United States

Large Language Models (LLMs) need to adapt to the continuous changes in data, tasks, and user preferences. Due to their massive size and the high costs associated with training, LLMs are not suitable for frequent retraining. However, updates are necessary to keep them in sync with rapidly evolving human knowledge. To address these challenges, this paper proposes the **C**ompression **M**emory **T**raining (**CMT**) method, an efficient and effective online adaptation framework for LLMs that features robust knowledge retention capabilities.
Inspired by human memory mechanisms, CMT compresses and extracts information from new documents to be stored in a memory bank. When answering to queries related to these new documents, the model aggregates these document memories from the memory bank to better answer user questions. The parameters of the LLM itself do not change during training and inference, reducing the risk of catastrophic forgetting. To enhance the encoding, retrieval, and aggregation of memory, we further propose three new general and flexible techniques, including memory-aware objective, self-matching and top-$k$ aggregation. Extensive experiments conducted on three continual learning datasets (i.e., StreamingQA, SQuAD and ArchivalQA) demonstrate that the proposed method improves model adaptability and robustness across multiple base LLMs (e.g., +4.07 EM \&amp; +4.19 F1 in StreamingQA with *Llama-2-7b*).

AAAI 2025

CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models

Large Language Models (LLMs) need to adapt to the continuous changes in data, tasks, and user preferences. Due to their massive size and the high costs associated with training, LLMs are not suitable for frequent retraining. However, updates are necessary to keep them in sync with rapidly evolving human knowledge. To address these challenges, this paper proposes the **C**ompression **M**emory **T**raining (**CMT**) method, an efficient and effective online adaptation framework for LLMs that features robust knowledge retention capabilities.
Inspired by human memory mechanisms, CMT compresses and extracts information from new documents to be stored in a memory bank. When answering to queries related to these new documents, the model aggregates these document memories from the memory bank to better answer user questions. The parameters of the LLM itself do not change during training and inference, reducing the risk of catastrophic forgetting. To enhance the encoding, retrieval, and aggregation of memory, we further propose three new general and flexible techniques, including memory-aware objective, self-matching and top-$k$ aggregation. Extensive experiments conducted on three continual learning datasets (i.e., StreamingQA, SQuAD and ArchivalQA) demonstrate that the proposed method improves model adaptability and robustness across multiple base LLMs (e.g., +4.07 EM \& +4.19 F1 in StreamingQA with *Llama-2-7b*).

poster

We are pleased to announce the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), which will be held in Philadelphia, Pennsylvania at the Pennsylvania Convention Center from February 25 to March 4, 2025.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

### [Invited Speakers](https://aaai.org/conference/aaai/aaai-25/aaai-25-invited-speakers/)

Register [here](https://aaai.org/conference/aaai/aaai-25/registration/)

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.



Generative AI presents transformative potential across various domains, from creative arts to scientific visualization. However, the utility of AI-generated imagery is often compromised by visual flaws, including anatomical inaccuracies, improper object placements, and misplaced textual elements. These imperfections pose significant challenges for practical applications. To overcome these limitations, we introduce Yuan, a novel framework that autonomously corrects visual imperfections in text-to-image synthesis. Yuan uniquely conditions on both the textual prompt and the segmented image, generating precise masks that identify areas in need of refinement without requiring manual intervention—a common constraint in previous methodologies. Following the automated masking process, an advanced inpainting module seamlessly integrates contextually coherent content into the identified regions, preserving the integrity and fidelity of the original image and associated text prompts. Through extensive experimentation on publicly available datasets such as ImageNet100 and Stanford Dogs, along with a custom-generated dataset, Yuan demonstrated superior performance in eliminating visual imperfections. Our approach consistently achieved higher scores in quantitative metrics, including NIQE, BRISQUE, and PI, alongside favorable qualitative evaluations. These results underscore Yuan's potential to significantly enhance the quality and applicability of AI-generated images across diverse fields. The source code will be made available upon acceptance.

Yuan: Yielding Unblemished Aesthetics Through a Unified Network for Visual Imperfections Removal in Generated Images

Speech watermarking techniques can proactively mitigate the potential harmful consequences of instant voice cloning techniques. These techniques involve the insertion of signals into speech that are imperceptible to humans but can be detected by algorithms. Previous approaches typically embed watermark messages into continuous space. However, intuitively, embedding watermark information into robust discrete latent space can significantly improve the robustness of watermarking systems. In this paper, we propose DiscreteWM, a novel speech watermarking framework that injects watermarks into the discrete intermediate representations of speech. Specifically, we map speech into discrete latent space with a vector-quantized autoencoder and inject watermarks by changing the modular arithmetic relation of discrete IDs. To ensure the imperceptibility of watermarks, we also propose a manipulator model to select the candidate tokens for watermark embedding. Experimental results demonstrate that our framework achieves state-of-the-art performance in robustness and imperceptibility, simultaneously. Moreover, our flexible frame-wise approach can serve as an efficient solution for both voice cloning detection and information hiding. Additionally, DiscreteWM can encode 1 to 150 bits of watermark information within a 1-second speech clip, indicating its encoding capacity. Audio samples are available at https://DiscreteWM.github.io/discrete_wm.

Speech Watermarking with Discrete Intermediate Representations

This paper studies solving the system of nonlinear equations $ {\bf F}(\bf x)={\bf 0}$, where ${\bf F}:{\mathbb R}^{d}\to{\mathbb R}^d$. 
We propose Gram-Reduced Levenberg--Marquardt (GRLM) method which updates the matrix ${\bf J}(\cdot)^\top{\bf J}(\cdot)$ in every $m$ iterations, where ${\bf J}(\cdot)$ is the Jacobian of ${\bf F}(\cdot)$.
Our method has  a global convergence guarantee without relying on any step of line-search or solving  sub-problems.  We prove our method takes at most $\mathcal{O}(m^2+m^{-0.5}\epsilon^{-2.5})$ iterations to find an $\epsilon$-stationary point of $\frac{1}{2}\|\|{\bf F}(\cdot)\|\|^2$, which leads to overall computation cost of $\mathcal{O}(d^3\epsilon^{-1}+d^2\epsilon^{-2})$ by taking $m=\Theta(\epsilon^{-1})$. Our results are strictly better than the cost of $\mathcal{O}(d^3\epsilon^{-2})$ for existing LM methods. We also show the proposed method enjoys local superlinear convergence rate under the non-degenerate assumption. We provide rigorous experiments on important applications of scientific computing and machine learning to validate the efficiency of the proposed methods.

An Enhanced Levenberg--Marquardt Method via Gram Reduction

LLMs obtain remarkable performance but suffer from hallucinations. Most research on detecting hallucination focuses on the questions with short and concrete correct answers that are easy to check the faithfulness. Hallucination detections for text generation with open-ended answers are more challenging. Some researchers use external knowledge to detect hallucinations in generated texts, but external resources for specific scenarios are hard to access. Recent studies on detecting hallucinations in long text without external resources conduct consistency comparison among multiple sampled outputs. To handle long texts, researchers split long texts into multiple facts and individually compare the consistency of each pairs of facts. However, these methods (1) hardly achieve alignment among multiple facts; (2) overlook dependencies between multiple contextual facts. In this paper, we propose a graph-based context-aware (GCA) hallucination detection for text generations, which aligns knowledge facts and considers the dependencies between contextual knowledge triples in consistency comparison. Particularly, to align multiple facts, we conduct a triple-oriented response segmentation to extract multiple knowledge triples. To model dependencies among contextual knowledge triple (facts), we construct contextual triple into a graph and enhance triples' interactions via message passing and aggregating via RGCN. To avoid the omission of knowledge triples in long text, we conduct a LLM-based reverse verification via reconstructing the knowledge triples. Experiments show that our model enhances hallucination detection and excels  all baselines.

Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling

Bundle recommendation aims to enhance user experience by suggesting complementary items that users are likely to purchase together. While recent advances in recommendation systems have shown promise, there are still significant challenges: i) The dynamic nature of user preferences and interactions introduces noise that can distort the effectiveness of recommendations. ii) Existing methods often show poor robustness when facing the sparsity of user interactions with bundles in real-world scenarios. To tackle these issues, we introduce a disentangled contrastive bundle recommendation (DCBR) framework with conditional diffusion. First, we propose a conditional bundle diffusion model for denoising the user-bundle interaction graph, introducing a bundle latent consistency constraint during the optimization process to mitigate the degradation of original interaction information. Subsequently, we design a triple-view denoised graph learning module to obtain effective representations from multiple views. Furthermore, we introduce a dual-view disentangled contrastive learning paradigm, considering the potential relationships between the user-bundle view and two additional views (user-item and bundle-item). By maximizing the similarity between positive samples in these contrastive views, we generate disentangled contrastive signals, overcoming interaction sparsity and alleviating noise issues. Our experimental evaluations on four benchmark datasets reveal that DCBR significantly outperforms state-of-the-art methods.

Disentangled Contrastive Bundle Recommendation with Conditional Diffusion

Recommender Systems (RS) are widely applied for navigating information, and Collaborative Filtering (CF) is one of prominent RS techniques due to the advantages of domain independence and easy interpretation.
Among the numerous CF methods,  Variational Autoencoders (VAE), benefiting from modeling in a probabilitistic way, stands out in capturing user preferences through representation learning. Despite the superiority, VAE-based CF models still suffers from two challenging problems: (1) Exposure bias: models in training state are narrowly exposed to a limited, biased sample of data, leading to a skewed understanding of users' true preferences; (2) Posterior collapse: models excessively simplify the learned latent variable distributions, generating na\"ive representations that are unable to encapsulate the complex data patterns and thereby resulting improper recommendations. In this paper, we propose a Debiased and Representation-enhanced Variational AutoEncoder (DR-VAE) framework for collaborative recommendations. Specifically, for exposure bias problem, DR-VAE incorporates a Debiasing Estimator, significantly mitigating the impact of exposure bias. For poster collapse issue, DR-VAE innovatively introduces a Flow-based Representation Enhancement module, ensuring us to encapsulate complex data patterns by fitting complex and intricate posterior distributions directly. We provide experimental validations over four datasets to substantiate the efficacy of our DR-VAE framework.

DR-VAE: Debiased and Representation-enhanced Variational Autoencoder for Collaborative Recommendation

For a data-generating process for random variables that can be described with a linear structural equation model, we consider a situation in which (i) a set of covariates satisfying the back-door criterion cannot be observed or (ii) such a set can be observed, but standard statistical estimation methods cannot be applied to estimate causal effects because of multicollinearity/high-dimensional data problems. We propose a novel two-stage penalized regression approach, the penalized covariate-mediator selection operator (PCM Selector), to estimate the causal effects in such scenarios. Unlike existing penalized regression analyses, when a set of intermediate variables is available, PCM Selector provides a consistent or less biased estimator of the causal effect. In addition, PCM Selector provides a variable selection procedure for intermediate variables to obtain better estimation accuracy of the causal effects than does the back-door criterion.

PCM Selector: Penalized Covariate-Mediator Selection Operator for Evaluating Linear Causal Effects

Temporal Action Detection (TAD) is fundamental yet challenging for real-world video applications. Leveraging the unique benefits of transformers, various DETR-based approaches have been adopted in TAD. However, it has recently been identified that the attention collapse in self-attention causes the performance degradation of DETR for TAD. Building upon previous research, this paper newly addresses the attention collapse problem in cross-attention within DETR-based TAD methods. Moreover, our findings reveal that cross-attention exhibits patterns distinct from predictions, indicating a short-cut phenomenon. To resolve this, we propose a new framework, Prediction-Feedback DETR (Pred-DETR), which utilizes predictions to restore the collapse and align the cross- and self-attention with predictions. Specifically, we devise novel prediction-feedback objectives using guidance from the relations of the predictions. As a result, Pred-DETR significantly alleviates the collapse and achieves state-of-the-art performance among DETR-based methods on various challenging benchmarks including THUMOS14, ActivityNet-v1.3, HACS, and FineAction.

Prediction-Feedback DETR for Temporal Action Detection

Universal domain adaptation (UniDA) transfers knowledge from a labelled source domain to an unlabelled target domain under domain-shift and category-shift for annotation. In reality, due to privacy protection or other limits, not only source data but also pre-trained models on it may be unavailable when training on target data. In this paper, we go a step further to explore the black-box universal domain adaptation (B$^{2}$-UniDA) problem. It requires tackling the labelling task under shifts by only accessing the interface of pre-trained source models. To this end, we introduce GSS which proposes a novel sample selection criterion based on gradient descent and Bayes' Theorem to identify samples of potential unknown classes. This criterion doesn't require manually-set thresholds depending on data used and is suitable for various datasets. GSS builds an open-set classifier and enables it to estimate probabilities of belonging to each class including the unknown category and adjust estimates adaptively. To overcome class-imbalance, especially imbalance between the unknown and known classes, we propose a balancing mechanism by measuring training status and estimating DA type. In addition to distilling knowledge from source model outputs, we focus on mining the categorical structure of target domain by self-training. Experiments on benchmarks show the state-of-the-art performance of GSS compared to typical methods, including source models or source data dependent methods.

Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation

Spike cameras, as innovative neuromorphic devices, generate continuous spike streams to capture high-speed scenes with lower bandwidth and higher dynamic range than traditional RGB cameras. However, reconstructing high-quality video from these inputs, especially under low-light conditions, remains challenging. Traditional methods often rely on synthetic data or basic reconstructions as supervision signals, but these approaches falter when dealing with noisy or low-quality spike signals, leading to performance degradation. This is primarily due to inadequate noise modeling, the domain gap between synthetic and real datasets, and the reliance on low-quality pseudo labels, resulting in images with unclear textures, excessive noise, and diminished brightness.
To address these challenges, we introduce a novel reconstruction framework that goes beyond traditional training paradigms. Instead of relying solely on visual data, we incorporate textual descriptions and unpaired high-quality datasets as new forms of supervision. Textual descriptions provide additional context that guides the network's feature reconstruction, while high-quality datasets help produce sharp latent images. Our experiments on real-world low-light datasets, such as U-CALTECH and U-CIFAR, demonstrate that this approach significantly enhances texture clarity and luminance balance. Furthermore, the reconstructed images are well-aligned with the broader visual features needed for downstream tasks, ensuring more robust and versatile performance in challenging environments.

Content not yet available

CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models

Downloads

Next from AAAI 2025

Yuan: Yielding Unblemished Aesthetics Through a Unified Network for Visual Imperfections Removal in Generated Images

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Content not yet available

CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAAI 2025

Yuan: Yielding Unblemished Aesthetics Through a Unified Network for Visual Imperfections Removal in Generated Images

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads