United States

Recent advances in large language model assistants have made them indispensable, raising significant concerns over managing their safety. Automated red teaming offers a promising alternative to the labor-intensive and error-prone manual probing for vulnerabilities, providing more consistent and scalable safety evaluations. However, existing approaches often compromise diversity by focusing on maximizing attack success rate. Additionally, methods that decrease the cosine similarity from historical embeddings with semantic diversity rewards lead to novelty stagnation as history grows. To address these issues, we introduce DiveR-CT, which relaxes conventional constraints on the objective and semantic reward, granting greater freedom for the policy to enhance diversity. Our experiments demonstrate DiveR-CT&#39;s marked superiority over baselines by 1) generating data that perform better in various diversity metrics across different attack success rate levels, 2) better-enhancing resiliency in blue team models through safety tuning based on collected data, 3) allowing dynamic control of objective weights for reliable and controllable attack success rates, and 4) reducing susceptibility to reward overoptimization. Overall, our method provides an effective and efficient approach to LLM red teaming, accelerating real-world deployment.

AAAI 2025

DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints

safety and robustness

snlp

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Recent advances in large language model assistants have made them indispensable, raising significant concerns over managing their safety. Automated red teaming offers a promising alternative to the labor-intensive and error-prone manual probing for vulnerabilities, providing more consistent and scalable safety evaluations. However, existing approaches often compromise diversity by focusing on maximizing attack success rate. Additionally, methods that decrease the cosine similarity from historical embeddings with semantic diversity rewards lead to novelty stagnation as history grows. To address these issues, we introduce DiveR-CT, which relaxes conventional constraints on the objective and semantic reward, granting greater freedom for the policy to enhance diversity. Our experiments demonstrate DiveR-CT's marked superiority over baselines by 1) generating data that perform better in various diversity metrics across different attack success rate levels, 2) better-enhancing resiliency in blue team models through safety tuning based on collected data, 3) allowing dynamic control of objective weights for reliable and controllable attack success rates, and 4) reducing susceptibility to reward overoptimization. Overall, our method provides an effective and efficient approach to LLM red teaming, accelerating real-world deployment.

technical paper

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.



To address catastrophic forgetting in Continual Relation Extraction (CRE), many current approaches rely on memory buffers to rehearse previously learned knowledge while acquiring new tasks. Recently, prompt-based methods have emerged as potent alternatives to rehearsal-based strategies, demonstrating strong empirical performance. However, upon analyzing existing prompt-based approaches for CRE, we identified several critical limitations, such as inaccurate prompt selection, inadequate mechanisms for mitigating forgetting in shared parameters, and suboptimal handling of cross-task and within-task variances. To overcome these challenges, we draw inspiration from the relationship between prefix tuning and mixture of experts, proposing a novel approach that employs a prompt pool for each task, capturing variations within each task while enhancing cross-task variances. Furthermore, we incorporate a generative model to consolidate prior knowledge within shared parameters, eliminating the need for explicit data storage. Extensive experiments validate the efficacy of our approach, demonstrating superior performance over state-of-the-art prompt-based and rehearsal-free methods in continual relation extraction.

Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective

The profusion of knowledge encoded in large language models (LLMs) and their ability to apply this knowledge zero-shot in a range of settings makes them promising candidates for use in decision-making. However, they are currently limited by their inability to provide outputs which can be faithfully explained and effectively contested to correct  mistakes. In this paper, we attempt to reconcile these strengths and weaknesses by introducing argumentative LLMs (aLLMs), a method for augmenting LLMs with argumentative reasoning. Concretely, aLLMs construct argumentation frameworks, which then serve as the basis for formal reasoning in support of decision-making. The interpretable nature of these argumentation frameworks and formal reasoning means that any decision made by aLLMs may be explained and contested. We evaluate aLLMs' performance experimentally in comparison with state-of-the-art techniques, in the context of the decision-making task of claim verification. We also define novel properties to characterise contestability and assess aLLMs formally in terms of these properties.

Argumentative Large Language Models for Explainable and Contestable Claim Verification

Video captioning automatically generates natural language phrases to explain the contents in video frames. When deploying captioning models in specialized domains, active learning can help to reduce the high annotation cost. However, existing active video caption methods rely on conventional uncertainty estimation, which can be  highly unreliable due to the novel challenge caused by the generative nature of the captioning process, which is much more complex than standard supervised learning tasks. Both our empirical evaluation and theoretical investigation reveal that entropy-based uncertainty estimation is in fact inflated, which will mislead active video sampling. Another challenge arises from the rich content of videos, as each video could be described in multiple ways. 
A single uncertainty score obtained from one possible caption does not capture the diversity induced by the rich content. To fill out this critical gap, we propose to identify multiple sources of uncertainty and perform novel hierarchical aggregation to integrate uncertainty from distinct sources mathematically. This innovates a holistic uncertainty metric to quantify the overall informativeness of video content for active sampling. The overall uncertainty is built upon conditional vacuity, which is a novel extension of the second-order uncertainty introduced along with the evidential learning framework to the generative setting, leading to more robust uncertainty estimation without inflation.  Both theoretical analysis and experimental evaluation are conducted to justify the effectiveness of the proposed framework for complex uncertainty estimation and interactive learning.

Hierarchical Multi-Source Uncertainty Aggregation for Interactive Video Captioning

Deep learning based dehazing networks trained on paired synthetic data have shown impressive performance, but they struggle with significant degradation in generalization ability on real-world hazy scenes. In this paper, we propose a lightweight Retinex-based Generative Adversarial \textbf{N}etwork (RetinexGAN) for real-world image dehazing using unpaired data. Our RetinexGAN consists of two stages: self-supervised pre-training and weakly-supervised fine-tuning. During the pre-training, we reduce the image dehazing task to an illumination-reflectance decomposition task based on the duality correlation between Retinex and dehazing. Specifically, a decomposition network named DecomNet is constructed to obtain an illumination and a reflectance, simultaneously. Moreover, a self-supervised learning strategy is developed to construct the connection between the preliminary dehazed result and the input hazy image, which constrains the solution space of DecomNet and accelerates training, leading to a more realistic dehazed result. In the fine-tuning stage, we develop a dual DTCWT-based attention module and embed it into the U-Net architecture to further improve the quality of preliminary result in the frequency domain. In addition, the adversarial learning is employed to constrain the relevance between the clean image and the final dehazed result in a weakly supervised manner, which can promote more natural performance. Extensive experiments on several real-world datasets demonstrate that our proposed RetinexGAN performs favorably over state-of-the-art dehazing methods in visual quality and quantitative evaluation.

Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network

Traffic prediction is critical for optimizing travel scheduling and enhancing public safety, yet the complex spatial and temporal dynamics within traffic data present significant challenges for accurate forecasting. In this paper, we introduce a novel model, the Spatiotemporal-aware Trend-Seasonality Decomposition Network (STDN). This model begins by constructing a dynamic graph structure to represent 
traffic flow and incorporates novel spatio-temporal embeddings to jointly capture global traffic dynamics. The  representations learned are further refined by a specially designed trend-seasonality decomposition module, which disentangles the trend-cyclical component and seasonal component for each traffic node at different times within the graph. These components are subsequently processed through an encoder-decoder network to generate the final predictions. Extensive experiments conducted on real-world traffic datasets demonstrate that STDN achieves superior performance with remarkable computation cost. Furthermore, we have released a new traffic dataset named JiNan, which features unique inner-city dynamics, thereby enriching the scenario comprehensiveness in traffic prediction evaluation. All source code and data are available.

Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting

The leaderboard of Large Language Models (LLMs) in mathematical tasks has been continuously updated. However, the majority of evaluations focus solely on the final results, neglecting the quality of the intermediate steps. This oversight can mask underlying problems, such as logical errors or unnecessary steps in the reasoning process. To measure reasoning beyond final-answer accuracy, we introduce ReasonEval, a new methodology for evaluating the quality of reasoning steps. ReasonEval employs $\textit{validity}$ and $\textit{redundancy}$ to characterize the reasoning quality, as well as accompanying LLMs to assess them automatically. We explore different design options for the LLM-based evaluators and empirically demonstrate that ReasonEval, when instantiated with base models possessing strong mathematical knowledge and trained with high-quality labeled data, consistently outperforms baseline methods in the meta-evaluation datasets. We also highlight the strong generalization capabilities of ReasonEval. By utilizing ReasonEval to evaluate LLMs specialized in math, we find that an increase in final-answer accuracy does not necessarily guarantee an improvement in the overall quality of the reasoning steps for challenging mathematical problems. Additionally, we observe that ReasonEval can play a significant role in data selection. We open-source the best-performing model, meta-evaluation script, and all evaluation results to facilitate future research.

Evaluating Mathematical Reasoning Beyond Accuracy

Given that AI systems are set to play a pivotal role in future decision-making processes, their trustworthiness and reliability are of critical concern. Due to their scale and complexity, modern AI systems resist direct interpretation, and alternative ways are needed to establish trust in those systems, and determine how well they align with human values. We argue that good measures of the information processing similarities between AI and humans, may be able to achieve these same ends.
While Representational alignment (RA) approaches measure similarity between the internal states of two systems, the associated data can be expensive and difficult to collect for human systems.
In contrast, Behavioural alignment (BA) comparisons are cheaper and easier, but questions remain as to their sensitivity and reliability.
We propose two new behavioural alignment metrics misclassification agreement which measures the similarity between the errors of two systems on the same instances, and class-level error similarity which measures the similarity between the error distributions of two systems. 
We show that our metrics correlate well with RA metrics, and provide complementary information to another BA metric, within a range of domains, and set the scene for a new approach to value alignment.

Measuring Error Alignment for Decision-Making Systems

Reinforcement Learning from Human Feedback (RLHF) aligns language models (LMs) with human values by training reward models (RMs) on binary preferences and using these RMs to fine-tune the base models. Despite its importance, the internal mechanisms of RLHF remain poorly understood. This paper introduces new metrics to evaluate RM effectiveness, focusing on feature imprint, feature resistance, and feature robustness. We categorize alignment datasets into target features (desired values) and spoiler features (undesired concepts). By regressing RM scores against these features, we quantify the extent to which RMs reward them -- feature imprint. We define alignment resistance as the proportion of the preference dataset where RMs fail to match human preferences, and we assess alignment robustness by analyzing RM responses to slightly perturbed texts. Our experiments, utilizing open-source components like the Anthropic/hh-rlhf preference dataset and OpenAssistant RMs, reveal significant imprints of target features and a notable sensitivity to spoiler features. We observed a 26% resistance incidence in portions of the dataset where LM labelers disagreed with human preferences. We also find that misalignment stems from confusing entries in the alignment dataset. These findings underscore the importance of scrutinizing both RMs and alignment datasets for a deeper understanding of value alignment.

SEAL: Systematic Error Analysis for Value ALignment

The increasing parameters and expansive dataset of large lan- guage models (LLMs) highlight the urgent demand for a technical solution to audit the underlying privacy risks and copyright issues associated with LLMs. Existing studies have partially addressed this need through an exploration of the pre-training data detection problem, which is an instance of a membership inference attack (MIA). This problem involves determining whether a given piece of text has been used during the pre-training phase of the target LLM. Although existing methods have designed various sophisticated MIA score functions to achieve considerable detection performance in pre-trained LLMs, how to achieve high-confidence detection and how to perform MIA on aligned LLMs remain challenging. In this paper, we propose MIA-Tuner, a novel instruction-based MIA method, which instructs LLMs themselves to serve as a more precise pre-training data detector internally, rather than design an external MIA score function. Furthermore, we design two instruction-based safeguards to respectively mitigate the privacy risks brought by the existing methods and MIA-Tuner. To comprehensively evaluate the most recent state-of-the-art LLMs, we collect a more up-to-date MIA benchmark dataset, named WIKIMIA-24, to replace the widely adopted benchmark WIKIMIA. We conduct extensive experiments across various aligned and unaligned LLMs over the two benchmark datasets. The results demonstrate that MIA-Tuner increases the AUC of MIAs from 0.7 to a significantly high level of 0.9.

MIA-Tuner: Adapting Large Language Models as Pre-training Text Detector

With the malicious use and dissemination of multi-modal deepfake videos, researchers start to investigate multi-modal deepfake detection. Unfortunately, most of the existing methods tune all parameters with limited speech video datasets, only exploit audio-visual inconsistencies and are trained under coarse-grained consistency supervision, which hinders the generalization ability in practical scenarios. In this paper, we propose the first multi-task audio-visual prompt learning method for multi-modal deepfake video detection, by exploiting multiple foundation models. Specifically, we construct a two-stream multi-task learning architecture and propose sequential visual prompts and short-time audio prompts to extract multi-modal features, which can be aligned at the frame level and utilized in subsequent fine-grained feature matching and fusion. Due to the natural alignment of lip movement and spoken phoneme in real data, we propose a frame-level cross-modal feature matching loss function to learn the fine-grained audio-visual consistency. Comprehensive experiments demonstrate the effectiveness and generalization ability of our method against the state-of-the-art methods.

Downloads

Next from AAAI 2025

Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective

Similar lecture

Nullspace Disentanglement for Red Teaming Language Models