United States

Large Vision-Language Models (LVLMs) often fail to align with human preferences, leading to issues like generating misleading content without proper visual context (also known as \textit{hallucination}).
A promising solution to address this problem is using human-preference alignment techniques, such as best-of-$n$ sampling and reinforcement learning.
However, these techniques face the difficulty arising from the scarcity of visual preference data required to train a visual reward model (VRM).
In this work, we continue the line of research.
We present a $\textbf{Ro}$bust $\textbf{V}$isual $\textbf{R}$eward$ \textbf{M}$odel (RoVRM) which improves human-preference alignment for LVLMs.
RoVRM leverages auxiliary textual preference data through a three-phase progressive training approach and optimal transport-based preference data selection to effectively mitigate the scarcity of visual preference data.
We experiment with RoVRM on the commonly used vision-language tasks based on the LLaVA-1.5-7B and -13B models.
Experimental results demonstrate that RoVRM consistently outperforms traditional VRMs. 
Furthermore, our three-phase progressive training and preference data selection approaches can yield consistent performance gains over ranking-based alignment techniques, such as direct preference optimization.

AAAI 2025

RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data

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.



Despite being empowered with alignment mechanisms, large language models (LLMs) are increasingly vulnerable to emerging jailbreak attacks that can compromise their alignment mechanisms. 
This vulnerability poses significant risks to the real-world applications.
Existing work faces challenges in both training efficiency and generalization capabilities (i.e., Reinforcement Learning from Human Feedback and Red-Teaming). 
Developing effective strategies to enable LLMs to resist continuously evolving jailbreak attempts represents a significant challenge. 
To address this challenge, we propose a novel defensive paradigm called GuidelineLLM, which assists LLMs in recognizing queries that may have harmful content. 
Before LLMs respond to a query, GuidelineLLM first identifies potential risks associated with the query, summarizes these risks into guideline suggestions, and then feeds these guidelines to the responding LLMs.
Importantly, our approach eliminates the necessity for additional safety fine-tuning of the LLMs themselves; only the GuidelineLLM requires fine-tuning. This characteristic enhances the general applicability of GuidelineLLM across various LLMs. 
Experimental results demonstrate that GuidelineLLM can significantly reduce the attack success rate (ASR) against the LLMs (an average reduction of 34.17\% ASR) while maintaining the helpfulness of the LLMs in handling benign queries. 
Code is available at $\texttt{Anonymous}$.

Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM

Time series forecasting is an important application in various domains such as energy management, traffic planning, financial markets, meteorology, and medicine. However, real-time series data often present intricate temporal variability and sharp fluctuations, which pose significant challenges for time series forecasting. Previous models that rely on 1D time series representations usually struggle with complex temporal variations. To address the limitations of 1D time series, this study introduces the Times2D method that transforms the 1D time series into 2D space. Times2D consists of three main parts: first, a Periodic Decomposition Block (PDB) that captures temporal variations within a period and between the same periods by converting the time series into a 2D tensor in the frequency domain. Second, the First and Second Derivative Heatmaps (FSDH) capture sharp changes and turning points, respectively. Finally, an Aggregation Forecasting Block (AFB) integrates the 2D tensors from PDB and FSDH for accurate forecasting. This 2D transformation enables the utilization of 2D convolutional operations to effectively capture long and short characteristics of the time series. Comprehensive experimental results across large-scale data in the literature demonstrate that the proposed Times2D model achieves state-of-the-art performance in both short-term and long-term forecasting.

Times2D: Multi-Period Decomposition and Derivative Mapping for General Time Series Forecasting

Substitute training-based data-free black-box attacks pose a significant threat to enterprise-deployed models. These attacks use a generator to synthesize data and query APIs, then train a substitute model to approximate the target model's decision boundary based on the returned results. However, existing attack methods often struggle to produce sufficiently diverse data, particularly for complex target models and extensive target data domains, severely limiting their practical application. To address this gap, we design domain-augmented learning to improve the quality of the synthetic data domain (SDD) generated by the generator from two perspectives. Specifically, (1) To broaden the SDD's coverage, we introduce textual semantic embeddings into the generator for the first time. (2) For enhancing the SDD's discretization, we propose a competitive optimization strategy that forces the generator to self-compete, along with heterogeneity excitation to overcome the constraints of information entropy on diversity. Comprehensive experiments demonstrate that our method is more effective. In non-targeted attacks on the CIFAR-10 and Tiny-ImageNet datasets, our method outperforms the state-of-the-art by 14\% and 7\% in attack success rate, respectively.

Power of Diversity: Enhancing Data-Free Black-Box Attack with Domain-Augmented Learning

To address the challenges of imbalanced multi-class datasets typically used for rare event detection in critical cyber-physical systems, we propose an optimal, efficient, and adaptable mixed integer programming (MIP) ensemble weighting scheme. Our approach leverages the diverse capabilities of the classifier ensemble on a granular per class basis, while optimizing the weights of classifier-class pairs using elastic net regularization for improved robustness and generalization. Additionally, it seamlessly and optimally selects a predefined number of classifiers from a given set. We evaluate and compare our MIP-based method against six well-established weighting schemes, using representative datasets and suitable metrics, under various ensemble sizes. The experimental results reveal that MIP outperforms all existing approaches, achieving an improvement in balanced accuracy ranging from 0.99% to 7.31%, with an overall average of 4.53% across all datasets and ensemble sizes. Furthermore, it attains an overall average increase of 4.63%, 4.60%, and 4.61% in macro-averaged precision, recall, and F1-score, respectively, while maintaining computational efficiency.

Rare Event Detection in Imbalanced Multi-Class Datasets Using an Optimal MIP-Based Ensemble Weighting Approach

Neural Radiance Fields (NeRF) often struggle with reconstructing and rendering highly reflective scenes. Recent advancements have developed various reflection-aware appearance models to enhance NeRF's capability to render specular reflections. However, the robust reconstruction of highly reflective scenes is still hindered by the inherent shape ambiguity on specular surfaces. Existing methods typically rely on additional geometry priors to regularize the shape prediction, but this can lead to oversmoothed geometry in complex scenes. Observing the critical role of surface normals in parameterizing reflections, we introduce a transmittance-gradient-based normal estimation technique that remains robust even under ambiguous shape conditions. Furthermore, we propose a dual activated densities module that effectively accommodates the disparity between smooth surface normals and sharp object boundaries. Combined with a reflection-aware appearance model, our proposed method achieves robust reconstruction and high-fidelity rendering of scenes featuring both highly specular reflections and intricate geometric structures. Extensive experiments demonstrate that our method outperforms existing state-of-the-art methods on various datasets. The code will be publicly released if the article is accepted.

Normal-NeRF: Ambiguity-Robust Normal Estimation for Highly Reflective Scenes

Image restoration, which reconstructs high-quality images from degraded ones, has been extensively studied using various datasets in computer vision.
However, existing datasets are based on images captured with professional equipment such as digital cameras and action cameras.
This limits their applicability to everyday situations, as they focus on specialized scenarios.
This paper focuses on the degradation of smartphone camera lenses and proposes a dataset of degraded images commonly encountered in daily smartphone use.
To simulate realistic image degradation, we employed a novel approach of applying physical degradation directly to the source.
Furthermore, through experiments with existing image restoration models, we demonstrated that our dataset is extremely challenging.
\fi
Smartphone cameras are ubiquitous in daily life, yet their performance can be severely impacted by dirty lenses, leading to degraded image quality. 
This issue is often overlooked in image restoration research, which typically focuses on ideal or controlled conditions. 
To address this gap, we introduce SIDL (Smartphone Images with Dirty Lenses), a novel dataset designed specifically for the task of restoring images captured through contaminated smartphone lenses. 
SIDL contains a diverse collection of real-world images taken under various lighting conditions and environments. 
These images feature a wide range of lens contaminants, including water drops, fingerprints, and dust. 
Each contaminated image is paired with a clean reference image, enabling supervised learning approaches for restoration tasks.

To evaluate the challenge posed by SIDL, we trained and compared the performance of various state-of-the-art restoration models on this dataset. 
Our results indicate that while these models can achieve some level of restoration, the diverse and realistic nature of the contaminations in SIDL presents significant challenges that are not adequately addressed by existing methods. 
This highlights the need for more robust and adaptable image restoration techniques.

SIDL: A Real-World Dataset for Restoring Smartphone Images with Dirty Lenses

In this study, we focus on $\textit{heterogeneous}$ knowledge transfer across entirely different model architectures, tasks, and modalities. Existing knowledge transfer methods ($\textit{e.g.}$, backbone sharing, knowledge distillation) often hinge on shared elements within model structures or task-specific features/labels, limiting transfers to complex model types or tasks. To overcome these challenges, we present MergeNet, which learns to bridge the gap of parameter spaces of heterogeneous models, facilitating the direct interaction, extraction, and application of knowledge within these parameter spaces. The core mechanism of MergeNet lies in the parameter adapter, which operates by querying the source model's low-rank parameters and adeptly learning to identify and map parameters into the target model. MergeNet is learned alongside both models, allowing our framework to dynamically transfer and adapt knowledge relevant to the current stage, including the training trajectory knowledge of the source model. Extensive experiments on heterogeneous knowledge transfer demonstrate significant improvements in challenging settings, where representative approaches may falter or prove less applicable.

MergeNet: Knowledge Migration across Heterogeneous Models,  Tasks,  and Modalities

In Hotelling's model of spatial competition, a unit mass of voters is distributed in the interval $[0,1]$ (with their location corresponding to their political persuasion), and each of $m$ candidates selects as a strategy their distinct position in this interval. Each voter votes for the nearest candidate, and candidates choose their strategy to maximize their votes. It is known that if there are more than two candidates, equilibria may not exist in this model. It was unknown, however, how close to an equilibrium one could get. Our work studies approximate equilibria in this model, where a strategy profile is an (additive) $\epsilon$-equilibria if no candidate can increase their votes by $\epsilon$, and provides tight or nearly-tight bounds on the approximation $\epsilon$ achievable.

We show that for 3 candidates, for any distribution of the voters, $\epsilon \ge 1/12$. Thus, somewhat surprisingly, for any distribution of the voters and any strategy profile of the candidates, at least $1/12$th of the total votes is always left ``on the table.'' Extending this, we show that in the worst case, there exist voter distributions for which $\epsilon \ge 1/6$, and this is tight: one can always compute a $1/6$-approximate equilibria in polynomial time. We then study the general case of $m$ candidates, and show that as $m$ grows large, we get closer to an exact equilibrium: one can always obtain a $1/(m+1)$-approximate equilibria in polynomial time. We show this bound is asymptotically tight, by giving voter distributions for which $\epsilon \ge 1/(m+3)$.

Nearly Tight Bounds on Approximate Equilibria in Spatial Competition on the Line

Recent advances in text-to-image diffusion models have spurred research on personalization, particularly focusing on customized image synthesis of subjects within reference images. Although existing personalization methods can modify the subjects' positions or personalize multiple subjects simultaneously, they often struggle with altering the behaviors of subjects or their dynamic interactions. This challenge arises from overfitting to reference images, which becomes more problematic when only a single reference image is available. To address these challenges, we propose DynASyn, an effective multi-subject personalization method that works from a single reference image. DynASyn preserves subject identity during the personalization process by aligning concept-based priors with subject appearances and actions. This alignment is achieved by regularizing the attention maps between the subject token and images through concept-based priors. Furthermore, we introduce concept-based prompt-and-image augmentation to enhance the trade-off between identity preservation and action diversity. We also propose an SDE-based editing technique guided by augmented prompts to generate diverse appearances and actions while maintaining identity consistency in the augmented images. Experiments demonstrate that DynASyn is capable of synthesizing highly realistic images of subjects in novel contexts with dynamic interactions with their surroundings and outperforms baseline methods in both quantitative and qualitative aspects.

DynASyn: Multi-Subject Personalization Enabling Dynamic Action Synthesis

LLM have achieved success in many fields but still troubled by problematic content in the training corpora. LLM unlearning aims at reducing their influence and avoid undesirable behaviours. However, existing unlearning methods remain vulnerable to adversarial queries and the unlearned knowledge resurfaces after the manually designed attack queries. As part of a red-team effort to proactively assess the vulnerabilities of unlearned models, we design \textbf{D}ynamic \textbf{U}nlearning \textbf{A}ttack  (\textbf{DUA}), a dynamic and automated framework to attack these models and evaluate their robustness. It optimizes adversarial suffixes to reintroduce the unlearned knowledge in various scenarios. We find that unlearned knowledge can be recovered in $55.2\%$  of the questions, even without revealing the unlearned model's parameters. In response to this vulnerability, we propose \textbf{L}atent \textbf{A}dversarial \textbf{U}nlearning (\textbf{LAU}), a universal framework that effectively enhances the robustness of the unlearned process. It formulates the unlearning process as a min-max optimization problem and resolves it through two stages: an attack stage, where perturbation vectors are trained and added to the latent space of LLMs to recover the unlearned knowledge, and a defense stage, where previously trained perturbation vectors are used to enhance unlearned model's robustness. With our LAU framework, we obtain two robust unlearning methods, \textbf{AdvGA} and \textbf{AdvNPO}. We conduct extensive experiments across multiple unlearning benchmarks and various models, and demonstrate that they improve the unlearning effectiveness by over $53.5\%$, cause only less than a $11.6\%$ reduction in neighboring knowledge, and have almost no impact on the model's general capabilities.

RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data

Downloads

Next from AAAI 2025

Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data

.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

Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads