LLMs have shown strong performance on human-centric reasoning tasks. While previous evaluations have explored whether LLMs can infer intentions or detect deception, they often overlook the individualized reasoning styles that influence how people interpret and act in social contexts. Social deduction games (SDGs) provide a natural testbed for evaluating individualized reasoning styles, where different players may adopt diverse but contextually valid reasoning strategies under identical conditions. To address this, we introduce InMind, a cognitively grounded evaluation framework designed to assess whether LLMs can capture and apply personalized reasoning styles in SDGs. InMind enhances structured gameplay data with round-level strategy traces and post-game reflections, collected under both Observer and Participant modes. It supports four cognitively motivated tasks that jointly evaluate both static alignment and dynamic adaptation. As a case study, we apply InMind to the game Avalon, evaluating 11 state-of-the-art LLMs. General-purpose LLMs, even GPT-4o frequently rely on lexical cues, struggling to anchor reflections in temporal gameplay or adapt to evolving strategies. In contrast, reasoning-enhanced LLMs like DeepSeek-R1 exhibit early signs of style-sensitive reasoning. These findings reveal key limitations in current LLMs’ capacity for individualized, adaptive reasoning, and position InMind as a step toward cognitively aligned human–AI interaction.

Phonetic Cloaking Replacement (PCR), defined as the deliberate use of homophonic or near-homophonic variants to hide toxic intent, has become a major obstacle to Chinese content moderation. While this problem is well-recognized, existing evaluations predominantly rely on rule-based, synthetic perturbations that ignore the creativity of real users. We organize PCR into a four-way surface-form taxonomy and compile PCR-ToxiCN, a dataset of 500 naturally occurring, phonetically cloaked offensive posts gathered from the RedNote platform. Benchmarking state-of-the-art LLMs on this dataset exposes a serious weakness: the best model reaches only an F1-score of 0.672, and zero-shot chain-of-thought prompting pushes performance even lower. Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy. This study offers the first comprehensive taxonomy of Chinese PCR, a realistic benchmark that reveals current detectors’ limits, and a lightweight mitigation technique that advances research on robust toxicity detection.

“Commonsense reasoning and moral understanding are crucial tasks in artificial intelligence (AI) and natural language processing (NLP). However, existing research often falls short in terms of faithfulness and informativeness during the reasoning process. We propose a novel framework for performing commonsense reasoning and moral understanding using large language models (LLMs), involving constructing guided prompts by incorporating relevant knowledge for commonsense reasoning and extracting facts from stories for moral understanding. We conduct extensive experiments on the Commonsense Reasoning and Moral Understanding in Children’s Stories (CRMUS) dataset with widely recognised LLMs under both zero-shot and fine-tuning settings, demonstrating the effectiveness of our proposed method. Furthermore, we analyse the adaptability of different LLMs in extracting facts for moral understanding performance.”

Safe and reliable natural language inference is critical for extracting insights from clinical trial reports but poses challenges due to biases in large pre-trained language models. This paper presents a novel data augmentation technique to improve model robustness for biomedical natural language inference in clinical trials. By generating synthetic examples through semantic perturbations and domain-specific vocabulary replacement and adding a new task for numerical and quantitative reasoning, we introduce greater diversity and reduce shortcut learning. Our approach, combined with multi-task learning and the DeBERTa architecture, achieved significant performance gains on the NLI4CT 2024 benchmark compared to the original language models. Ablation studies validate the contribution of each augmentation method in improving robustness. Our best-performing model ranked 12th in terms of faithfulness and 8th in terms of consistency, respectively, out of the 32 participants.

Emotion detection from text is a crucial task in understanding natural language with wide-ranging applications. Existing approaches for multilingual emotion detection from text face challenges with data scarcity across many languages and a lack of interpretability. We propose a novel method that leverages both monolingual and multilingual pre-trained language models to improve performance and interpretability. Our approach involves 1) training a high-performing English monolingual model in parallel with a multilingual model and 2) using knowledge distillation to transfer the emotion detection capabilities from the monolingual teacher to the multilingual student model. Experiments on a multilingual dataset demonstrate significant performance gains for refined multilingual models like XLM-RoBERTa and E5 after distillation. Furthermore, our approach enhances interpretability by enabling better identification of emotion-trigger words. Our work presents a promising direction for building accurate, robust and explainable multilingual emotion detection systems.

In recent years, pre-trained language models have garnered significant attention due to their effectiveness, which stems from the rich knowledge acquired during pre-training. To mitigate the inconsistency issues between pre-training tasks and downstream tasks and to facilitate the resolution of language-related issues, prompt-based approaches have been introduced, which are particularly useful in low-resource scenarios. However, existing approaches mostly rely on verbalizers to translate the predicted vocabulary to task-specific labels. The major limitations of this approach are the ignorance of potentially relevant domain-specific words and being biased by the pre-training data. To address these limitations, we propose a framework that incorporates conceptual knowledge for text classification in the extreme zero-shot setting. The framework includes prompt-based keyword extraction, weight assignment to each prompt keyword, and final representation estimation in the knowledge graph embedding space. We evaluated the method on four widely-used datasets for sentiment analysis and topic detection, demonstrating that it consistently outperforms recently-developed prompt-based approaches in the same experimental settings.

This paper describes a working system which performs natural language inference using polarity-marked parse trees. The system handles all of the instances of monotonicity inference in the FraCaS data set. Except for the initial parse, it is entirely deterministic. It handles multi-premise arguments, and the kind of inference performed is essentially “logical”, but it goes beyond what is representable in first-order logic. In any case, the system works on surface forms rather than on representations of any kind.

The lack of labeled data is one of the main challenges when building a task-oriented dialogue system. Existing dialogue datasets usually rely on human labeling, which is expensive, limited in size, and in low coverage. In this paper, we instead propose our framework auto-dialabel to automatically cluster the dialogue intents and slots. In this framework, we collect a set of context features, leverage an autoencoder for feature assembly, and adapt a dynamic hierarchical clustering method for intent and slot labeling. Experimental results show that our framework can promote human labeling cost to a great extent, achieve good intent clustering accuracy (84.1%), and provide reasonable and instructive slot labeling results.