The widespread success of large language models (LLMs) on NLP benchmarks has been accompanied by concerns that LLMs function primarily as stochastic parrots that reproduce texts similar to what they saw during pre-training, often erroneously. But what is the nature of their errors, and do these errors exhibit any regularities? In this work, we examine irrelevant context hallucinations, in which models integrate misleading contextual cues into their predictions. Through behavioral analysis, we show that these errors result from a structured yet flawed mechanism that we term class-based (mis)generalization, in which models combine abstract class cues with features extracted from the query or context to derive answers. Furthermore, mechanistic interpretability experiments on Llama-3, Mistral, and Pythia across 39 factual recall relation types reveal that this behavior is reflected in the model's internal computations: (i) abstract class representations are constructed in lower layers before being refined into specific answers in higher layers, (ii) feature selection is governed by two competing circuits -- one prioritizing direct query-based reasoning, the other incorporating contextual cues -- whose relative influences determine the final output. Our findings provide a more nuanced perspective on the stochastic parrot argument: through form-based training, LLMs can exhibit generalization leveraging abstractions, albeit in unreliable ways based on contextual cues -- what we term stochastic chameleons.

Large language models (LLMs) exhibit excellent performance in natural language processing (NLP), but remain highly sensitive to the quality of input queries, especially when these queries contain misleading or inaccurate information. Existing methods focus on correcting the output, but they often overlook the potential of improving the ability of LLMs to detect and correct misleading content in the input itself. In this paper, we propose a novel three-stage fine-tuning method that enhances the ability of LLMs to detect and correct misleading information in the input, further improving response accuracy and reducing hallucinations. Specifically, the three stages include (1) training LLMs to identify misleading information, (2) training LLMs to correct the misleading information using built-in or external knowledge, and (3) training LLMs to generate accurate answers based on the corrected queries. To evaluate our method, we conducted experiments on three datasets for the hallucination detection task and the question answering~(QA) task, as well as two datasets containing misleading information that we constructed. The experimental results demonstrate that our method significantly improves the accuracy and factuality of LLM responses, while also enhancing the ability to detect hallucinations and reducing the generation of hallucinations in the output, particularly when the query contains misleading information.

Endowing dialogue agents with persona information has proven to significantly improve the consistency and diversity of their generations. While much focus has been placed on aligning dialogues with provided personas, the adaptation to the interlocutor's profile remains largely underexplored. In this work, we investigate three key aspects: (1) a model's ability to align responses with both the provided persona and the interlocutor's; (2) its robustness when dealing with familiar versus unfamiliar interlocutors and topics, and (3) the impact of additional fine-tuning on specific persona-based dialogues. We evaluate dialogues generated with diverse speaker pairings and topics, framing the evaluation as an author identification task and employing both LLM-as-a-judge and human evaluations. By systematically masking or disclosing information about the interlocutor, we assess its impact on dialogue generation. Results show that access to the interlocutor's persona improves the recognition of the target speaker, while masking it does the opposite. Although models generalise well across topics, they struggle with unfamiliar interlocutors. Finally, we found that in zero-shot settings, LLMs often copy biographical details, facilitating identification but trivialising the task.

Summarizing long-form narratives--such as books, movies, and TV scripts--requires capturing intricate plotlines, character interactions, and thematic coherence, a task that remains challenging for existing LLMs. We introduce NexusSum, a multi-agent LLM framework for narrative summarization that processes long-form text through a structured, sequential pipeline--without requiring fine-tuning. Our approach introduces two key innovations: (1) Dialogue-to-Description Transformation: A narrative-specific preprocessing method that standardizes character dialogue and descriptive text into a unified format, improving coherence. (2) Hierarchical Multi-LLM Summarization: A structured summarization pipeline that optimizes chunk processing and controls output length for accurate, high-quality summaries. Our method establishes a new state-of-the-art in narrative summarization, achieving up to a 30.0% improvement in BERTScore (F1) across books, movies, and TV scripts. These results demonstrate the effectiveness of multi-agent LLMs in handling long-form content, offering a scalable approach for structured summarization in diverse storytelling domains.

Recent advancements in Generative AI and Large Language Models (LLMs) have enabled the creation of highly realistic synthetic content, raising concerns about the potential for malicious use, such as misinformation and manipulation. Moreover, detecting Machine-Generated Text (MGT) remains challenging due to the lack of robust benchmarks that assess generalization to real-world scenarios. In this work, we present a pipeline to test the resilience of state-of-the-art MGT detectors (e.g., Mage, Radar, LLM-DetectAIve) to linguistically informed adversarial attacks. To challenge the detectors, we fine-tune language models using Direct Preference Optimization (DPO) to shift the MGT style toward human-written text (HWT). This exploits the detectors' reliance on stylistic clues, making new generations more challenging to detect. Additionally, we analyze the linguistic shifts induced by the alignment and which features are used by detectors to detect MGT texts. Our results show that detectors can be easily fooled with relatively few examples, resulting in a significant drop in detection performance. This highlights the importance of improving detection methods and making them robust to unseen in-domain texts.

The rapid spread of misinformation, further amplified by recent advances in generative AI, poses significant threats to society, impacting public opinion, democratic stability, and national security. Understanding and proactively assessing these threats requires exploring methodologies that enable structured and scalable misinformation generation. In this paper, we propose a novel approach that leverages knowledge graphs (KGs) as structured semantic resources to systematically generate fake triplets. By analyzing the structural properties of KGs, such as the distance between entities and their predicates, we identify plausibly false relationships. These triplets are then used to guide large language models (LLMs) in generating misinformation statements with varying degrees of credibility. By utilizing structured semantic relationships, our deterministic approach produces misinformation inherently challenging for humans to detect, drawing exclusively upon publicly available KGs (e.g., WikiGraphs). Additionally, we investigate the effectiveness of LLMs in distinguishing between genuine and artificially generated misinformation. Our analysis highlights significant limitations in current LLM-based detection methods, underscoring the necessity for enhanced detection strategies and a deeper exploration of inherent biases in generative models.

Although multi-agent systems based on large language models show strong capabilities on multiple tasks, they are still limited by high computational overhead, information loss, and robustness. Inspired by ResNet's residual learning, we propose Residual Mixture-of-Agents (RMoA), integrating residual connections to optimize efficiency and reliability. To maximize information utilization from model responses while minimizing computational costs, we innovatively design an embedding-based diversity selection mechanism that greedily selects responses via vector similarity. Furthermore, to mitigate iterative information degradation, we introduce a Residual Extraction Agent to preserve cross-layer incremental information by capturing inter-layer response differences, coupled with a Residual Aggregation Agent for hierarchical information integration. Additionally, we propose an adaptive termination mechanism that dynamically halts processing based on residual convergence, further improving inference efficiency. RMoA achieves state-of-the-art performance on the benchmarks of across alignment, mathematical reasoning, code generation, and multitasking understanding, while significantly reducing computational overhead. Code is available at https://github.com/mindhunter01/RMoA.

Accurate modeling of subjective phenomena such as emotion expression requires data annotated with authors' intentions. Commonly such data is collected by asking study participants to donate and label genuine content produced in the real world, or create content fitting particular labels during the study. Asking participants to create content is often simpler to implement and presents fewer risks to participant privacy than data donation. However, it is unclear if and how study-created content may differ from genuine content, and how differences may impact models. We collect study-created and genuine multimodal social media posts labeled for emotion and compare them on several dimensions, including model performance. We find that compared to genuine posts, study-created posts are longer, rely more on their text and less on their images for emotion expression, and focus more on emotion-prototypical events. The samples of participants willing to donate versus create posts are demographically different. Study-created data is valuable to train models that generalize well to genuine data, but realistic effectiveness estimates require genuine data.

Large language models and vision-language models (which we jointly call LMs) have transformed NLP and CV, demonstrating remarkable potential across various fields. However, their capabilities in affective analysis (i.e. sentiment analysis and emotion detection) remain underexplored. This gap is largely due to the absence of comprehensive evaluation benchmarks, and the inherent complexity of affective analysis tasks. In this paper, we introduce MMAFFBen, the first extensive open-source benchmark for multilingual multimodal affective analysis. MMAFFBen encompasses text, image, and video modalities across 35 languages, covering four key affective analysis tasks: sentiment polarity, sentiment intensity, emotion classification, and emotion intensity. Moreover, we construct the MMAFFIn dataset for fine-tuning LMs on affective analysis tasks, and further develop MMAFFLM-3b and MMAFFLM-7b based on it. We evaluate various representative LMs, including GPT-4o-mini, providing a systematic comparison of their affective understanding capabilities. This project is available at https://github.com/lzw108/MMAFFBen.

This study introduces a novel conceptual framework distinguishing problem-seeking from problem-solving to clarify the unique features of human intelligence in contrast to AI. Problem-seeking refers to the embodied, emotionally grounded process by which humans identify and set goals, while problem-solving denotes the execution of strategies aimed at achieving such predefined objectives. The framework emphasizes that while AI excels at efficiency and optimization, it lacks the orientation derived from experiential grounding and the embodiment flexibility intrinsic to human cognition. To empirically explore this distinction, the research analyzes metadata from 157 YouTube videos discussing AI. Conducting a thematic analysis combining qualitative insights with keyword-based quantitative metrics, this mixed-methods approach uncovers recurring themes in public discourse, including privacy, job displacement, misinformation, optimism, and ethical concerns. The results reveal a dual sentiment: public fascination with AI's capabilities coexists with anxiety and skepticism about its societal implications. The discussion critiques the orthogonality thesis, which posits that intelligence is separable from goal content, and instead argues that human intelligence integrates goal-setting and goal-pursuit. It underscores the centrality of embodied cognition in human reasoning and highlights how AI's limitations come from its current reliance on computational processing. The study advocates for enhancing emotional and digital literacy to foster responsible AI engagement. It calls for reframing public discourse to recognize AI as a tool that augments -- rather than replaces -- human intelligence. By positioning problem seeking at the core of cognition and as a critical dimension of intelligence, this research offers new perspectives on ethically aligned and human-centered AI development.