United States

Safety alignment is indispensable for Large language models (LLMs) to defend threats from malicious instructions. However, recent researches reveal safety-aligned LLMs prone to reject benign queries due to the exaggerated safety issue, limiting their helpfulness. In this paper, we propose a $\underline{\textbf{S}}$afety-$\underline{\textbf{C}}$onscious $\underline{\textbf{A}}$ctivatio$\underline{\textbf{n}}$ $\underline{\textbf{S}}$teering ($\textbf{SCANS}$) method to mitigate the exaggerated safety concerns in aligned LLMs. First, SCANS extracts the refusal steering vectors within the activation space and utilizes vocabulary projection to anchor some specific safety-critical layers which influence model refusal behavior. Second, by tracking the hidden state transition, SCANS identifies the steering direction and steers the model behavior accordingly, achieving a balance between exaggerated safety and adequate safety. Experiments show that SCANS achieves new state-of-the-art performance on XSTest and OKTest benchmarks, without impairing their defense capability against harmful queries and maintaining almost unchanged model capability.

AAAI 2025

SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

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 ACL 2025". 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 need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

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.



Deep imbalanced regression (DIR), where the target values have a highly skewed distribution and are also continuous, is an intriguing yet under-explored problem in machine learning. 
  While recent works have already shown that incorporating various classification-based regularizers can produce enhanced outcomes, the role of classification remains elusive in DIR.
  Moreover, such regularizers (e.g., contrastive penalties) merely focus on learning discriminative features of data, which inevitably results in ignorance of either continuity or similarity across the data.
  To address these issues, we first bridge the connection between the objectives of DIR and classification from a Bayesian perspective. 
  Consequently, this motivates us to decompose the objective of DIR into a combination of classification and regression tasks, which naturally guides us toward a divide-and-conquer manner to solve the DIR problem.
  Specifically, by aggregating the data at nearby labels into the same groups, we introduce an ordinal group-aware contrastive learning loss along with a multi-experts regressor to tackle the different groups of data thereby maintaining the data continuity.
  Meanwhile, considering the similarity between the groups, we also propose a symmetric descending soft labeling strategy to exploit the intrinsic similarity across the data, which allows classification to facilitate regression more effectively.
  Extensive experiments on real-world datasets also validate the effectiveness of our method.

Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression

Majority illusion is a phenomenon in social networks wherein the decision by the majority of the network is not the same as one's personal social circle's majority, leading to an incorrect perception of the majority in a large network. For the model of [Grandi et al. 2023], we present polynomial-time algorithms which completely eliminate majority illusion by altering as few connections in the network as possible. Eliminating majority illusion ensures each neighbourhood in the network has at least a $\frac12$-fraction of the majority winner. This result is surprising as [Grandi et al. 2023] prove that partially eliminating majority illusion is NP-hard. We generalize the majority illusion problem to an arbitrary fraction $p$ and show that the problem of ensuring all neighbourhoods in the network contain at least a $p$-fraction of nodes consistent with a given preference is NP-hard, for nearly all values of $p$.

Eliminating Majority Illusion Is Easy

Malicious users attempt to replicate commercial models functionally at low cost by training a clone model with query responses. It is challenging to timely prevent such model-stealing attacks to achieve strong protection and maintain utility. In this paper, we propose a novel non-parametric detector called Account-aware Distribution Discrepancy (ADD) to recognize queries from malicious users by leveraging account-wise local dependency. We formulate each class as a Multivariate Normal distribution (MVN) in the feature space and measure the malicious score as the sum of weighted class-wise distribution discrepancy. The ADD detector is combined with random-based prediction poisoning to yield a plug-and-play defense module named D-ADD for image classification models. Results of extensive experimental studies show that D-ADD achieves strong defense against different types of attacks with little interference in serving benign users for both soft and hard-label settings.

Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy

Adaptive learning, also known as adaptive teaching, relies on learning path recommendations that sequentially suggest personalized learning items (such as lectures and exercises) to meet the unique needs of each learner. Despite the extensive research in this field, previous approaches have primarily modeled the interaction sequences between learners and items using simple indexing, leading to three issues: (1) The utilization of information from both learners and items is not sufficient. For instance, these models are unable to leverage the semantic information contained within the textual content of the items. (2) Models need to be retrained on different datasets separately, which makes it difficult to adapt to the continuously expanding item pool in online educational scenarios. 
(3) The existing recommendation paradigm based on trained reinforcement learning frameworks, suffers from unstable recommendation performance in sparse learning logs.
To address these challenges, we propose a generalized \textbf{Gen}erative Agent for \textbf{A}daptive \textbf{L}earning (\textbf{GenAL}), which integrates educational tools with LLMs' semantic understanding to enable effective and generalizable learning path recommendations across diverse data distributions. Specifically, our framework consists of two components: the Global Thinking Agent, which updates the learner profile and reflects on recommendation outcomes based on the learner's historical learning records. The other is the Local Teaching Agent, which recommends items using educational prior knowledge. Leveraging the LLM's robust semantic understanding, our framework does not rely on item indexing but instead extracts relevant information from the textual content. We evaluated our approach on three real-world datasets, and the experimental results demonstrate that our GenAL not only consistently outperforms all baselines but also exhibits strong generalization ability.

GenAL: Generative Agent for Adaptive Learning

Hallucination continues to be one of the most critical challenges in the institutional adoption journey of Large Language Models (LLMs). While prior studies have primarily focused on the post-generation analysis and refinement of outputs, this paper centers on the effectiveness of queries in eliciting accurate responses from LLMs. We present HalluciBot, a model that estimates the query's propensity to hallucinate before generation, without invoking any LLMs during inference. HalluciBot can serve as a proxy reward model for query rewriting, offering a general framework to estimate query quality based on accuracy and consensus. In essence, HalluciBot investigates how poorly constructed queries can lead to erroneous outputs - moreover, by employing query rewriting guided by HalluciBot's empirical estimates, we demonstrate that 95.7% output accuracy can be achieved for Multiple Choice questions. The training procedure for HalluciBot consists of perturbing 369,837 queries $n$ times, employing $n+1$ independent LLM agents, sampling an output from each query, conducting a Multi-Agent Monte Carlo Simulation on the sampled outputs, and training an encoder classifier. The idea of perturbation is the outcome of our ablation studies that measures the increase in output diversity (+12.5 agreement spread) by perturbing a query in lexically different but semantically similar ways. Therefore, HalluciBot paves the way to ratiocinate (76.0% test F1 score, 46.6% in saved computation on hallucinatory queries), rewrite (+30.2% positive class transition from hallucinatory to non-hallucinatory), rank (+50.6% positive class transition from hallucinatory to non-hallucinatory), and route queries to effective pipelines.

Is There No Such Thing as a Bad Question? H4R: HalluciBot for Ratiocination, Rewriting, Ranking, and Routing

We propose an energy amplification technique to address the issue that existing models easily overlook low-energy components in time series forecasting. This technique comprises an energy amplification block and an energy restoration block. The energy amplification block enhances the energy of low-energy components to improve the model's learning efficiency for these components, while the energy restoration block returns the energy to its original level. Moreover, considering that the energy-amplified data typically displays two distinct energy peaks in the frequency spectrum, we integrate the energy amplification technique with a seasonal-trend forecaster to model the temporal relationships of these two peaks independently, serving as the backbone for our proposed model, Amplifier. Additionally, we propose a  semi-channel interaction temporal relationship enhancement block for Amplifier, which enhances the model's ability to capture temporal relationships from the perspective of the commonality and specificity of each channel in the data. Extensive experiments on eight time series forecasting benchmarks consistently demonstrate our model's superiority in both effectiveness and efficiency compared to state-of-the-art methods.

Amplifier: Bringing Attention to Neglected Low-Energy Components in Time Series Forecasting

Identifying and linking the same users across different social platforms is crucial for understanding user behavior and preferences. However, cross-domain datasets exhibit diverse characteristics, such as varying check-in frequencies, significant disparities in data precision, and distinct distributions. Existing trajectory representations rely on recurrent neural network, which fails to dynamically learn multi-dimensional feature relations and capture high-order associations. Furthermore, current methods for integrating trajectory information fails to capture the complex relations and dynamic variations among cross-domain mobility trajectories. To this end, we propose the Hierarchical Spatio-Temporal Enhanced Attention Hypergraph Network (StarNet). This model dynamically regulates the multi-dimensional features of trajectories through a locally enhanced spatiotemporal graph neural network. Meanwhile, StarNet employs a hypergraph network enhanced by a global spatiotemporal to capture high-order associations between cross-domain trajectories. The fusion enhancement association integrates local and global information, which enables this model to link user identities. Extensive experiments on two well-known LBSN cross-domain datasets reveal that StarNet outperforms state-of-the-art baselines in the accuracy of user identity linkage.

Cross-Domain Trajectory Association Based on Hierarchical Spatiotemporal Enhanced Attention Hypergraph

The Large Vision-Language Model (LVLM) integrates computer vision and natural language processing techniques, offering substantial application potential. However, these models demand extensive resources during inference. Adaptive attention techniques can dynamically reduce computational redundancy and thus improve efficiency. Although current adaptive attention methods significantly reduce the memory requirements of Transformer-based language models, they are not tailored for LVLMs. We observe that LVLMs generate responses from both remote image tokens and local text tokens, and different modalities have different attention patterns. This observation inspires us to manage the attention for each modality separately. Specifically, for visual input, we store the cache of potentially useful information but only compute the most critical parts. For language input, we care more about local information. Based on our observation and analysis of vision-language attention patterns, we develop A-VL, a plug-and-play adaptive attention tailored for LVLM inference. Extensive evaluations on three vision-language tasks and five datasets show the effectiveness of our designs. Our approach A-VL outperforms existing adaptive attention methods in reducing memory usage and computational load without compromising performance.

A-VL: Adaptive Attention for Large Vision-Language Models

Vehicle-to-infrastructure (V2I) cooperative perception systems can enhance the sensing abilities of autonomous vehicles. Existing V2I solutions often consider LiDARs devices instead of cameras, the most prevalent sensors with low cost and wide installation. In addition, a major challenge that has been underexplored is the time asynchrony between image frames from different sources. This asynchrony arises because of clock differences, varying times involved in data processing and transmission, causing uncertain delays that complicate data alignment and potentially reduce perception accuracy. We propose BEVSync, a camera-based V2I cooperative perception system that adaptively aligns frames from the ego-vehicle and infrastructure by compensating for motion deviations. Specifically, we develop an extractor-compensator model to extract and predict perceptual features using historical frames, thereby smoothing out the data misalignment. Experiments on the real-world dataset DAIR-V2X show that our approach surpasses existing methods in terms of performance and robustness.

BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays

Existing ToM benchmarks for LLMs diverge from real-world scenarios in three aspects: 1) they assess a limited range of mental states, such as beliefs, 2) false beliefs are not comprehensively explored, and 3) the diverse personality traits of characters are overlooked.
To address these challenges, we introduce ToMATO, a new ToM benchmark that utilizes LLM-LLM conversations featuring information asymmetry.
ToMATO is formulated as multiple-choice QA over conversations generated by LLMs.
By employing a prompting method that requires LLMs to verbalize their thoughts before each utterance, we capture both first- and second-order mental states across five categories: beliefs, intentions, desires, emotions, and knowledge.
These verbalized thoughts serve as answers to questions designed to assess the mental states of characters within conversations.
Furthermore, the information asymmetry introduced by concealing thoughts from others induces the generation of false beliefs about various mental states.
Assigning distinct personality traits to LLM agents further diversifies both utterances and thoughts.
ToMATO consists of 5.4k multiple-choice questions, 753 conversations, and 15 personality trait patterns.
Our analysis shows that this dataset construction approach frequently generates false beliefs due to the information asymmetry between LLM agents, and effectively reflect diverse personalities.
We evaluate nine LLMs on ToMATO and find that even GPT-4o mini lags behind human performance, especially in understanding false beliefs, and lacks robustness to various personality traits.

Downloads

Next from AAAI 2025

Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression

Similar lecture

LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES