Cognitive Alignment in Personality Reasoning: Leveraging Prototype Theory for MBTI Inference

Haoyuan Li, Yuanbo Tong, Yuchen Li, Zirui Wang, Chunhou Liu, Jiamou LiuSubmission to ICLR 2026 · Desk rejected

TL;DR

Current NLP models treat personality detection as a rigid, hard-label classification problem, ignoring how humans actually categorize traits using central exemplars. ProtoMBTI bridges this gap by grounding Myers-Briggs Type Indicator (MBTI) prediction in cognitive prototype theory, using a lightweight encoder, a verified prototype bank, and a "retrieve–reuse–revise–retain" inference cycle. This psychologically aligned approach substantially outperforms existing methods on both 4-dimension and fine-grained 16-type prediction tasks while providing interpretable case-based rationales.

What Problem Does This Paper Address?

Predicting personality from unstructured text (such as social media posts) is valuable for personalized agents, recommendation engines, and adaptive tutors. Most prior approaches treat MBTI classification as flat, binary labeling across four axes: Introversion/Extraversion (I/E), Sensing/Intuition (S/N), Thinking/Feeling (T/F), and Judging/Perceiving (J/P).

However, this formulation suffers from three major flaws:

  1. Psychological Misalignment: Human cognition categorizes fuzzy concepts via prototypes (central, highly representative exemplars) rather than rigid rule sets.
  2. Severe Class Imbalance: Online self-reporting is heavily skewed toward intuitive, introverted types (e.g., INFP, INFJ, INTP), while sensor types (e.g., ESFJ, ESTJ) comprise less than 1–2% of standard corpora.
  3. Loss of Fine-Grained Fidelity: Prior work largely evaluates models by averaging across the four binary axes. Full 16-type prediction—which introduces significant boundary confusion between adjacent personalities (e.g., ENTP vs. ENTJ)—remains notoriously difficult and largely overlooked.

Methods

ProtoMBTI operationalizes prototype theory within a three-stage pipeline:

[Raw Posts] ──► [Balanced LLM Augmentation] ──► [4D Classifier Filter] ──► [LoRA Encoder (≤2B)]
                                                                                  │
                                                                           [Prototype Bank]
                                                                                  │
 [Test Post] ──► [Retrieve Top-k] ──► [LLM Case-Based Inference] ──► [Adaptive Retain]
  1. Class-Balanced & Multi-Dimensional Augmentation: The authors balance the dataset by prompting LLMs to generate style-consistent synthetic posts for underrepresented types. A trained 4D Classifier (an encoder with four dichotomy heads and one full-type head) acts as a strict gatekeeper: a synthetic post is kept only if all four dichotomies and the overall 16-type prediction match the target label. Survived posts undergo semantic, linguistic, and sentiment enrichment.
  2. Prototype Bank Construction: A compact (≤2B parameter) language model (e.g., Qwen2.5-1.5B or LLaMA-1B) is fine-tuned with LoRA under 16-type cross-entropy loss to map texts into a discriminative embedding space. Every verified instance is saved as a prototype triplet: $\langle \text{text}, \text{embedding}, \text{label} \rangle$.
  3. Retrieve–Reuse–Revise–Retain Reasoning: At test time:
  4. Retrieve: The query text is embedded, and the top-$k$ most similar prototypes are pulled via cosine similarity.
  5. Reuse & Revise: An LLM inspects the query alongside the retrieved prototypes, evaluating tone, logicality, and emotionality in an in-context case-based reasoning prompt.
  6. Retain: Verified correct predictions are embedded and appended back into the prototype library, allowing dynamic, lifelong adaptation.

Datasets

The authors evaluate on two established benchmark corpora:

  • Kaggle MBTI: 8,675 users from PersonalityCafe, with each user represented by 50 sampled posts.
  • Pandora: 9,067 Reddit users, featuring a broader and more linguistically varied distribution.

Both datasets are split 8:1:1 into training, validation, and test splits. Crucially, data augmentation is applied exclusively to train and validation sets, leaving test splits with their original, highly skewed distributions to benchmark genuine generalization under real-world distribution shift.

Key Insights and Ideas

  • Prototypes Must Be Typical, Not Just Semantically Close: In ablation experiments, retrieving random prototypes (50.77% 16-type accuracy) or retrieving via standard semantic similarity without personality-space encoding (50.15%) performed worse than using no prototypes at all (54.15%). Irrelevant exemplars act as distractor noise, demonstrating that prototype typicality (graded membership) is foundational.
  • The Sweet Spot of Context ($k=3$): Sweeping $k \in [1, 9]$ revealed an inverted-U curve peaking at $k=3$ (59.38%). Single exemplars ($k=1$) make predictions brittle, whereas large sets ($k \ge 7$) introduce peripheral outliers that blur boundary distinctions.
  • 16-Type Metric as the True Stress Test: While dichotomy averages mask errors, full 16-type accuracy plummeted by ~26 percentage points when prototype reasoning was removed, underscoring that exemplar matching is vital for resolving cross-dichotomy boundary ambiguities.

Key Results

  • Single-Dataset SOTA: On Kaggle, ProtoMBTI (using Qwen) achieved 85.14% average 4-dichotomy accuracy (vs. previous SOTA ETM at 77.79%) and 71.42% on direct 16-type classification (compared to ETM's theoretical 35.89%). On Pandora, ProtoMBTI achieved 71.41% average dichotomy accuracy and 60.22% 16-type accuracy (vs. ETM's 65.77% and 18.68%).
  • Multi-Source Training & Generalization: When trained on a mixed Kaggle–Pandora corpus, ProtoMBTI reached 85.54% (Kaggle) and 92.13% (Pandora) 16-type accuracy.
  • Cross-Domain Transfer: In swapped-domain tests (training/validating across different distributions), the model preserved strong generalization, yielding 81.23% on Kaggle and 81.15% on Pandora for 16-type categorization.

Limitations and Future Work

  • Reliance on Synthetic Augmentation: Achieving class balance requires LLM data synthesis, which introduces potential model-specific stylistic artifacts despite classifier filtering.
  • Baseline Comparability on 16 Types: Because earlier works only evaluated individual binary dichotomies, the paper's comparison to baselines on 16-type classification relies on computing the product of the four individual dichotomy accuracies rather than comparing against direct multi-class implementations.
  • Scope of Personality Models: The work focuses on the categorical MBTI taxonomy; extending prototype-driven reasoning to dimensional models (e.g., the Big Five trait scales) and multimodal inputs (audio/video) remains open for future research.

Summary updated

Abstract

Personality recognition from text is typically cast as hard-label classification, which obscures the graded, prototype-like nature of human personality judgments. We present \textsc{ProtoMBTI}, a cognitively aligned framework for MBTI inference that operationalizes prototype theory within an LLM-based pipeline. First, we construct a balanced, quality-controlled corpus via LLM-guided multi-dimensional augmentation (semantic, linguistic, sentiment). Next, we LoRA-fine-tune a lightweight ($\leq$2B) encoder to learn discriminative embeddings and to standardize a bank of ``personality prototypes''. At inference, we retrieve top-$k$ prototypes for a query post and perform a retrieve–reuse–revise–retain cycle: the model aggregates prototype evidence via prompt-based voting, revises when inconsistencies arise, and, upon correct prediction, retains the sample to continually enrich the prototype library. Across Kaggle and Pandora benchmarks, \textsc{ProtoMBTI} improves over baselines on both the four MBTI dichotomies and the full 16-type task, and exhibits robust cross-dataset generalization. Our results indicate that aligning the inference process with psychological prototype reasoning yields gains in accuracy, interpretability, and transfer for text-based personality modeling.
BibTeX
@misc{li2025cognitive,
  author = {Haoyuan Li and YUANBO TONG and Yuchen Li and Zirui Wang and Chunhou Liu and Jiamou Liu},
  title = {{Cognitive Alignment in Personality Reasoning: Leveraging Prototype Theory for MBTI Inference}},
  year = {2025},
  eprint = {2511.00115},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.00115},
}
Figure