Singapore

Multimodal large reasoning models (MLRMs) have advanced visual-textual integration, enabling sophisticated human-AI interaction. While prior work has exposed MLRMs to visual jailbreaks, it remains underexplored how their reasoning capabilities reshape the security landscape under adversarial inputs. To fill this gap, we conduct a systematic security assessment of MLRMs and uncover a security-reasoning paradox: 
although deeper reasoning boosts cross‑modal risk recognition, it also creates cognitive blind spots that adversaries can exploit. 
We observe that MLRMs oriented toward human-centric service are highly susceptible to users&#39; emotional cues during the deep-thinking stage, often overriding safety protocols or built‑in safety checks under high emotional intensity.
Inspired by this key insight, we propose \textbf{EmoAgent}, an autonomous adversarial emotion-agent framework that orchestrates exaggerated affective prompts to hijack reasoning pathways.
Even when visual risks are correctly identified, models can still produce harmful completions through emotional misalignment. We further identify persistent high-risk failure modes in transparent deep-thinking scenarios, such as MLRMs generating harmful reasoning masked behind seemingly safe responses. These failures expose misalignments between internal inference and surface-level behavior, eluding existing content-based safeguards. To quantify these risks, we introduce three metrics: (1) \emph{Risk-Reasoning Stealth Score (RRSS)} for harmful reasoning beneath benign outputs; (2) \emph{Risk-Visual Neglect Rate (RVNR)} for unsafe completions despite visual risk recognition; and (3) \emph{Refusal Attitude Inconsistency (RAIC)} for evaluating refusal unstability under prompt variants.
Extensive experiments on advanced MLRMs demonstrate the effectiveness of EmoAgent and reveal deeper emotional cognitive misalignments in model safety behavior.
\textbf{ Warning: This paper contains examples that may be offensive or harmful.}

AAAI 2026

The Emotional Baby Is Truly Deadly: Does Your Multimodal Large Reasoning Model Have Emotional Flattery Towards Humans?

multimodal large reasoning models

jailbreaks

ai security

## 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.

Multimodal large reasoning models (MLRMs) have advanced visual-textual integration, enabling sophisticated human-AI interaction. While prior work has exposed MLRMs to visual jailbreaks, it remains underexplored how their reasoning capabilities reshape the security landscape under adversarial inputs. To fill this gap, we conduct a systematic security assessment of MLRMs and uncover a security-reasoning paradox: 
although deeper reasoning boosts cross‑modal risk recognition, it also creates cognitive blind spots that adversaries can exploit. 
We observe that MLRMs oriented toward human-centric service are highly susceptible to users' emotional cues during the deep-thinking stage, often overriding safety protocols or built‑in safety checks under high emotional intensity.
Inspired by this key insight, we propose \textbf{EmoAgent}, an autonomous adversarial emotion-agent framework that orchestrates exaggerated affective prompts to hijack reasoning pathways.
Even when visual risks are correctly identified, models can still produce harmful completions through emotional misalignment. We further identify persistent high-risk failure modes in transparent deep-thinking scenarios, such as MLRMs generating harmful reasoning masked behind seemingly safe responses. These failures expose misalignments between internal inference and surface-level behavior, eluding existing content-based safeguards. To quantify these risks, we introduce three metrics: (1) \emph{Risk-Reasoning Stealth Score (RRSS)} for harmful reasoning beneath benign outputs; (2) \emph{Risk-Visual Neglect Rate (RVNR)} for unsafe completions despite visual risk recognition; and (3) \emph{Refusal Attitude Inconsistency (RAIC)} for evaluating refusal unstability under prompt variants.
Extensive experiments on advanced MLRMs demonstrate the effectiveness of EmoAgent and reveal deeper emotional cognitive misalignments in model safety behavior.
\textbf{ Warning: This paper contains examples that may be offensive or harmful.}

poster

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.

To address the limitations of Transformer decoders in capturing edge details, recognizing local textures and modeling spatial continuity, this paper proposes a novel decoder framework specifically designed for medical image segmentation, comprising three core modules. First, the Adaptive Cross-Fusion Attention (ACFA) module integrates channel feature enhancement with spatial attention mechanisms and introduces learnable guidance in three directions (planar, horizontal, and vertical) to enhance responsiveness to key regions and structural orientations. Second, the Triple Feature Fusion Attention (TFFA) module fuses features from Spatial, Fourier and Wavelet domains, achieving joint frequency-spatial representation that strengthens global dependency and structural modeling while preserving local information such as edges and textures, making it particularly effective in complex and blurred boundary scenarios. Finally, the Structural-aware Multi-scale Masking Module (SMMM) optimizes the skip connections between encoder and decoder by leveraging multi-scale context and structural saliency filtering, effectively reducing feature redundancy and improving semantic interaction quality. Working synergistically, these modules not only address the shortcomings of traditional decoders but also significantly enhance performance in high-precision tasks such as tumor segmentation and organ boundary extraction, improving both segmentation accuracy and model generalization. Experimental results demonstrate that this framework provides an efficient and practical solution for medical image segmentation.

Decoding with Structured Awareness: Integrating Directional, Frequency-Spatial, and Structural Attention for Medical Image Segmentation

Optical satellites, with their diverse band layouts and ground sampling distances, supply indispensable evidence for tasks ranging from ecosystem surveillance to emergency response. However, significant discrepancies in band composition and spatial resolution across different optical sensors present major challenges for existing Remote Sensing Foundation Models (RSFMs). These models are typically pretrained on fixed band configurations and resolutions, making them vulnerable to real-world scenarios involving missing bands, cross-sensor fusion, and unseen spatial scales, thereby limiting their generalization and practical deployment.
To address these limitations, we propose Any-Optical-Model ($AOM$), the first universal RSFM explicitly designed to accommodate arbitrary band compositions, sensor types, and resolution scales. To preserve distinctive spectral characteristics even when bands are missing or newly introduced, $AOM$ introduces a spectrum-independent tokenizer that assigns each channel a dedicated band embedding, enabling explicit encoding of spectral identity. To effectively capture texture and contextual patterns from sub-meter to hundred-meter imagery, we design a multi-scale adaptive patch embedding mechanism that dynamically modulates the receptive field. Furthermore, to maintain global semantic consistency across varying resolutions, $AOM$ incorporates a multi-scale semantic alignment mechanism alongside a channel-wise self-supervised masking and reconstruction pretraining strategy that jointly models spectral-spatial relationships.
Extensive experiments on over 10 public datasets, including those from Sentinel-2, Landsat, and HLS, demonstrate that $AOM$ consistently achieves state-of-the-art (SOTA) performance under challenging conditions such as band-missing, cross-sensor, and cross-resolution settings. These results highlight $AOM$ as a crucial step toward building truly general-purpose RSFMs.

Any-Optical-Model: A Universal Foundation Model for Optical Remote Sensing

Multimodal emotion recognition plays a crucial role in enhancing the intelligence of human-computer interaction and emotional understanding. However, conventional approaches face challenges such as scarcity of annotated data, significant modality heterogeneity, and temporal misalignment. To address these issues, we propose DHCM-CACL, a novel self-supervised emotion recognition framework integrating EEG and facial expressions. During the pre-training phase, we propose a Dynamic Hierarchical Cross-modal Mamba module (DHCM), which models long-term dependencies through dynamic state matrices, incorporates forgetting gates for noise suppression, and constructs a hierarchical cross-modal interaction structure, effectively achieving cross-modal temporal alignment and mitigating modality heterogeneity. Subsequently, we propose a Confidence-Adaptive Contrastive Learning module (CACL) that dynamically adjusts sample weights using gated confidence signals derived from DHCM to compute loss, prioritizing reliable samples while suppressing noisy instances through adaptive weighting, thereby enhancing representation reliability and low-sample generalization. During the fine-tuning phase, we integrate a cross-modal attention gating mechanism to reinforce temporal associations and adopt an evidence-aware joint optimization objective, providing probabilistic credibility outputs for emotion prediction. Experimental results on the DEAP and MAHNOB-HCI datasets demonstrate that our approach achieves state-of-the-art performance in emotion classification under both subject-dependent and subject-independent settings.

DHCM-CACL: Dynamic Hierarchical Cross-modal Mamba with Confidence-Adaptive Contrastive Learning for Multimodal Emotion Recognition

Recently, End-to-End Speech Translation (E2E-ST) methods leveraging large language models (LLMs) have demonstrated strong generalization capabilities and excellent scalability by integrating pre-trained speech encoders with LLMs, where Low-Rank Adaptation (LoRA) is commonly used for parameter-efficient fine-tuning to reduce training costs. However, LoRA's low-rank assumption often fails in multilingual tasks, as the inherent complexity of cross-lingual semantic relationships and syntactic variations exceeds the representational capacity of low-rank matrices. This leads to parameter conflicts across languages, resulting in suboptimal performance. To address this issue, we propose Mixture of Low-Rank Adaptations (MoLoRA), which integrates the Mixture of Experts (MoE) mechanism with LoRA. MoLoRA effectively enhances the model's expressive capacity while maintaining parameter efficiency during training. Specifically, we treat multiple LoRA modules as low-rank experts and introduce a routing mechanism to dynamically activate language-specific experts. Additionally, shared experts are incorporated and consistently activated to model cross-lingual general knowledge. Furthermore, to enhance the robustness and accuracy of speech representations, we propose a Multi-Granularity Representation Fusion module (MGRF). This module mitigates local distortions in frame-level speech representations caused by noise by fusing frame-level and sentence-level features, thereby providing the LLM with more accurate high-level semantic information. We conduct multilingual experiments on the MuST-C and CoVoST-2 datasets. Our method achieves an average BLEU score of 32.2 across eight language pairs on the MuST-C dataset and an average of 36.3 across three language pairs on the CoVoST-2 dataset, establishing a new state-of-the-art (SOTA) performance.

MoLoRA: Boosting LLM-based End-to-end Speech Translation with Mixture of Low-rank Experts

Recommending event schedules is a key issue in Event-based Social Networks (EBSNs) in order to maintain user activity. An effective recommendation is required to maximize the user's preference, subjecting to both time and geographical constraints. Existing methods face an inherent trade-off among efficiency, effectiveness, and generalization, due to the NP-hard nature of the problem. This paper proposes the Chain-of-Scheduling (CoS) framework, which activates the event scheduling capability of Large Language Models (LLMs) through a guided, efficient scheduling process. CoS enhances LLM by formulating the schedule task into three atomic stages, i.e., exploration, verification and integration. Then we enable the LLMs to generate CoS autonomously via Knowledge Distillation (KD). Experimental results show that CoS achieves near-theoretical optimal effectiveness with high efficiency on three real-world datasets in a interpretable manner. Moreover, it demonstrates strong zero-shot learning ability on out-of-domain data.

CoS: Towards Optimal Event Scheduling via Chain-of-Scheduling

Rejoining fragment images of precious artifacts is a meaningful task because complete artifacts could provide valuable clues for the research of human civilization. However, existing rejoining methods face several challenges including time-consuming manual annotation, insufficient rejoining accuracy, and prohibitive computation cost. For rejoining fragment images of bone sticks (a precious artifact), we propose a lightweight vision graph neural network called RejoinViG to address these challenges. First, our method avoids time-consuming manual annotation of ballast contour data by experts. Specifically, our method directly takes a pair of fragment images as input and then determines whether the image pair is rejoinable. Second, our method improves rejoining accuracy by contour, script, and texture through dynamically constructing local and global graphs. Third, our method improves rejoining accuracy while reducing computation cost by introducing a new attention mechanism named node self-attention. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods significantly. For example, the Top-1 accuracy of our method is 3.9 times that of SFF-Siam. Surprisingly, our method successfully rejoins a pair of previously unknown but rejoinable fragment images of bone sticks in a real-world scenario.

Rejoining Precious Artifacts: Efficiently Bone Stick Rejoining Based Massive Fragment Images by Contour, Script, and Texture

Neuro-symbolic learning has emerged as a promising paradigm for interpretable visual reasoning, where mapping natural language questions to executable programs plays a central role. However, most existing methods focus exclusively on the forward program generation from questions while overlooking the reverse process of reconstructing questions from programs. In this paper, we propose BiPaR (Bidirectional Parsing and Reconstruction), a Transformer-based framework that jointly models both program parsing and question reconstruction within a unified architecture. Unlike previous approaches that only perform forward parsing, BiPaR introduces reverse program-to-question reconstruction as a powerful auxiliary signal, which improves program generation quality and accelerates convergence, particularly under limited supervision. We further provide a theoretical analysis showing how reverse reconstruction facilitates faster optimization during training. The bidirectional modeling makes BiPaR well-suited for both supervised and semi-supervised learning scenarios. We present two architectural variants: BiPaR-Full, which employs encoder-decoder Transformers for both modules; and BiPaR-DOnly, a lightweight variant that employs a decoder-only structure for question reconstruction, reducing model complexity. Experiments on CLEVR and a GQA subset demonstrate that BiPaR significantly outperforms standard Transformer baselines. Furthermore, in the semi-supervised learning setting, BiPaR achieves notable improvements by leveraging additional questions without program annotations.

Learning to Parse and Reconstruct: Bidirectional Modeling of Question-to-Program Mapping

Deep learning models excel in visual recognition but suffer severe performance drops when training labels are corrupted by noise. Under label noise prior work cannot learn accurate similarities and thus misguide the learning process. In this paper, we uncover a complementary and novel phenomenon, Dissimilarity Invariance, whereby semantic dissimilarity between unrelated samples remains stable despite label noise. Leveraging this insight, we propose NegScale, a plug-and-play framework that shifts focus from fragile similarity to robust dissimilarity. NegScale integrates: (1) Structured Negative Orthogonality Penalty (SNOP), enforcing subspace orthogonality for unrelated samples; and (2) Dissimilarity-Calibrated Similarity Adjustment (DCSA), suppressing spurious similarity using dissimilarity anchors. We also give theoretical analysis that proves Dissimilarity Invariance and the effectiveness of NegScale. Empirical results demonstrate that NegScale consistently outperforms state-of-the-art baselines, establishing new benchmarks on CIFAR with synthetic noise and real-world datasets.

Leveraging Dissimilarity Invariance as a Robust Anchor for Learning with Noisy Labels

Imitation learning for robotic manipulation faces a fundamental challenge: the scarcity of large-scale, high-quality robot demonstration data. Recent robotic foundation models often pre-train on cross-embodiment robot datasets to increase data scale, while they face significant limitations as the diverse morphologies and action spaces across different robot embodiments make unified training challenging. In this paper, we present H-RDT (Human to Robotics Diffusion Transformer), a novel approach that leverages human manipulation data to enhance robot manipulation capabilities. Our key insight is that large-scale egocentric human manipulation videos with paired 3D hand pose annotations provide rich behavioral priors that capture natural manipulation strategies and can benefit robotic policy learning. We introduce a two-stage training paradigm: (1) pre-training on large-scale egocentric human manipulation data, and (2) cross-embodiment fine-tuning on robot-specific data with modular action encoders and decoders. Built on a diffusion transformer architecture with 2B parameters, H-RDT uses flow matching to model complex action distributions. The modular design of action encoder and decoder components enables effective knowledge transfer from the unified human embodiment to diverse robot platforms through efficient fine-tuning. Extensive evaluations encompassing both simulation and real-world experiments, single-task and multitask scenarios, as well as few-shot learning and robustness assessments, demonstrate that H-RDT outperforms training from scratch and existing state-of-the-art methods, including $\boldsymbol{\pi}_0$ and RDT, achieving significant improvements of 13.9% and 40.5% over training from scratch in simulation and real-world experiments, respectively. The results validate our core hypothesis that human manipulation data can serve as a powerful foundation for learning bimanual robotic manipulation policies.

H-RDT: Human Manipulation Enhanced Bimanual Robotic Manipulation

Anchor-based 3D Gaussian Splatting (GS), exemplified by Scaffold-GS, achieves remarkable storage efficiency through a hybrid explicit-implicit representation. However, their reliance on a single, monolithic network to decode anchor features imposes a severe bottleneck on model capacity, often resulting in blurred details and view-dependent artifacts in complex scenes. To break this bottleneck, we introduce the concept of Scene Experts: a strategy that decomposes the task of modeling a complex scene across a collection of specialized sub-models. To realize the paradigm, we propose MoE-GS. Our approach designs the decoder as a Sparsely-Gated Mixture of Experts (MoE), which dramatically increases the model's total capacity while maintaining comparable inference cost via sparse activation. To effectively train this high-capacity model, we propose two key innovations: (1) A progressive curriculum learning strategy that first trains all experts on a robust baseline before encouraging them to specialize on different scene components. (2) A novel opacity-aware regularization that penalizes inactive neural Gaussians, ensuring the expanded capacity is efficiently used. Extensive experiments demonstrate that MoE-GS substantially outperforms state-of-the-art methods on diverse benchmarks, significantly improving reconstruction fidelity while requiring a smaller or comparable Gaussian model size. The code is included in the supplementary material.

Content not yet available

Next from AAAI 2026

Decoding with Structured Awareness: Integrating Directional, Frequency-Spatial, and Structural Attention for Medical Image Segmentation

Similar lecture

VisCRA: A Visual Chain Reasoning Attack for Jailbreaking Multimodal Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES