks, growing concerns have emerged over the misuse of sensitive, copyrighted, or harmful data during training. To address these concerns, unlearning techniques have been developed to remove the influence of specific data without retraining from scratch. However, this paper reveals a critical vulnerability in fine-tuning-based unlearning: a malicious user can craft a manipulated forgetting request that stealthily degrades the model's utility for benign users. We demonstrate this risk through a red-teaming Stealthy Attack (SA), which is inspired by two key limitations of existing unlearning (the inability to constrain the scope of unlearning effect and the failure to distinguish benign tokens from unlearning signals). Prior work has shown that unlearned models tend to memorize forgetting data as unlearning signals, and respond with hallucinations or feigned ignorance when unlearning signals appear in the input. By subtly increasing the presence of common benign tokens in the forgetting data, SA enhances the connection between benign tokens and unlearning signals. As a result, when normal users include such tokens in their prompts, the model exhibits unlearning behaviors, leading to unintended utility degradation. To address this vulnerability, we propose Scope-aware Unlearning (SU), a lightweight enhancement that introduces a scope term into the unlearning objective, encouraging the model to localize the forgetting effect. Our method requires no additional data processing, integrates seamlessly with existing fine-tuning frameworks, and significantly improves robustness against SA. Extensive experiments validate the effectiveness of both SA and SU.

Unlearning seeks to remove specific knowledge from large language models (LLMs), but its effectiveness remains contested. On one side, "forgotten" knowledge can often be recovered through interventions such as light fine-tuning; on the other side, unlearning may induce catastrophic forgetting that degrades general capabilities. Despite active exploration of unlearning methods, interpretability analyses of the mechanism are scarce due to the difficulty of tracing knowledge in LLMs' complex architectures. We address this gap by proposing unPact, an interpretable framework for unlearning via prompt attribution and contribution tracking. Typically, it quantifies each prompt token's influence on outputs, enabling pre- and post-unlearning comparisons to reveal what changes. Across six mainstream unlearning methods, three LLMs, and three benchmarks, we find that: (1) Unlearning appears to be effective by disrupting focus on keywords in prompt; (2) Much of the knowledge is not truly erased and can be recovered by simply emphasizing these keywords in prompts, without modifying the model's weights; (3) Catastrophic forgetting arises from indiscriminate penalization of all tokens. Taken together, our results suggest an unlearning dilemma: existing methods tend either to be insufficient - knowledge remains recoverable by keyword emphasis, or overly destructive - general performance collapses due to catastrophic forgetting, still leaving a gap to reliable unlearning.

Research on stereotypes in large language models (LLMs) has largely focused on English-speaking contexts, due to the lack of datasets in other languages and the high cost of manual annotation in underrepresented cultures. To address this gap, we introduce a cost-efficient human-LLM collaborative annotation framework and apply it to construct EspanStereo, a Spanish-language stereotype dataset spanning multiple Spanish-speaking countries across Europe and Latin America. EspanStereo captures both well-documented stereotypes from prior literature and culturally specific biases absent from English-centric resources. Using LLMs to generate candidate stereotypes and in-culture annotators to validate them, we demonstrate the framework's effectiveness in identifying nuanced, region-specific biases. Our evaluation of Spanish-supporting LLMs using EspanStereo reveals significant variation in stereotypical behavior across countries, highlighting the need for more culturally grounded assessments. Beyond Spanish, our framework is adaptable to other languages and regions, offering a scalable path toward multilingual stereotype benchmarks. This work broadens the scope of stereotype analysis in LLMs and lays the groundwork for comprehensive cross-cultural bias evaluation.

Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle RL's Razor: among all ways to solve a new task, RL prefers those closest in KL to the original model.

In Continual learning (CL) balancing effective adaptation while combating catastrophic forgetting is a central challenge. Many of the recent best-performing methods utilize various forms of prior task data, e.g. a replay buffer, to tackle the catastrophic forgetting problem. Having access to previous task data can be restrictive in many real-world scenarios, for example when task data is sensitive or proprietary. To overcome the necessity of using previous tasks' data, in this work, we start with strong representation learning methods that have been shown to be less prone to forgetting. We propose a holistic approach to jointly learn the representation and class prototypes while maintaining the relevance of old class prototypes and their embedded similarities. Specifically, samples are mapped to an embedding space where the representations are learned using a supervised contrastive loss. Class prototypes are evolved continually in the same latent space, enabling learning and prediction at any point. To continually adapt the prototypes without keeping any prior task data, we propose a novel distillation loss that constrains class prototypes to maintain relative similarities as compared to new task data. This method yields state-of-the-art performance in the task-incremental setting, outperforming methods relying on large amounts of data, and provides strong performance in the class-incremental setting without using any stored data points.

As LLMs are increasingly integrated into user-facing applications, addressing biases that perpetuate societal inequalities is crucial. While much work has gone into measuring or mitigating biases in these models, fewer studies have investigated their origins. Therefore, this study examines the correlation between gender-occupation bias in pre-training data and their manifestation in LLMs, focusing on the Dolma dataset and the OLMo model. Using zero-shot prompting and token co-occurrence analyses, we explore how biases in training data influence model outputs. Our findings reveal that biases present in pre-training data are amplified in model outputs. The study also examines the effects of prompt types, hyperparameters, and instruction-tuning on bias expression, finding instruction-tuning partially alleviating representational bias while still maintaining overall stereotypical gender associations, whereas hyperparameters and prompting variation have a lesser effect on bias expression. Our research traces bias throughout the LLM development pipeline and underscores the importance of mitigating bias at the pretraining stage.

Concept erasure aims to remove specified features from a representation. It can be used to improve fairness (e.g. preventing a classifier from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). In this paper, we introduce LEAst-squares Concept Erasure (LEACE), a closed-form method which provably prevents all linear classifiers from detecting a concept while inflicting the least possible damage to the representation. We apply LEACE to large language models with a novel procedure called "concept scrubbing," which erases target concept information from every layer in the network. We demonstrate the usefulness of our method on two tasks: measuring the reliance of language models on part-of-speech information, and reducing gender bias in BERT embeddings. Code is available at https://github.com/EleutherAI/concept-erasure.

Despite their impressive performance, generative image models trained on large-scale datasets frequently fail to produce images with seemingly simple concepts -- e.g., human hands or objects appearing in groups of four -- that are reasonably expected to appear in the training data. These failure modes have largely been documented anecdotally, leaving open the question of whether they reflect idiosyncratic anomalies or more structural limitations of these models. To address this, we introduce a systematic approach for identifying and characterizing "conceptual blindspots" -- concepts present in the training data but absent or misrepresented in a model's generations. Our method leverages sparse autoencoders (SAEs) to extract interpretable concept embeddings, enabling a quantitative comparison of concept prevalence between real and generated images. We train an archetypal SAE (RA-SAE) on DINOv2 features with 32,000 concepts -- the largest such SAE to date -- enabling fine-grained analysis of conceptual disparities. Applied to four popular generative models (Stable Diffusion 1.5/2.1, PixArt, and Kandinsky), our approach reveals specific suppressed blindspots (e.g., bird feeders, DVD discs, and whitespaces on documents) and exaggerated blindspots (e.g., wood background texture and palm trees). At the individual datapoint level, we further isolate memorization artifacts -- instances where models reproduce highly specific visual templates seen during training. Overall, we propose a theoretically grounded framework for systematically identifying conceptual blindspots in generative models by assessing their conceptual fidelity with respect to the underlying data-generating process.

LLMs can be conveniently adapted to a diverse set of tasks, e.g, prediction, question-answering tasks, etc, using appropriate prompts with few-shot examples. Biased or harmful concepts, e.g. gender or bio-weapons, present in pre-trained LLMs can lead to unsafe or unethical responses for many such prompts. Removing such undesirable concepts robustly across different prompt types remains a challenging problem, since existing unlearning methods typically ignore the impact of prompt variation. In this paper, we explore a novel adversarial approach to use a joint prompt for the main task and concept task prediction. We show that fine-tuning using the ``worst prompt type'' for concept prediction (with the highest concept accuracy) improves the average unlearning performance over a fine-tuning method that uses a combination of all prompt types. Our proposed method, MPSelectTune, is a two-stage approach that minimizes the concept accuracy of the highest accuracy-prompt type, after fine-tuning using a novel multi-task loss using multiple prompt types. Experimental results on four benchmarks show 2 - 15% main task accuracy improvements over recent baselines and while reducing the worst-case concept accuracy by up to 17% compared to recent baselines.

Large Language Models (LLMs) increasingly require selective removal of harmful or sensitive knowledge, called unlearning, yet existing methods and benchmarks fail to evaluate this capability completely. Current approaches rely on disjoint forget and retain sets composed of independent facts, and measure success using simple and direct factual recall. This framing fails to capture a key requirement of unlearning, namely the ability to eliminate harmful behaviors while preserving benign and beneficial knowledge. We argue that effective unlearning must operate at the level of concepts, ensuring complete removal of unsafe applications while maintaining their correct and useful usage, thereby achieving conceptually meaningful and complete unlearning. To better evaluate unlearning techniques from such a practical viewpoint, we introduce the notion of dual-use concepts: concepts that can be used in both harmful and benign contexts. Building on these concepts, we construct a benchmark called ConceptGuard where forget and retain sets are explicitly complementary in concept usage. Our benchmark uniquely enables unlearning to be explored and gauged at the level of concepts, instead of sparse facts, and evaluation is intent-sensitive with the goal of maximizing contextual separation to promote safer behavior. We demonstrate that current unlearning techniques perform poorly under this setting, showing weak contextual separation alongside poor performance in ROUGE and concept-level metrics. Our results reveal strong forgetting-utility trade-offs, limited gains in contextual sensitivity, and poor consistency in concept-level control across methods, and provide ideas for unlearning approaches that better align with real-world safety requirements. Our dataset is publicly available.

Machine unlearning in large language models aims to remove unwanted knowledge while preserving the model's remaining capabilities. Although existing methods use retention objectives or restrict where edits occur, achieving the desired forgetting level can still leave collateral changes that impair non-target behavior. Our recovery comparisons suggest that some of these changes can be reversed while preserving observed forgetting performance. In this work, we present Propose-Then-Project Unlearning (PTP-U), a framework that combines targeted forgetting with the recovery of non-target capabilities. PTP-U first applies local analytic edits to weaken target knowledge associations, then aligns non-target output distributions with those of the original model to recover capabilities while maintaining fixed forgetting constraints. Both stages serve a common goal: satisfying the forgetting requirements while preserving fluent generation and performance on non-target tasks. Across three benchmarks, PTP-U achieves the strongest forgetting-retention trade-off among evaluated methods, reaching 81.22%-91.03% forgetting while preserving 94.20% non-target utility on average. At matched forgetting, PTP-U consistently retains higher non-target utility.

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

Machine unlearning is rapidly becoming a practical requirement, driven by privacy regulations, data errors, and the need to remove harmful or corrupted training samples. Despite this, most existing methods tackle the problem purely from a post-hoc perspective. They attempt to erase the influence of targeted training samples through parameter updates that typically require access to the full training data. This creates a mismatch with real deployment scenarios where unlearning requests can be anticipated, revealing a fundamental limitation of post-hoc approaches. We propose unlearning by design, a novel paradigm in which models are directly trained to support forgetting as an inherent capability. We instantiate this idea with Machine UNlearning via KEY deletion (MUNKEY), a memory augmented transformer that decouples instance-specific memorization from model weights. Here, unlearning corresponds to removing the instance-identifying key, enabling direct zero-shot forgetting without weight updates or access to the original samples or labels. Across natural image benchmarks, fine-grained recognition, and medical datasets, MUNKEY outperforms all post-hoc baselines. Our results establish that unlearning by design enables fast, deployment-oriented unlearning while preserving predictive performance.

Text-to-image generative AI models such as Stable Diffusion are used daily by millions worldwide. However, the extent to which these models exhibit racial and gender stereotypes is not yet fully understood. Here, we document significant biases in Stable Diffusion across six races, two genders, 32 professions, and eight attributes. Additionally, we examine the degree to which Stable Diffusion depicts individuals of the same race as being similar to one another. This analysis reveals significant racial homogenization, e.g., depicting nearly all middle eastern men as dark-skinned, bearded, and wearing a traditional headdress. We then propose novel debiasing solutions that address the above stereotypes. Finally, using a preregistered experiment, we show that being presented with inclusive AI-generated faces reduces people's racial and gender biases, while being presented with non-inclusive ones increases such biases. This persists regardless of whether the images are labeled as AI-generated. Taken together, our findings emphasize the need to address biases and stereotypes in AI-generated content.

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluates model responses at two levels: (i) given an under-informative context, we test how strongly responses reflect social biases, and (ii) given an adequately informative context, we test whether the model's biases override a correct answer choice. We find that models often rely on stereotypes when the context is under-informative, meaning the model's outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conflicts, with this difference widening to over 5 points on examples targeting gender for most models tested.

Editing knowledge in large language models is an attractive capability to have which allows us to correct incorrectly learnt facts during pre-training, as well as update the model with an ever-growing list of new facts. While existing model editing techniques have shown promise, they are usually evaluated using metrics for reliability, specificity and generalization over one or few edits. We argue that for model editing to have practical utility, we must be able to make multiple edits to the same model. With this in mind, we evaluate the current model editing methods at scale, focusing on two state of the art methods: ROME and MEMIT. We find that as the model is edited sequentially with multiple facts, it continually forgets previously edited facts and the ability to perform downstream tasks. This forgetting happens in two phases -- an initial gradual but progressive forgetting phase followed by abrupt or catastrophic forgetting phase. Both gradual and catastrophic forgetting limit the usefulness of model editing methods at scale -- the former making model editing less effective as multiple edits are made to the model while the latter caps the scalability of such model editing methods. Our analysis also highlights other key limitations of ROME and MEMIT at scale. With our work, we push for the development and evaluation of model editing methods keeping scalability in mind.

The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.

Large language models deployed in sensitive applications increasingly require the ability to unlearn specific knowledge, such as user requests, copyrighted materials, or outdated information, without retraining from scratch to ensure regulatory compliance, user privacy, and safety. This task, known as machine unlearning, aims to remove the influence of targeted data (forgetting) while maintaining performance on the remaining data (retention). A common approach is to formulate this as a multi-objective problem and reduce it to a single-objective problem via scalarization, where forgetting and retention losses are combined using a weighted sum. However, this often results in unstable training dynamics and degraded model utility due to conflicting gradient directions. To address these challenges, we propose OFMU, a penalty-based bi-level optimization framework that explicitly prioritizes forgetting while preserving retention through a hierarchical structure. Our method enforces forgetting via an inner maximization step that incorporates a similarity-aware penalty to decorrelate the gradients of the forget and retention objectives, and restores utility through an outer minimization step. To ensure scalability, we develop a two-loop algorithm with provable convergence guarantees under both convex and non-convex regimes. We further provide a rigorous theoretical analysis of convergence rates and show that our approach achieves better trade-offs between forgetting efficacy and model utility compared to prior methods. Extensive experiments across vision and language benchmarks demonstrate that OFMU consistently outperforms existing unlearning methods in both forgetting efficacy and retained utility.

In this paper, we introduce Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), a novel approximate unlearning method. SCAR efficiently eliminates specific information while preserving the model's test accuracy without using a retain set, which is a key component in state-of-the-art approximate unlearning algorithms. Our approach utilizes a modified Mahalanobis distance to guide the unlearning of the feature vectors of the instances to be forgotten, aligning them to the nearest wrong class distribution. Moreover, we propose a distillation-trick mechanism that distills the knowledge of the original model into the unlearning model with out-of-distribution images for retaining the original model's test performance without using any retain set. Importantly, we propose a self-forget version of SCAR that unlearns without having access to the forget set. We experimentally verified the effectiveness of our method, on three public datasets, comparing it with state-of-the-art methods. Our method obtains performance higher than methods that operate without the retain set and comparable w.r.t the best methods that rely on the retain set.

Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity.