PCoKG: Personality-aware Commonsense Reasoning with Debate
Weijie Li, Zhongqing Wang, Guodong Zhou · 2026
Read PCoKG: Personality-aware Commonsense Reasoning with Debate by Weijie Li, Zhongqing Wang, Guodong Zhou (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06234.txt.
Introduction
What a person is like shapes how they think through daily life. Most AI common-sense tools miss that. Weijie Li and 2 co-authors build a new bank of facts for it in a 2026 paper. Talk tools that chat with users need this feel for traits to stay in role.
Their bank holds more than half a million linked facts. They picked events from an old data set that split views by trait type. Large AI models then played roles and gave views for each type. A debate step, with one side for, one side against, and a judge, cleaned the work. Tests showed big base models gained most from the bank. Used in trait-based chat, replies matched the source text more closely than before.
Catalog record
| CREATOR | Weijie Li, Zhongqing Wang, Guodong Zhou |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-06234
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PCoKG: Personality-aware Commonsense Reasoning with Debate Weijie Li, Zhongqing Wang, Guodong Zhou Abstract Most commonsense reasoning models overlook the influence of personality traits, limiting their effectiveness in personalized systems such as dialogue generation. To address this limitation, we introduce the Personality-aware Commonsense Knowledge Graph (PCoKG), a structured dataset comprising 521,316 quadruples. We begin by employing three evaluators to score and filter events from the ATOMIC dataset, selecting those that are likely to elicit diverse reasoning patterns across different personality types. For knowledge graph construction, we leverage the role-playing capabilities of large language models (LLMs) to perform reasoning tasks. To enhance the quality of the generated knowledge, we incorporate a debate mechanism consisting of a proponent, an opponent, and a judge, which iteratively refines the outputs through feedback loops. We evaluate the dataset from multiple perspectives and conduct fine-tuning and ablation experiments using multiple LLM backbones to assess PCoKG’s robustness and the effectiveness of its construction pipeline. Our LoRA-based fine-tuning results indicate a positive correlation between model performance and the parameter scale of the base models. Finally, we apply PCoKG to persona-based dialogue generation, where it demonstrates improved consistency between generated responses and reference outputs. This work bridges the gap between commonsense reasoning and individual cognitive differences, enabling the development of more personalized and context-aware AI systems. Code — https://github.com/silverbeats/pcs˙v2 1 Introduction Commonsense reasoning remains a key challenge in machine intelligence ( 17 ) . To advance this capability, NLP research has developed datasets such as ATOMIC ( 14 ; 7 ) , which focus on if-then reasoning about events, including their causes ( xIntent ) and effects ( xEffect ). COMET ( 2 ) , built on ATOMIC, has been applied to emotion recognition ( 25 ) , empathetic dialogue generation ( 20 ; 13 ; 19 ) , where contextual understanding is essential. Figure 1: The upper panel shows how COMET is commonly used, while the lower panel illustrates a more realistic application incorporating personal traits. Commonsense knowledge in ATOMIC is structured as (e,r,t) triplets, where e represents an event, r an inference dimension, and t the outcome. While useful for general reasoning, this framework overlooks individual differences—particularly personality traits—that shape how people interpret and respond to events. As a result, it fails to model real-world cognitive diversity and limits its applicability in personalized AI. As illustrated in the upper part of Figure 1 , existing approaches like COMET generate plausible but generic inferences without considering personal traits. In contrast, the lower part illustrates the significant influence of individual perspectives, resulting in more diverse and personalized responses. Current frameworks, therefore, fall short in capturing the richness of human cognition. To address this limitation and model individual differences in commonsense reasoning, we introduce the Personality-aware Commonsense Knowledge Graph (PCoKG), which enhances traditional knowledge graphs by integrating person-specific personality traits. We expand the conventional (e,r,t) triplet structure into a quadruple format: (e,p,r,t) , where p represents personality information based on the Myers-Briggs Type Indicator (MBTI) ( 11 ) . This design enables more realistic modeling of how individuals with different personality types perceive and respond to the same event. For example, an introverted person may view social interactions very differently compared to an extroverted individual. By incorporating personality, our approach facilitates more personalized and context-aware inferences, enabling AI systems better to reflect the cognitive diversity inherent in human reasoning. Constructing a high-quality commonsense knowledge graph enriched with personality information is challenging. Traditional crowdsourcing methods are limited by the difficulty and cost of recruiting individuals with specific personality traits at scale. To address this, we propose a novel LLM-driven approach that simulates diverse personality types using the MBTI framework. Our method introduces two key components: an evaluator-based mechanism that filters events and inference dimensions to ensure relevance and diversity, and a role-playing prompting strategy enhanced with a multi-agent debate mechanism to improve reasoning consistency and quality. Based on this approach, we construct PCoKG — the first large-scale, personality-aware commonsense knowledge graph, which contains 521,316 high-quality (e,p,r,t) quadruples across 9 inference dimensions and all 16 MBTI types. Compared to existing resources, PCoKG offers three major advantages. First, it is the first large-scale dataset and modeling framework that explicitly incorporates personality into commonsense reasoning, enabling personalized and contextually appropriate inferences. Second, our LLM-based pipeline eliminates the need for labour-intensive human annotation, supporting scalable data expansion. Third, although the dataset is built upon the MBTI framework, the pipeline is designed to be generalizable, allowing adaptation to other personality theories or the incorporation of additional role attributes. To validate the effectiveness of our approach, we conducted comprehensive experiments using multiple LLM backbones, demonstrating the robustness and generalisation of the generated knowledge across different architectures. We further performed extensive ablation studies to assess the contributions of each component in our pipeline and analyzed performance across model sizes to understand the impact of parameter scale. Finally, we applied PCoKG to a persona-based dialogue generation task, showcasing its practical utility in real-world applications. These results validate both the effectiveness of our methodology and the value of incorporating personality into commonsense knowledge modeling. 2 Related Work 2.1 Commonsense Knowledge Bases ConceptNet ( 16 ) is a widely used commonsense knowledge base (Commonsense Knowledge Base, CKB), though its Chinese version contains relatively limited knowledge ( 9 ) . To address such limitations, 24 automatically constructed a large-scale commonsense knowledge base called TransOMCS by transforming syntactic parses of web sentences into structured triples. However, most existing CKBs primarily focus on taxonomic relationships, such as isA and Synonym ( 5 ) , which inevitably restricts their applicability. In contrast, another line of commonsense knowledge bases—represented by ATOMIC—requires human annotators to infer the causes and consequences of given events based on personal commonsense knowledge, thereby enriching the knowledge graph with tail entities ( 14 ; 7 ) . Inspired by ATOMIC’s organisational structure, 10 introduced C 3 KG, the first Chinese conversational commonsense knowledge graph built upon four types of conversational flows linking headers (events) and tails (inference results). Following the ATOMIC framework, 21 built an emotional commonsense knowledge graph. 22 constructed a Chinese dataset for personalized-aware commonsense reasoning, which is the most closely related to our work. However, their dataset was built manually, making it difficult to scale up in terms of size and coverage. 2.2 Myers-Briggs Type Indicator The Myers-Briggs Type Indicator (MBTI) ( 11 ) is a widely recognized personality model that categorizes individuals along four dichotomous dimensions: Introversion/Extraversion, Sensing/Intuition, Thinking/Feeling, and Judging/Perceiving. Despite psychometric critiques ( 1 ) , it is widely used in both professional and personal contexts. Due to its intuitive appeal, the MBTI has gained traction in computational research. Recent studies have leveraged MBTI for personality-aware natural language processing and human-computer interaction. Examples include MBTI-labeled Reddit datasets ( 6 ) , personalized emotional support systems ( 18 ) , embedding stable MBTI traits into LLMs ( 4 ) , simulating MBTI types to evaluate LLM decision-making, affective computing datasets integrating MBTI and emotion ( 26 ) , conversational MBTI and gender inference models ( 15 ) , investigations into personality-adaptive dialogue agents ( 3 ) , and analyses of MBTI-induced bias in hate speech detection ( 23 ) . Collectively, these studies highlight the potential of MBTI-aware