The 2025 Global Risks Report identifies state-based armed conflict and societal polarisation among the most pressing global threats, with social media playing a central role in amplifying toxic discourse. Memes, as a widely used mode of online communication, often serve as vehicles for spreading harmful content. However, limitations in data accessibility and the high cost of dataset curation hinder the development of robust meme moderation systems. To address this challenge, in this work, we introduce a first-of-its-kind dataset of 6,300 real-world meme-based posts annotated in two stages: (i) binary classification into toxic and normal, and (ii) fine-grained labelling of toxic memes as hateful, dangerous, or offensive. A key feature of this dataset is that it is enriched with auxiliary metadata of socially relevant tags, enhancing the context of each meme. In addition, we propose a tag generation module that produces socially grounded tags, because most in-the-wild memes often do not come with tags. Experimental results show that incorporating these tags substantially enhances the performance of state-of-the-art VLMs detection tasks. Our contributions offer a novel and scalable foundation for improved content moderation in multimodal online environments.

Consider one or more AI agents performing daily living tasks in collaboration with a human they have not worked with before. Figure 1 shows snapshots of a motivating scenario in which two AI agents (male, blue shirt; female, red dress) and a human agent (female, green top) are preparing breakfast and setting up a workstation. The agents (AI, human) have a limited view of the environment and do not communicate with each other, although each of them is aware of the state of the domain, including the location of teammates and the outcomes of their actions (e.g., change in location of an object moved by a teammate). The AI agents have to reason with different descriptions of domain knowledge and uncertainty that include qualitative statements ("eggs are usually in the fridge") and quantitative measures of uncertainty ("I am 90% sure I saw the eggs on the kitchen table"), adapting their actions to changes in the domain and teammates' behavior. These characteristics correspond to Ad Hoc T eamwork (AHT), which requires cooperation "on the fly" without prior coordination [1]; many practical problems such as disaster rescue are AHT problems. The state of the art in AHT has moved from using preset protocols that define specific actions to be performed in specific states, to methods that use a long history of prior experiences to build a deep network model of the behavior of other agents (or agent types) and optimize the ad hoc agent's behavior [2]. However, it is difficult to gather large datasets of different situations in complex domains. Also, these methods are opaque and make it difficult to revise the existing models over time. In a departure from existing work, we design an architecture for AHT that bridges knowledge-based and data-driven reasoning and learning, enabling an ad hoc agent to: Leverage the ability of a Large Language Model (LLM) to anticipate future high-level tasks to be completed, revising and adapting the LLM's output to domain-specific knowledge and experience; Perform non-monotonic logical reasoning with prior commonsense domain knowledge at different abstractions, and learned models predicting the behavior of other agents, toward achieving current and anticipated tasks as joint goals; and Rapidly identify the need for, learn, and revise the models predicting the behavior of each teammate to facilitate scalable collaboration in complex domains.

Gradient inversion attacks pose significant privacy threats to distributed training frameworks such as federated learning, enabling malicious parties to reconstruct sensitive local training data from gradient communications between clients and an aggregation server during the aggregation process. While traditional encryption-based defenses, such as homo-morphic encryption, offer strong privacy guarantees without compromising model utility, they often incur prohibitive computational overheads. To mitigate this, selective encryption has emerged as a promising approach, encrypting only a subset of gradient data based on the data's significance under a certain metric. However, there have been few systematic studies on how to specify this metric in practice. This paper systematically evaluates selective encryption methods with different significance metrics against state-of-the-art attacks. Our findings demonstrate the feasibility of selective encryption in reducing computational overhead while maintaining resilience against attacks. We propose a distance-based significance analysis framework that provides theoretical foundations for selecting critical gradient elements for encryption. Through extensive experiments on different model architectures (LeNet, CNN, BERT, GPT-2) and attack types, we identify gradient magnitude as a generally effective metric for protection against optimization-based gradient inversions. However, we also observe that no single selective encryption strategy is universally optimal across all attack scenarios, and we provide guidelines for choosing appropriate strategies for different model architectures and privacy requirements.

Aligning large language models (LLMs) with human preferences is a critical challenge in AI research. While methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) are widely used, they often rely on large, costly preference datasets. The current work lacks methods for high-quality data selection specifically for preference data. In this work, we introduce a novel difficulty-based data selection strategy for preference datasets, grounded in the DPO implicit reward mechanism. By selecting preference data examples with smaller DPO implicit reward gaps, which are indicative of more challenging cases, we improve data efficiency and model alignment. Our approach consistently outperforms five strong baselines across multiple datasets and alignment tasks, achieving superior performance with only 10\% of the original data. This principled, efficient selection method offers a promising solution for scaling LLM alignment with limited resources.

Reinforcement learning has significantly enhanced the reasoning capabilities of Large Language Models (LLMs) in complex problem-solving tasks. Recently, the introduction of DeepSeek R1 has inspired a surge of interest in leveraging rule-based rewards as a low-cost alternative for computing advantage functions and guiding policy optimization. However, a common challenge observed across many replication and extension efforts is that when multiple sampled responses under a single prompt converge to identical outcomes, whether correct or incorrect, the group-based advantage degenerates to zero. This leads to vanishing gradients and renders the corresponding samples ineffective for learning, ultimately limiting training efficiency and downstream performance. To address this issue, we propose a consistency-aware policy optimization framework that introduces a structured global reward based on outcome consistency, the global loss based on it ensures that, even when model outputs show high intra-group consistency, the training process still receives meaningful learning signals, which encourages the generation of correct and self-consistent reasoning paths from a global perspective. Furthermore, we incorporate an entropy-based soft blending mechanism that adaptively balances local advantage estimation with global optimization, enabling dynamic transitions between exploration and convergence throughout training. Our method introduces several key innovations in both reward design and optimization strategy. We validate its effectiveness through substantial performance gains on multiple mathematical reasoning benchmarks, highlighting the proposed framework's robustness and general applicability. Code of this work has been released at https://github.com/hijih/copo-code.git.

The application of Large Language Models (LLMs) is growing in the productive completion of Software Engineering tasks. Yet, studies investigating the productive prompting techniques often employed a limited problem space, primarily focusing on well-known prompting patterns and mainly targeting function-level SE practices. We identify significant gaps in real-world workflows that involve complexities beyond class-level (e.g., multi-class dependencies) and different features that can impact Human-LLM Interactions (HLIs) processes in code generation. To address these issues, we designed an experiment that comprehensively analyzed the HLI features regarding the code generation productivity. Our study presents two project-level benchmark tasks, extending beyond function-level evaluations. We conducted a user study with 36 participants from diverse backgrounds, asking them to solve the assigned tasks by interacting with the GPT assistant using specific prompting patterns. We also examined the participants' experience and their behavioral features during interactions by analyzing screen recordings and GPT chat logs. Our statistical and empirical investigation revealed (1) that three out of 15 HLI features significantly impacted the productivity in code generation; (2) five primary guidelines for enhancing productivity for HLI processes; and (3) a taxonomy of 29 runtime and logic errors that can occur during HLI processes, along with suggested mitigation plans.

Automated grading systems can efficiently score short-answer responses, yet they often fail to indicate when a grading decision is uncertain or potentially contentious. We introduce semantic entropy, a measure of variability across multiple GPT-4-generated explanations for the same student response, as a proxy for human grader disagreement. By clustering rationales via entailment-based similarity and computing entropy over these clusters, we quantify the diversity of justifications without relying on final output scores. We address three research questions: (1) Does semantic entropy align with human grader disagreement? (2) Does it generalize across academic subjects? (3) Is it sensitive to structural task features such as source dependency? Experiments on the ASAP-SAS dataset show that semantic entropy correlates with rater disagreement, varies meaningfully across subjects, and increases in tasks requiring interpretive reasoning. Our findings position semantic entropy as an interpretable uncertainty signal that supports more transparent and trustworthy AI-assisted grading workflows.

Recent studies have extended the application of large language models (LLMs) to geographic problems, revealing surprising geospatial competence even without explicit spatial supervision. However, LLMs still face challenges in spatial consistency, multi-hop reasoning, and geographic bias. To address these issues, we propose GeoSR, a self-refining agentic reasoning framework that embeds core geographic principles -- most notably Tobler's First Law of Geography -- into an iterative prediction loop. In GeoSR, the reasoning process is decomposed into three collaborating agents: (1) a variable-selection agent that selects relevant covariates from the same location; (2) a point-selection agent that chooses reference predictions at nearby locations generated by the LLM in previous rounds; and (3) a refine agent that coordinates the iterative refinement process by evaluating prediction quality and triggering further rounds when necessary. This agentic loop progressively improves prediction quality by leveraging both spatial dependencies and inter-variable relationships. We validate GeoSR on tasks ranging from physical-world property estimation to socioeconomic prediction. Experimental results show consistent improvements over standard prompting strategies, demonstrating that incorporating geostatistical priors and spatially structured reasoning into LLMs leads to more accurate and equitable geospatial predictions. The code of GeoSR is available at https://github.com/JinfanTang/GeoSR.

Controllable Text Generation (CTG) is a vital subfield in Natural Language Processing (NLP), aiming to generate text that aligns with desired attributes. However, previous studies commonly focus on the quality of controllable text generation for short sequences, while the generation of long-form text remains largely underexplored. In this paper, we observe that the controllability of texts generated by the powerful prefix-based method Air-Decoding tends to decline with increasing sequence length, which we hypothesize primarily arises from the observed decay in attention to the prefixes. Meanwhile, different types of prefixes including soft and hard prefixes are also key factors influencing performance. Building on these insights, we propose a lightweight and effective framework called Dynamic Token-level Prefix Augmentation (DTPA) based on Air-Decoding for controllable text generation. Specifically, it first selects the optimal prefix type for a given task. Then we dynamically amplify the attention to the prefix for the attribute distribution to enhance controllability, with a scaling factor growing exponentially as the sequence length increases. Moreover, based on the task, we optionally apply a similar augmentation to the original prompt for the raw distribution to balance text quality. After attribute distribution reconstruction, the generated text satisfies the attribute constraints well. Experiments on multiple CTG tasks demonstrate that DTPA generally outperforms other methods in attribute control while maintaining competitive fluency, diversity, and topic relevance. Further analysis highlights DTPA's superior effectiveness in long text generation.

Jailbreaking -- bypassing built-in safety mechanisms in AI models -- has traditionally required complex technical procedures or specialized human expertise. In this study, we show that the persuasive capabilities of large reasoning models (LRMs) simplify and scale jailbreaking, converting it into an inexpensive activity accessible to non-experts. We evaluated the capabilities of four LRMs (DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, Qwen3 235B) to act as autonomous adversaries conducting multi-turn conversations with nine widely used target models. LRMs received instructions via a system prompt, before proceeding to planning and executing jailbreaks with no further supervision. We performed extensive experiments with a benchmark of harmful prompts composed of 70 items covering seven sensitive domains. This setup yielded an overall attack success rate across all model combinations of 97.14%. Our study reveals an alignment regression, in which LRMs can systematically erode the safety guardrails of other models, highlighting the urgent need to further align frontier models not only to resist jailbreak attempts, but also to prevent them from being co-opted into acting as jailbreak agents.