Singapore

Verbatim memorization in Large Language Models (LLMs) is a multifaceted phenomenon involving distinct underlying mechanisms. We introduce a novel method to analyze the different forms of memorization described by the existing taxonomy. Specifically, we train Convolutional Neural Networks (CNNs) on the attention weights of the LLM and evaluate the alignment between this taxonomy and the attention weights involved in decoding.

We find that the existing taxonomy performs poorly and fails to reflect distinct mechanisms within the attention blocks. We propose a new taxonomy that maximizes alignment with the attention weights, consisting of three categories: memorized samples that are guessed using language modeling abilities, memorized samples that are recalled due to high duplication in the training set, and non-memorized samples. Our results reveal that few-shot verbatim memorization does not correspond to a distinct attention mechanism. We also show that a significant proportion of extractable samples are in fact guessed by the model and should therefore be studied separately. Finally, we develop a custom visual interpretability technique to localize the regions of the attention weights involved in each form of memorization.

AAAI 2026

Guess or Recall? Training CNNs to Classify and Localize Memorization in LLMs

training data memorization

large language models

privacy

security

## Welcome to EMNLP 2022!
I am delighted to welcome you to EMNLP 2022! I believe this conference has been complicated beyond any precedent. Over the past year, it’s been thrilling to see the organization team approach each new puzzle with creativity and enthusiasm. We hope that those participating in Abu Dhabi as well as those joining remotely will leave the conference feeling newly inspired by the program and newly connected to our ever-growing community. Following EMNLP 2021 and major NLP conferences since, EMNLP 2022 is “hybrid,” serving both virtual and in-person participants.

Our key innovations for EMNLP 2022 include:

* EMNLP 2022 is “hybrid” in a second sense, as well: we allowed both direct and rolling review paper submissions, building on the pilot experiment of EMNLP 2021, which considered a small number of ARR submissions. 
* Familiar from NAACL but new to EMNLP, we’ve added an industry track.
* During the conference, “portals” will link virtual poster sessions to in-person conference participants during poster sessions each day.
* The first ACL-family conference in the United Arab Emirates.

 *Message from Noah A. Smith, University of Washington and Allen Institute for AI, Seattle, Washington, USA* 
***EMNLP 2022 General Chair***
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/9eec7d4a287ee18c278b08229290aa83.png)](https://drive.google.com/file/d/1OlPv6QBeo62VVTughj2jkiLeyHd1WnUt/view)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/a3db7a768409f05192210d98601edb25.png)](https://emnlp2022.rocket.chat/)

To access this site you need to register. Please register [here](https://2022.emnlp.org/registration/).

Register here

EMNLP 2022

Welcome!
EMNLP 2022 will take place in Abu Dhabi from December 7th to December 11th, 2022. And it will be held in hybrid mode, both online and offline.

technical paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

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-26 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.

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

In the proposed talk, I will present an overview of my
research under the theme “Towards Trustworthy Multimodal AI
Systems.” My talk will consist of two main parts. The first
section will survey my work on multimodal explainable
artificial intelligence (XAI), including challenges of
scaling current techniques (specifically, mechanistic
interpretability tools) to multimodal models and proposed
algorithms to mitigate them. The second section will focus
on my research in addressing safety and reasoning problems
of frontier medical models.

Towards Trustworthy Multimodal AI Systems

Organic photovoltaic (OPV) materials offer a promising pathway for sustainable energy generation. However, their development is hindered by the challenge of identifying high-performance donor-acceptor pairs with optimal power conversion efficiencies (PCEs). Most existing design strategies focus exclusively on either the donor or the acceptor, rather than employing a unified model capable of designing both components. In this work, we introduce a dual-pronged machine learning framework for OPV discovery, integrating predictive modeling and generative molecular design. In this study, we propose the newly curated Organic Photovoltaic Donor-Acceptor Dataset (OPV$^2$D), the largest of its kind, comprising 2,000 experimentally characterized donor-acceptor pairs. This dataset serves as a comprehensive foundation for model training and evaluation. To enable accurate property prediction in organic photovoltaic (OPV) materials, we first introduce the Organic Photovoltaic Classifier (OPVC) to predict the likelihood that a given material exhibits OPV behavior. Complementing this, we develop a hierarchical graph neural network framework that integrates multi-task learning and cross-modal donor–acceptor interaction modeling. This framework includes the Molecular Orbital Energy Estimator (MOE$^2$) for predicting the highest occupied molecular orbital–lowest unoccupied molecular orbital (HOMO–LUMO) energy levels, and the Photovoltaic Performance Predictor (P$^3$) for estimating power conversion efficiency (PCE), both achieving state-of-the-art accuracy. In addition, we introduce the Material Generative Pretrained Transformer (MatGPT) to generate synthetically accessible organic semiconductors. Building on this, we propose a reinforcement learning strategy with three-objective policy optimization guides molecular generation while preserving chemical validity. By bridging molecular representation learning with device performance prediction, our framework advances computational OPV material discovery. All code and data will be open-sourced upon publication.

CycleChemist: A Dual-Pronged Machine Learning Framework for Organic Photovoltaic Discovery

Recent advances in context optimization (CoOp) guided by large language model (LLM)–distilled medical semantic priors offer a scalable alternative to manual prompt engineering and full fine-tuning for adapting biomedical CLIP-based vision-language models (VLMs). However, prompt learning in this context is challenged by semantic misalignment between LLMs and CLIP variants due to divergent training corpora and model architectures; it further lacks scalability across continuously evolving families of foundation models. More critically, pairwise multimodal alignment via conventional Euclidean-space optimization lacks the capacity to model unified representations or apply localized geometric constraints, which tends to amplify modality gaps in complex biomedical imaging and destabilize few-shot adaptation. In this work, we propose vMFCoOp, a framework that inversely estimates von Mises–Fisher (vMF) distributions on a shared Hyperspherical Manifold, aligning semantic biases between arbitrary LLMs and CLIP backbones via Unified Semantic Anchors to achieve robust biomedical prompting and superior few-shot classification. Grounded in three complementary constraints, vMFCoOp demonstrates consistent improvements across 14 medical datasets, 12 medical imaging modalities, and 13 anatomical regions, outperforming state-of-the-art methods in accuracy, generalization, and clinical applicability.

vMFCoOp: Towards Equilibrium on a Unified Hyperspherical Manifold for Prompting Biomedical VLMs

Modern planning systems utilize various plan representations - sequential, parallel, partially ordered (PO), and partial-order causal link (POCL) - each with different models for concurrency.
These formalisms are often implicitly assumed to have the same base properties, particularly regarding makespan.
We challenge this assumption, proving the relationship between them is fundamentally asymmetric. Our analysis shows conversions from plans with rigid concurrency layers (sequential, parallel) to those with flexible partial orders (PO, POCL) can preserve makespan.
However, the reverse generally fails; the flexible orderings in PO/POCL plans can yield shorter makespans for solutions that cannot be represented in parallel plans without serialization.
We prove that finding an optimal parallel representation for a given POCL plan is $\textsf{NP}$-complete, resolving a key question about their practical interchangeability.
We also provide tight complexity bounds for makespan-bounded plan existence.
Notably, our results disprove a claim in the literature that planning graph-based planners maximize concurrency by minimizing the critical path in derived PO plans.

Makespan Investigations of Sequential, Parallel, PO, and POCL Plans

In this paper, we rethink model agent behaviors from a geometric structure perspective in multi-agent reinforcement learning.
Modeling agent behaviors is essential for understanding how agents interact and facilitating effective decisions. The key lies in capturing the dependencies and sequential relationships among agent decisions. Since each decision influences the subsequent choices, this forms a hierarchical and nested tree-like structure of interdependencies. While modeling tree-like data in Euclidean spaces could cause distortion, which results in a loss of agent decision structure information. Motivated by this, we reconsider model agent behaviors in hyperbolic space and propose the Hyperbolic Multi-Agent Representations (HMAR) method, which projects the agent behaviors into a Poincaré ball and leverages hyperbolic neural networks to learn agent policy representations. Additionally, we designed a contrastive loss function to train this network, minimizing the distance in feature space between different representations of the same agent while maximizing the distance between representations of distinct agents. Experimental results provide empirical evidence for the effectiveness of the HMAR method in cooperative and competitive environments, demonstrating the potential of hyperbolic agent representations for effective decision-making in multi-agent environments.

Exploiting Geometric Structures for Modeling Multi-Agent Behaviors: A New Thinking

Chain-of-Thought (CoT) prompting has recently shown significant promise across various NLP and computer vision tasks by explicitly generating intermediate reasoning steps. However, we find that reinforcement learning (RL)-based fine-tuned CoT reasoning can paradoxically degrade performance in Visual Grounding tasks, particularly as CoT outputs become lengthy or complex. Additionally, our analysis reveals that increased dataset size does not always enhance performance due to varying data complexities. Motivated by these findings, we propose Curriculum-based Relative Policy Optimization (CuRPO), a novel training strategy that leverages CoT length and generalized Intersection over Union (gIoU) rewards as complexity indicators to progressively structure training data from simpler to more challenging examples. Extensive experiments on RefCOCO, RefCOCO+, RefCOCOg, and LISA datasets demonstrate the effectiveness of our approach. CuRPO consistently outperforms existing methods, including Visual-RFT, with notable improvements of up to +15.49 mAP on RefCOCO. Moreover, CuRPO exhibits exceptional efficiency and robustness, delivering strong localization performance even in few-shot learning scenarios, particularly benefiting tasks characterized by ambiguous and intricate textual descriptions.

Start Small, Think Big: Curriculum-based Relative Policy Optimization for Visual Grounding

Zero-Shot Relation Triplet Extraction (ZSRTE) aims to extract head-tail entity pairs and their corresponding relations from sentences, where the relations available during inference are not seen during training. Existing methods typically assume that entities are continuous; however, in practice, entities can be discontinuous, which poses challenges to these approaches. To address this issue, we are the first to discuss and study the ZSRTE task involving discontinuous entities, and propose an innovative BoG framework, which is based on our proposed Boundary Token Graph structure. This method first predicts and adds edges between boundary tokens of (dis)continuous entities to construct a token graph, and then innovatively transforms the relation triplet extraction task into a process of finding paths in the graph. Additionally, we design a Boundary Token-Aware Prompt for each relation to further enhance the interaction between boundary tokens and relation semantics. Experimental results on four ZSRTE datasets—with or without discontinuous entities—consistently demonstrate that our method outperforms previous approaches, achieving state-of-the-art results.

A Boundary Token Graph for Zero-Shot Relation Triplet Extraction Involving Discontinuous Entities

Covariance matrix estimation in high-dimensional settings is a fundamental problem in machine learning and data science. To mitigate the challenges posed by high dimensionality, a common structural assumption is the sparsity, which posits that most variable pairs exhibit negligible correlations. We revisit the seminal positive definite sparse covariance estimation (PDSCE) formulation proposed by Rothman (2012). In contrast to many earlier approaches, this formulation guarantees that the estimator is positive definite, thereby ensuring the well-posedness of the problem. However, the incorporation of a log-determinant barrier also introduces significant optimization challenges. Existing algorithms for solving the PDSCE problem suffer from several limitations, such as high computational complexity, poor scalability, and the need for explicit mechanisms to maintain positive definiteness. To address these issues, we propose a dual proximal gradient method (DPGM) that, unlike existing primal-space approaches, operates in the dual space. This dual method offers several key advantages. First, DPGM inherently preserves positive definiteness, eliminating the need for complex mechanisms to enforce it. Second, DPGM achieves improved iteration complexity over primal-space methods. Third, the worst-case step size for DPGM is a constant independent of the ambient dimension, making the method particularly suitable for high-dimensional applications. Extensive numerical experiments demonstrate that DPGM consistently outperforms existing algorithms, validating its effectiveness and scalability for high-dimensional sparse covariance estimation.

Positive Definite Sparse Covariance Estimation via Dual Space Optimization

Existing backdoor attacks on multivariate time series (MTS) forecasting enforce strict temporal and dimensional coupling between triggers and target patterns, requiring synchronous activation at fixed positions across variables. However, realistic scenarios often demand delayed and variable-specific activation.
We identify this critical unmet need and propose TDBA, a temporally decoupled backdoor attack framework for MTS forecasting. By injecting triggers that encode the expected location of the target pattern, TDBA enables the activation of the target pattern at any positions within the forecasted data, with the activation position flexibly varying across different variable dimensions.
TDBA introduces two core modules: (1) a position-guided trigger generation mechanism that leverages smoothed Gaussian priors to generate triggers that are position-related to the predefined target pattern; and (2) a position-aware optimization module that assigns soft weights based on trigger completeness, pattern coverage, and temporal offset, facilitating targeted and stealthy attack optimization.
Extensive experiments on real-world datasets show that TDBA consistently outperforms existing baselines in both effectiveness and stealthiness. Ablation studies confirm the controllability and robustness of its design.

Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting

Modern manufacturing environments demand not only accurate predictions but also interpretable insights to process anomalies, root causes, and potential interventions. Existing AI systems often function as isolated black boxes, lacking the seamless integration of prediction, explanation, and causal reasoning required for a unified decision-support solution. This fragmentation limits their trustworthiness and practical utility in high-stakes industrial environments. In this work, we present CausalTrace, a neurosymbolic causal analysis module integrated into the SmartPilot industrial CoPilot. CausalTrace performs data-driven causal analysis enriched by industrial ontologies and knowledge graphs, including advanced functions such as causal discovery, counterfactual reasoning, and root cause analysis (RCA). It supports real-time operator interaction and is designed to complement existing agents by offering transparent, explainable decision support. We conducted a comprehensive evaluation of CausalTrace using multiple causal assessment methods and the C3AN framework (i.e. Custom, Compact, Composite AI with Neurosymbolic Integration), which spans principles of robustness, intelligence, and trustworthiness. In an academic rocket assembly testbed, CausalTrace achieved substantial agreement with domain experts (ROUGE-1: 0.91 in ontology QA) and strong RCA performance (MAP@3: 94%, PR@2: 97%, MRR: 0.92, Jaccard: 0.92). It also attained 4.59/5 in the C3AN evaluation, demonstrating precision and reliability for live deployment.

Downloads

Next from AAAI 2026

Towards Trustworthy Multimodal AI Systems

Similar lecture

Finding Memo: Extractive Memorization in Constrained Sequence Generation Tasks