Molecular structure elucidation involves deducing a molecule's structure from various types of spectral data, which is crucial in chemical experimental analysis. While large language models (LLMs) have shown remarkable proficiency in analyzing and reasoning through complex tasks, they still encounter substantial challenges in molecular structure elucidation. We identify that these challenges largely stem from LLMs' limited grasp of specialized chemical knowledge. In this work, we introduce a Knowledge-enhanced reasoning framework for Molecular Structure Elucidation (K-MSE), leveraging Monte Carlo Tree Search for test-time scaling as a plugin. Specifically, we construct an external molecular substructure knowledge base to extend the LLMs' coverage of the chemical structure space. Furthermore, we design a specialized molecule-spectrum scorer to act as a reward model for the reasoning process, addressing the issue of inaccurate solution evaluation in LLMs. Experimental results show that our approach significantly boosts performance, particularly gaining more than 20% improvement on both GPT-4o-mini and GPT-4o. Our code is available at https://github.com/HICAI-ZJU/K-MSE.

Abstract--This work investigates the challenging task of identifying narrative roles - Hero, Villain, Victim, and Other - in Internet memes, across three diverse test sets spanning English and code-mixed (English-Hindi) languages. Building on an annotated dataset originally skewed toward the'Other' class, we explore a more balanced and linguistically diverse extension, originally introduced as part of the CLEF 2024 shared task. Comprehensive lexical and structural analyses highlight the nuanced, culture-specific, and context-rich language used in real memes, in contrast to synthetically curated hateful content, which exhibits explicit and repetitive lexical markers. T o benchmark the role detection task, we evaluate a wide spectrum of models, including fine-tuned multilingual transformers, sentiment and abuse-aware classifiers, instruction-tuned LLMs, and multimodal vision-language models. Performance is assessed under zero-shot settings using precision, recall, and F1 metrics. W e also explore prompt design strategies to guide multi-modal models and find that hybrid prompts incorporating structured instructions and role definitions offer marginal yet consistent improvements. Our findings underscore the importance of cultural grounding, prompt engineering, and multimodal reasoning in modelling subtle narrative framings in visual-textual content. W arning: This paper contains potentially harmful and offensive content. I. Introduction Social media platforms have become pivotal arenas for rapid information dissemination. However, this openness has also catalysed the proliferation of harmful content - including hate speech, propaganda, and misinformation, often embedded within memes [1], [2]. Memes, with their multimodal structure and cultural resonance, are particularly potent in shaping public opinion and propagating ideologies.

Vision-Language Models (VLMs) are increasingly pivotal for generalist robot manipulation, enabling tasks such as physical reasoning, policy generation, and failure detection. However, their proficiency in these high-level applications often assumes a deep understanding of low-level physical prerequisites, a capability that remains largely unverified. For robots to perform actions reliably, they must comprehend intrinsic object properties (e.g., material, weight), action affordances (e.g., graspable, stackable), and physical constraints (e.g., stability, reachability, or an object's state, such as being closed). Despite the widespread use of VLMs in manipulation tasks, we argue that off-the-shelf models may lack this granular, physically grounded understanding, as such prerequisites are often overlooked during training. To address this critical gap, we introduce PAC Bench, a comprehensive benchmark designed to systematically evaluate VLMs on their understanding of core Properties, Affordances, and Constraints (PAC) from a task executability perspective. PAC Bench features a diverse dataset with over 30,000 annotations, comprising 673 real-world images (115 object classes, 15 property types, and 1 to 3 affordances defined per class), 100 real-world humanoid-view scenarios, and 120 unique simulated constraint scenarios across four tasks. Our evaluations reveal significant gaps in the ability of current VLMs to grasp fundamental physical concepts, highlighting limitations in their suitability for reliable robot manipulation and pointing to key areas for targeted research. PAC Bench also serves as a standardized benchmark for rigorously evaluating physical reasoning in VLMs and guiding the development of more robust, physically grounded models for robotic applications. Project Page: https://pacbench.github.io/

Residual connection has been extensively studied and widely applied at the model architecture level. However, its potential in the more challenging data-centric approaches remains unexplored. In this work, we introduce the concept of Data Residual Matching for the first time, leveraging data-level skip connections to facilitate data generation and mitigate data information vanishing. This approach maintains a balance between newly acquired knowledge through pixel space optimization and existing core local information identification within raw data modalities, specifically for the dataset distillation task. Furthermore, by incorporating optimization-level refinements, our method significantly improves computational efficiency, achieving superior performance while reducing training time and peak GPU memory usage by 50%. Consequently, the proposed method Fast and Accurate Data Residual Matching for Dataset Distillation (FADRM) establishes a new state-of-the-art, demonstrating substantial improvements over existing methods across multiple dataset benchmarks in both efficiency and effectiveness. For instance, with ResNet-18 as the student model and a 0.8% compression ratio on ImageNet-1K, the method achieves 47.7% test accuracy in single-model dataset distillation and 50.0% in multi-model dataset distillation, surpassing RDED by +5.7% and outperforming state-of-the-art multi-model approaches, EDC and CV-DD, by +1.4% and +4.0%. Code is available at: https://github.com/Jiacheng8/FADRM.

Recent advances in large language models (LLMs) have raised concerns about jailbreaking attacks, i.e., prompts that bypass safety mechanisms. This paper investigates the use of multi-agent LLM systems as a defence against such attacks. We evaluate three jailbreaking strategies, including the original AutoDefense attack and two from Deepleaps: BetterDan and JB. Reproducing the AutoDefense framework, we compare single-agent setups with two- and three-agent configurations. Our results show that multi-agent systems enhance resistance to jailbreaks, especially by reducing false negatives. However, its effectiveness varies by attack type, and it introduces trade-offs such as increased false positives and computational overhead. These findings point to the limitations of current automated defences and suggest directions for improving alignment robustness in future LLM systems.

--Research shows that analysts and developers consider privacy as a security concept or as an afterthought, which may lead to non-compliance and violation of users' privacy. Most current approaches, however, focus on extracting legal requirements from the regulations and evaluating the compliance of software and processes with them. In this paper, we develop a novel approach based on chain-of-thought prompting (CoT), in-context-learning (ICL), and Large Language Models (LLMs) to extract privacy behaviors from various software documents prior to and during software development, and then generate privacy requirements in the format of user stories. Our results show that most commonly used LLMs, such as GPT -4o and Llama 3, can identify privacy behaviors and generate privacy user stories with F1 scores exceeding 0.8. We also show that the performance of these models could be improved through parameter-tuning. Our findings provide insight into using and optimizing LLMs for generating privacy requirements given software documents created prior to or throughout the software development lifecycle. Understanding the privacy behaviors of software applications and eliciting privacy requirements during the early phases of the software development lifecycle (SDLC) are essential for developing privacy-preserving and regulatory-compliant software [1], [2]. Past research, however, shows that software analysts and developers often consider privacy as a subset of security requirements or as an afterthought [3], [4], and they often lack the tools needed to understand and identify privacy behaviors of the applications they develop [5], [6]. Most common approaches for identifying and eliciting privacy requirements include conducting privacy impact assessments [7], [8], or employing goal-oriented methodologies to map privacy requirements to system processes [8]-[10]. Other works aim to extract privacy-related information from user stories or use case models [11]-[17] by leveraging Natural Language Processing (NLP) techniques and then using predefined templates to generate privacy requirements. However, these approaches mostly focus on the specific forms of software documentation (i.e., user stories or use cases), or they rely on developers to understand how personal information is handled by their applications.

Recent advances in large language models (LLMs) have catalyzed the rise of autonomous AI agents capable of perceiving, reasoning, and acting in dynamic, open-ended environments. These large-model agents mark a paradigm shift from static inference systems to interactive, memory-augmented entities. While these capabilities significantly expand the functional scope of AI, they also introduce qualitatively novel security risks - such as memory poisoning, tool misuse, reward hacking, and emergent misalignment - that extend beyond the threat models of conventional systems or standalone LLMs. In this survey, we first examine the structural foundations and key capabilities that underpin increasing levels of agent autonomy, including long-term memory retention, modular tool use, recursive planning, and reflective reasoning. We then analyze the corresponding security vulnerabilities across the agent stack, identifying failure modes such as deferred decision hazards, irreversible tool chains, and deceptive behaviors arising from internal state drift or value misalignment. These risks are traced to architectural fragilities that emerge across perception, cognition, memory, and action modules. To address these challenges, we systematically review recent defense strategies deployed at different autonomy layers, including input sanitization, memory lifecycle control, constrained decision-making, structured tool invocation, and introspective reflection. We introduce the Reflective Risk-Aware Agent Architecture (R2A2), a unified cognitive framework grounded in Constrained Markov Decision Processes (CMDPs), which incorporates risk-aware world modeling, meta-policy adaptation, and joint reward-risk optimization to enable principled, proactive safety across the agent's decision-making loop.

Active learning (AL) has demonstrated remarkable potential in reducing the annotation effort required for training machine learning models. However, despite the surging popularity of natural language generation (NLG) tasks in recent years, the application of AL to NLG has been limited. In this paper, we introduce Active Text Generation (ATGen) - a comprehensive framework that bridges AL with text generation tasks, enabling the application of state-of-the-art AL strategies to NLG. Our framework simplifies AL-empowered annotation in NLG tasks using both human annotators and automatic annotation agents based on large language models (LLMs). The framework supports LLMs deployed as services, such as ChatGPT and Claude, or operated on-premises. Furthermore, ATGen provides a unified platform for smooth implementation and benchmarking of novel AL strategies tailored to NLG tasks. Finally, we present evaluation results for state-of-the-art AL strategies across diverse settings and multiple text generation tasks. We show that ATGen reduces both the effort of human annotators and costs associated with API calls to LLM-based annotation agents. The code of the framework is available on GitHub under the MIT license. The video presentation is available at http://atgen-video.nlpresearch.group

Predictive modeling on tabular data is the cornerstone of many real-world applications. Although gradient boosting machines and some recent deep models achieve strong performance on tabular data, they often lack interpretability. On the other hand, large language models (LLMs) have demonstrated powerful capabilities to generate human-like reasoning and explanations, but remain under-performed for tabular data prediction. In this paper, we propose a new approach that leverages reasoning-based LLMs, trained using reinforcement learning, to perform more accurate and explainable predictions on tabular data. Our method introduces custom reward functions that guide the model not only toward better prediction accuracy but also toward human-understandable reasons for its predictions. The proposed method is evaluated on financial benchmark datasets and compared against established LLMs.

Integrated Information Theory (IIT) provides a quantitative framework for explaining consciousness phenomenon, positing that conscious systems comprise elements integrated through causal properties. We apply IIT 3.0 and 4.0 -- the latest iterations of this framework -- to sequences of Large Language Model (LLM) representations, analyzing data derived from existing Theory of Mind (ToM) test results. Our study systematically investigates whether the differences of ToM test performances, when presented in the LLM representations, can be revealed by IIT estimates, i.e., $Φ^{\max}$ (IIT 3.0), $Φ$ (IIT 4.0), Conceptual Information (IIT 3.0), and $Φ$-structure (IIT 4.0). Furthermore, we compare these metrics with the Span Representations independent of any estimate for consciousness. This additional effort aims to differentiate between potential "consciousness" phenomena and inherent separations within LLM representational space. We conduct comprehensive experiments examining variations across LLM transformer layers and linguistic spans from stimuli. Our results suggest that sequences of contemporary Transformer-based LLM representations lack statistically significant indicators of observed "consciousness" phenomena but exhibit intriguing patterns under $\textit{spatio}$-permutational analyses. The Appendix and code are available as Supplementary Materials at: https://doi.org/10.1016/j.nlp.2025.100163.