After calibrating in the first session, the slow efficient tuning parameters can capture more informativefeatures, improving generalization to incoming classes. Moreover, to further incorporate novel concepts, we strikeabalance between stability and plasticity byfixing slowefficient tuning parameters and continuously updating the fast ones. Specifically, a cross-classification loss with feature alignment is proposed to circumvent catastrophic forgetting.

Large Language Models increasingly possess capabilities that carry dual-use risks. While data filtering has emerged as a pretraining-time mitigation, it faces significant challenges: labeling whether data is harmful is expensive at scale, and given improving sample efficiency with larger models, even small amounts of mislabeled content could give rise to dangerous capabilities. To address risks associated with mislabeled harmful content, prior work proposed Gradient Routing (Cloud et al., 2024) -- a technique that localizes target knowledge into a dedicated subset of model parameters so they can later be removed. We explore an improved variant of Gradient Routing, which we call Selective GradienT Masking (SGTM), with particular focus on evaluating its robustness to label noise. SGTM zero-masks selected gradients such that target domain examples only update their dedicated parameters. We test SGTM's effectiveness in two applications: removing knowledge of one language from a model trained on a bilingual synthetic dataset, and removing biology knowledge from a model trained on English Wikipedia. In both cases SGTM provides better retain/forget trade-off in the presence of labeling errors compared to both data filtering and a previously proposed instantiation of Gradient Routing. Unlike shallow unlearning approaches that can be quickly undone through fine-tuning, SGTM exhibits strong robustness to adversarial fine-tuning, requiring seven times more fine-tuning steps to reach baseline performance on the forget set compared to a finetuning-based unlearning method (RMU). Our results suggest SGTM provides a promising pretraining-time complement to existing safety mitigations, particularly in settings where label noise is unavoidable.

A quarter century ago, Wikipedia's decentralized, crowdsourced, and consensus-driven model replaced the centralized, expert-driven, and authority-based standard for encyclopedic knowledge curation. The emergence of generative AI encyclopedias, such as Grokipedia, possibly presents another potential shift in epistemic evolution. This study investigates whether AI- and human-curated encyclopedias rely on the same foundations of authority. We conducted a multi-scale comparative analysis of the citation networks from 72 matched article pairs, which cite a total of almost 60,000 sources. Using an 8-category epistemic classification, we mapped the "epistemic profiles" of the articles on each platform. Our findings reveal several quantitative and qualitative differences in how knowledge is sourced and encyclopedia claims are epistemologically justified. Grokipedia replaces Wikipedia's heavy reliance on peer-reviewed "Academic & Scholarly" work with a notable increase in "User-generated" and "Civic organization" sources. Comparative network analyses further show that Grokipedia employs very different epistemological profiles when sourcing leisure topics (such as Sports and Entertainment) and more societal sensitive civic topics (such as Politics & Conflicts, Geographical Entities, and General Knowledge & Society). Finally, we find a "scaling-law for AI-generated knowledge sourcing" that shows a linear relationship between article length and citation density, which is distinct from collective human reference sourcing. We conclude that this first implementation of an LLM-based encyclopedia does not merely automate knowledge production but restructures it. Given the notable changes and the important role of encyclopedias, we suggest the continuation and deepening of algorithm audits, such as the one presented here, in order to understand the ongoing epistemological shifts.

Accurately answering a question about a given image requires combining observations with general knowledge. While this is effortless for humans, reasoning with general knowledge remains an algorithmic challenge.

Recently, the general-to-customized paradigm has emerged as the dominant approach for Cross-Modal Retrieval (CMR), which reconciles the distribution shift problem between the source domain and the target domain. However, existing general-to-customized CMR methods typically assume that the entire target-domain data is available, which is easily violated in real-world scenarios and thus inevitably suffer from the query shift (QS) problem. Specifically, query shift embraces the following two characteristics and thus poses new challenges to CMR. i) Online Shift: real-world queries always arrive in an online manner, rendering it impractical to access the entire query set beforehand for customization approaches; ii) Diverse Shift: even with domain customization, the CMR models struggle to satisfy queries from diverse users or scenarios, leaving an urgent need to accommodate diverse queries. In this paper, we observe that QS would not only undermine the well-structured common space inherited from the source model, but also steer the model toward forgetting the indispensable general knowledge for CMR. Inspired by the observations, we propose a novel method for achieving online and harmonious adaptation against QS, dubbed Robust adaptation with quEry ShifT (REST). To deal with online shift, REST first refines the retrieval results to formulate the query predictions and accordingly designs a QS-robust objective function on these predictions to preserve the well-established common space in an online manner. As for tackling the more challenging diverse shift, REST employs a gradient decoupling module to dexterously manipulate the gradients during the adaptation process, thus preventing the CMR model from forgetting the general knowledge. Extensive experiments on 20 benchmarks across three CMR tasks verify the effectiveness of our method against QS.

Multimodal Large Language Models (MLLMs) extend foundation models to real-world applications by integrating inputs such as text and vision. However, their broad knowledge capacity raises growing concerns about privacy leakage, toxicity mitigation, and intellectual property violations. Machine Unlearning (MU) offers a practical solution by selectively forgetting targeted knowledge while preserving overall model utility. When applied to MLLMs, existing neuron-editing-based MU approaches face two fundamental challenges: (1) forgetting becomes inconsistent across modalities because existing point-wise attribution methods fail to capture the structured, layer-by-layer information flow that connects different modalities; and (2) general knowledge performance declines when sensitive neurons that also support important reasoning paths are pruned, as this disrupts the model's ability to generalize. To alleviate these limitations, we propose a multimodal influential neuron path editor (MIP-Editor) for MU. Our approach introduces modality-specific attribution scores to identify influential neuron paths responsible for encoding forget-set knowledge and applies influential-path-aware neuron-editing via representation misdirection. This strategy also enables effective and coordinated forgetting across modalities while preserving the model's general capabilities. Experimental results demonstrate that MIP-Editor achieves a superior unlearning performance on multimodal tasks, with a maximum forgetting rate of 87.75% and up to 54.26% improvement in general knowledge retention. On textual tasks, MIP-Editor achieves up to 80.65% forgetting and preserves 77.9% of general performance.

Large Language Models (LLMs) possess remarkable generalization capabilities but struggle with multi-task adaptation, particularly in balancing knowledge retention with task-specific specialization. Conventional fine-tuning methods suffer from catastrophic forgetting and substantial resource consumption, while existing parameter-efficient methods perform suboptimally in complex multi-task scenarios. To address this, we propose Contextual Attention Modulation (CAM), a novel mechanism that dynamically modulates the representations of self-attention modules in LLMs. CAM enhances task-specific features while preserving general knowledge, thereby facilitating more effective and efficient adaptation. For effective multi-task adaptation, CAM is integrated into our Hybrid Contextual Attention Modulation (HyCAM) framework, which combines a shared, full-parameter CAM module with multiple specialized, lightweight CAM modules, enhanced by a dynamic routing strategy for adaptive knowledge fusion. Extensive experiments on heterogeneous tasks, including question answering, code generation, and logical reasoning, demonstrate that our approach significantly outperforms existing approaches, achieving an average performance improvement of 3.65%. The implemented code and data are available to ease reproducibility at https://github.com/Applied-Machine-Learning-Lab/HyCAM.

Conventional continual pretraining (CPT) for large language model (LLM) domain adaptation often suffers from catastrophic forgetting and limited domain capacity. Existing strategies adopt layer expansion, introducing additional trainable parameters to accommodate new knowledge. However, the uniform expansion and updates still entangle general and domain learning, undermining its effectiveness. Our pilot studies reveal that LLMs exhibit functional specialization, where layers and units differentially encode general-critical capabilities, suggesting that parameter expansion and optimization should be function-aware. We then propose ADEPT, Adaptive Expansion and Dynamic Decoupled Tuning for continual pretraining, a two-stage framework for domain-adaptive CPT. ADEPT first performs General-Competence Guided Selective Layer Expansion, duplicating layers least critical for the general domain to increase representational capacity while minimizing interference with general knowledge. It then applies Adaptive Unit-Wise Decoupled Tuning, disentangling parameter units within expanded layers according to their general-domain importance and assigning asymmetric learning rates to balance knowledge injection and retention. Experiments on mathematical and medical benchmarks show that ADEPT outperforms full-parameter CPT by up to 5.76% on the general domain and 5.58% on the target domain with only 15% of parameters tuned and less than 50% training time. Ablation studies, theoretical analysis, and extended investigations further demonstrate the necessity of targeted expansion and decoupled optimization, providing new principles for efficient and robust domain-adaptive CPT. Our code is open-sourced at https://github.com/PuppyKnightUniversity/ADEPT

Automated approaches to answer patient-posed health questions are rising, but selecting among systems requires reliable evaluation. The current gold standard for evaluating the free-text artificial intelligence (AI) responses--human expert review--is labor-intensive and slow, limiting scalability. Automated metrics are promising yet variably aligned with human judgments and often context-dependent. To address the feasibility of automating the evaluation of AI responses to hospitalization-related questions posed by patients, we conducted a large systematic study of evaluation approaches. Across 100 patient cases, we collected responses from 28 AI systems (2800 total) and assessed them along three dimensions: whether a system response (1) answers the question, (2) appropriately uses clinical note evidence, and (3) uses general medical knowledge. Using clinician-authored reference answers to anchor metrics, automated rankings closely matched expert ratings. Our findings suggest that carefully designed automated evaluation can scale comparative assessment of AI systems and support patient-clinician communication.