arxiv: 2605.04761 · v1 · submitted 2026-05-06 · 💻 cs.LG · cs.AI· cs.HC

Recognition: 3 theorem links

Cognitive Twins: Investigating Personalized Thinking Model Building and Its Performance Enhancement with Human-in-the-Loop

Wu-Yuin Hwang , Nur Alif Ilyasa , Muhammad Irfan Luthfi , Yuniar Indrihapsari

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:53 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.HC

keywords personalized thinking modelcognitive twinlearner journal analysishuman-in-the-loop refinementeducational AIMarzano taxonomysemantic abstractionLLM-based modeling

0 comments

The pith

A five-layer model built from learner journals using LLMs represents individual thinking patterns with 75 percent fidelity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces the Personalized Thinking Model as a hierarchical learner representation that structures evidence from journals into five layers based on an educational taxonomy, from specific behaviors up to core values. It describes a pipeline that processes these journals through large language model inference, embeddings, reduction, and clustering to construct the model as a cognitive twin. With data from forty participants over seven weeks, evaluations show an F1 score of 75.48 percent after human-in-the-loop refinement, user ratings of 4.30 on a five-point scale, and rising semantic coherence in higher layers. The work supports using such models to enable AI in education to align with personal cognitive processes rather than generic patterns.

Core claim

The Personalized Thinking Model organizes evidence from learner journals into a five-layer structure covering behavioral instances, behavioral patterns, cognitive routines, metacognitive tendencies, and self-system values. Grounded in Marzano's New Taxonomy of Educational Objectives, it is constructed using large language model inference combined with sentence embeddings, dimensionality reduction, and consensus clustering. Evaluations across automatic atomic information matching, user Likert ratings, and semantic alignment verification yield an overall F1 score of 75.48 percent after human-in-the-loop refinement, mean ratings of 4.30, and a pattern of increasing topic coherence from the base

What carries the argument

The five-layer PTM hierarchy that abstracts journal evidence into self-system values through LLM extraction and consensus clustering.

If this is right

PTM enables AI tutoring systems to align feedback with a learner's specific cognitive routines and tendencies rather than broad averages.
Human-in-the-loop refinement produces measurable gains in model fidelity as seen in the F1 score rise and stable user ratings.
The observed increase in semantic abstraction from lower to higher layers indicates the model successfully separates surface behaviors from deeper thinking patterns.
The pipeline supports interpretable, hierarchical learner representations that can be updated over time in educational settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Cognitive twins built this way could be tested in live tutoring loops to check whether alignment with the model improves learner outcomes over time.
The approach might extend to other personal data streams such as discussion transcripts to maintain an evolving thinking model.
If the abstraction pattern proves stable, it could guide designs for AI systems that reason at multiple cognitive depths mirroring human layers.
Iterative human feedback loops may allow the model to track changes in a learner's thinking as education progresses.

Load-bearing premise

The assumption that LLM-based extraction and clustering from journals accurately captures a learner's actual thinking model without introducing systematic biases or artifacts.

What would settle it

A study in which participants review the generated PTM layers against their original journals and report consistent mismatches at the metacognitive or value layers, or where personalized tutoring using the PTM shows no learning gains over generic AI support.

Figures

Figures reproduced from arXiv: 2605.04761 by Muhammad Irfan Luthfi, Nur Alif Ilyasa, Wu-Yuin Hwang, Yuniar Indrihapsari.

**Figure 1.** Figure 1: PTM Model Construction Pipeline Overview view at source ↗

**Figure 2.** Figure 2: Phase 1: Raw Data to L1 Behavioral Node Algorithm 1 Applied Weighted Consensus Clustering with Temporal Penalty Require: Set of L0 Instances X, threshold τ = 4 Ensure: Set of L1 Behavioral Pattern clusters C 1: Phase 1: Base Clustering 2: for each 5W1H attribute a do 3: Generate SBERT embeddings and reduce to 25 dimensions via UMAP 4: Apply HDBSCAN to generate assignments πa 5: end for 6: Phase 2: Consensu… view at source ↗

**Figure 3.** Figure 3: Phase 2: Dimension Model Generation Process view at source ↗

**Figure 4.** Figure 4: illustrates this recursive synthesis detail. This prompt acts as a generative function fID(c (n) d,p , d). Given a specific analytical dimension d and its corresponding cluster c (n) d,p (where p represents the specific cluster index), the LLM synthesizes one to three higher level interpretations that explain why these patterns appear together. Let V (n) d,p be the set of generated higher layer nodes for t… view at source ↗

**Figure 5.** Figure 5: Personalized Thinking Model (PTM) Overall Architecture view at source ↗

read the original abstract

This paper presents the Personalized Thinking Model (PTM), a hierarchical and interpretable learner representation designed for AI supported education. PTM organizes evidence from learner journals into a five-layer structure covering behavioral instances, behavioral patterns, cognitive routines, metacognitive tendencies, and self-system values. PTM is grounded in Marzano's New Taxonomy of Educational Objectives and tries to clone learner's thinking model and build cognitive twin. It was constructed using a pipeline that combines large language model inference (Gemini 2.5 Pro), sentence embeddings, dimensionality reduction, and consensus clustering. This paper evaluates PTM fidelity through three methods applied to 40 participants in a seven-week study. First, automatic evaluation using atomic information point matching yielded an overall F1 score of 74.57% before human-in-the-loop (HITL) refinement and 75.48% after refinement. Second, user evaluation using a Likert scale produced mean ratings of 4.26 and 4.30 on a five-point scale for pre and post-HITL conditions respectively. Third, semantic alignment verification showed that topic coherence increased from 0.436 at the behavioral layer to 0.626 at the core value layer, while lexical overlap with journal vocabulary decreased from 0.114 to 0.007 across those same layers. These results suggest that the PTM produces outputs with acceptable fidelity, was generally perceived by users as reflecting their thinking, and showed a pattern consistent with semantic abstraction across layers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The PTM paper layers LLM outputs from student journals into a Marzano-based hierarchy but its fidelity checks stay inside the same model loop and do not test against independent evidence of actual thinking.

read the letter

The paper introduces a five-layer Personalized Thinking Model that pulls evidence from learner journals, runs it through Gemini 2.5 Pro extraction plus embeddings and consensus clustering, and maps the results onto Marzano's taxonomy from behavioral instances up to self-system values. They added a human-in-the-loop refinement step and tested the whole thing on 40 participants across seven weeks. The pipeline itself is laid out clearly enough to follow and replicate if someone wanted to try it in an education setting. They also report three concrete checks: atomic-point F1 scores near 75 percent, average Likert ratings of 4.26–4.30, and the expected rise in topic coherence with a drop in lexical overlap as layers get more abstract. Those numbers give readers something to look at rather than pure description. The main limitation is that none of the checks escape the LLM's own interpretive frame. The F1 matching uses the same model family to score points, so it largely measures internal consistency. The Likert ratings come after users have seen the structured output, which leaves room for demand characteristics or simple preference for tidy summaries. The coherence and overlap shifts are direct results of the clustering method that forces abstraction, not independent confirmation that the layers match real thinking. Without a separate human-annotated baseline or a behavioral test that shows the twin predicts learner responses outside the original journals, the claim that this builds a faithful cognitive twin rests on an assumption that the pipeline introduces no systematic artifacts. This work sits in the AI-for-education space where people want structured ways to represent individual cognition. It has enough method and data to be worth sending to referees, though the validation section will need more external anchors before the central claim can be taken as demonstrated.

Referee Report

3 major / 3 minor

Summary. The paper introduces the Personalized Thinking Model (PTM), a five-layer hierarchical representation (behavioral instances to self-system values) grounded in Marzano's New Taxonomy, constructed from learner journals via LLM inference (Gemini 2.5 Pro), embeddings, dimensionality reduction, and consensus clustering to create 'cognitive twins.' In a seven-week study with 40 participants, it reports three evaluations: automatic atomic information point matching (F1 74.57% pre-HITL to 75.48% post), user Likert ratings (means 4.26 and 4.30), and semantic metrics (topic coherence rising 0.436 to 0.626, lexical overlap falling 0.114 to 0.007 across layers). The central claim is that PTM achieves acceptable fidelity to individual thinking models, is perceived as reflective by users, exhibits expected abstraction patterns, and benefits modestly from human-in-the-loop refinement.

Significance. If the PTM pipeline produces unbiased representations of unobserved learner cognition, the work would advance interpretable, taxonomy-grounded cognitive modeling for AI-supported education, moving beyond flat user profiles toward hierarchical 'twins.' Strengths include the explicit grounding in established educational theory, the multi-method evaluation design, and the practical HITL component. However, the significance is limited by the absence of independent human-annotated ground truth, which is required to substantiate claims of cloning actual thinking rather than LLM-mediated abstractions.

major comments (3)

[§4.2] §4.2 (Automatic Evaluation): The F1 scores (74.57% to 75.48%) are computed via atomic information point matching performed by the same LLM pipeline used for initial extraction and clustering; this measures internal consistency with the model's interpretive lens rather than independent fidelity to the learner's actual (unobserved) thinking, directly undermining the central fidelity claim without a separate human-annotated baseline.
[§4.3] §4.3 (User Evaluation): The Likert means (4.26 pre-HITL, 4.30 post) are post-exposure self-reports with no reported statistical significance tests, no comparison to control conditions (e.g., generic or shuffled taxonomies), and no controls for demand characteristics; this leaves open whether ratings reflect genuine reflection of personal thinking or presentation effects.
[§4.4] §4.4 (Semantic Alignment Verification): The reported rise in topic coherence and drop in lexical overlap are direct, expected outcomes of the abstraction-inducing consensus clustering step described in §3; they do not constitute external validation of semantic abstraction in the learner's thinking and cannot support the claim of pattern consistency with the learner's model.

minor comments (3)

[§1] The abstract and introduction use 'cognitive twin' and 'PTM' interchangeably without a clear definitional distinction; this should be clarified in §1 or §2.
[§3] Details on the exact prompts, temperature settings, and few-shot examples used with Gemini 2.5 Pro for evidence extraction are missing from §3; these are necessary for reproducibility.
[§3] The paper should report inter-annotator agreement or validation metrics for the consensus clustering step and the atomic information point extraction.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have carefully considered each major comment and provide point-by-point responses below. Where appropriate, we outline planned revisions to address the concerns while maintaining the integrity of our reported findings.

read point-by-point responses

Referee: [§4.2] §4.2 (Automatic Evaluation): The F1 scores (74.57% to 75.48%) are computed via atomic information point matching performed by the same LLM pipeline used for initial extraction and clustering; this measures internal consistency with the model's interpretive lens rather than independent fidelity to the learner's actual (unobserved) thinking, directly undermining the central fidelity claim without a separate human-annotated baseline.

Authors: We agree that the automatic evaluation relies on the same LLM pipeline and therefore primarily measures internal consistency rather than independent fidelity to unobserved learner cognition. This is a genuine limitation for substantiating claims of cloning actual thinking models. In the revised manuscript, we will update §4.2 to explicitly frame the F1 scores as an internal consistency metric, add a dedicated limitations paragraph discussing the absence of separate human-annotated ground truth, and clarify that the user evaluation and semantic metrics provide complementary (though not fully independent) support. We will also note the practical difficulties of obtaining unbiased human annotations for hierarchical cognitive structures. We maintain that the multi-method design offers useful initial evidence of PTM utility, but we will not overstate the automatic results as definitive external validation. revision: partial
Referee: [§4.3] §4.3 (User Evaluation): The Likert means (4.26 pre-HITL, 4.30 post) are post-exposure self-reports with no reported statistical significance tests, no comparison to control conditions (e.g., generic or shuffled taxonomies), and no controls for demand characteristics; this leaves open whether ratings reflect genuine reflection of personal thinking or presentation effects.

Authors: We acknowledge that the reported Likert means lack statistical tests, control conditions, and explicit discussion of demand characteristics. In the revision, we will add appropriate statistical analyses (e.g., Wilcoxon signed-rank tests for pre/post comparison and one-sample tests against the neutral midpoint) and report effect sizes. We will also expand the limitations section to address potential demand characteristics and the exploratory nature of the study, recommending future work with control groups using generic or randomized taxonomies. While the positive ratings provide preliminary indication of user-perceived reflection, we agree they cannot alone confirm genuine fidelity without such controls. revision: yes
Referee: [§4.4] §4.4 (Semantic Alignment Verification): The reported rise in topic coherence and drop in lexical overlap are direct, expected outcomes of the abstraction-inducing consensus clustering step described in §3; they do not constitute external validation of semantic abstraction in the learner's thinking and cannot support the claim of pattern consistency with the learner's model.

Authors: We recognize that the increases in topic coherence and decreases in lexical overlap are direct, expected results of the consensus clustering procedure. These metrics were intended to verify that the generated hierarchy exhibits the abstraction gradient predicted by Marzano's taxonomy, not to provide external validation of the learner's internal model. In the revised §4.4, we will reframe the presentation to clarify this purpose, emphasize that the results support structural consistency with the theoretical framework, and avoid language suggesting independent validation of the learner's thinking. This will prevent overstatement while preserving the value of demonstrating that the pipeline produces the intended hierarchical patterns. revision: yes

standing simulated objections not resolved

The absence of an independent human-annotated ground truth for the cognitive models, which would require a separate, resource-intensive annotation study not performed in the current work and cannot be retroactively added without new data collection.

Circularity Check

0 steps flagged

No significant circularity; PTM construction and evaluations are empirically grounded

full rationale

The paper constructs PTM by applying an external Marzano taxonomy via LLM extraction, embeddings, and clustering to learner journals, then measures fidelity through separate automatic atomic matching (F1), user Likert ratings, and independent semantic metrics (coherence, lexical overlap) on the same study cohort. These steps compare model outputs directly to source journals and participant perceptions without any derivation that reduces by construction to fitted parameters, self-definitions, or self-citation chains. No equations or load-bearing premises collapse into the inputs; the reported scores are straightforward empirical results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that the taxonomy applies and that LLM processing faithfully extracts the layers without major distortion.

axioms (1)

domain assumption Marzano's New Taxonomy of Educational Objectives accurately structures learner thinking into the five layers.
The model is grounded in this taxonomy as stated in the abstract.

invented entities (1)

Personalized Thinking Model (PTM) / Cognitive Twin no independent evidence
purpose: To clone learner's thinking model hierarchically from journals.
New model introduced; no independent evidence outside the paper's evaluations.

pith-pipeline@v0.9.0 · 5589 in / 1214 out tokens · 34535 ms · 2026-05-08T17:53:06.568538+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

86 extracted references · 3 canonical work pages

[1]

Aggarwal and ChengXiang Zhai

Charu C. Aggarwal and ChengXiang Zhai. A survey of text clustering algorithms. InMining Text Data, pages 77–128. Springer, 2012

2012
[2]

Power to the people: The role of humans in interactive machine learning.AI Magazine, 35(4):105–120, 2014

Saleema Amershi et al. Power to the people: The role of humans in interactive machine learning.AI Magazine, 35(4):105–120, 2014

2014
[3]

Anderson, David R

Lorin W. Anderson, David R. Krathwohl, et al.A Taxonomy for Learning, Teaching, and Assessing. Longman, New York, 2001

2001
[4]

Springer Science & Business Media, 2013

Roger Azevedo and Vincent Aleven, eds.International Handbook of Metacognition and Learn- ing Technologies. Springer Science & Business Media, 2013

2013
[5]

Hubness reduction improves clustering and trajectory inference in single- cell transcriptomic data.Bioinformatics, 38(4):1045–1051, 2022

Elise Amblard et al. Hubness reduction improves clustering and trajectory inference in single- cell transcriptomic data.Bioinformatics, 38(4):1045–1051, 2022

2022
[6]

Bennet et al

M. Bennet et al. A new framework for contextual multilayer knowledge embedding in large language models. 2024

2024
[7]

Bloom.Taxonomy of Educational Objectives

Benjamin S. Bloom.Taxonomy of Educational Objectives. Longmans Green, New York, 1956

1956
[8]

Student models that invite the learner in.International Journal of Artificial Intelligence in Education, 17(2):89–120, 2007

Susan Bull and Judy Kay. Student models that invite the learner in.International Journal of Artificial Intelligence in Education, 17(2):89–120, 2007

2007
[9]

Open learner models

Susan Bull and Judy Kay. Open learner models. InAdvances in Intelligent Tutoring Systems, pages 301–322. Springer, 2010

2010
[10]

Ricardo J. G. B. Campello et al. Density based clustering based on hierarchical density estimates. InAdvances in Knowledge Discovery and Data Mining, pages 160–172. Springer, 2013

2013
[11]

Association for Computational Linguistics, 1998

Nancy Chinchor and Beth Sundheim.MUC 7 Named Entity Task Definition. Association for Computational Linguistics, 1998

1998
[12]

How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings

Kawin Ethayarajh. How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. InProceedings of EMNLP-IJCNLP, pages 55–65, 2019

2019
[13]

Morgan & Claypool, 2015

Atefeh Farzindar and Diana Inkpen.Natural Language Processing for Social Media. Morgan & Claypool, 2015

2015
[14]

Learning analytics should not promote one size fits all: The effects of instructional conditions in predicting academic success.The Internet and Higher Education, 28:68–84, 2016

Dragan Gašević et al. Learning analytics should not promote one size fits all: The effects of instructional conditions in predicting academic success.The Internet and Higher Education, 28:68–84, 2016

2016
[15]

A comprehensive empirical comparison of hubness re- duction methods.Knowledge and Information Systems, 2019

Roman Feldbauer and Arthur Flexer. A comprehensive empirical comparison of hubness re- duction methods.Knowledge and Information Systems, 2019

2019
[16]

Gemini 2.5 Pro technical report

Google DeepMind. Gemini 2.5 Pro technical report. 2025. 23

2025
[17]

Cognitive structure generation for student modeling.Computers and Education: Artificial Intelligence, 2025

Tianlong Gu et al. Cognitive structure generation for student modeling.Computers and Education: Artificial Intelligence, 2025

2025
[18]

Interactive machine learning for health informatics.Brain Informatics, 2016

Andreas Holzinger. Interactive machine learning for health informatics.Brain Informatics, 2016

2016
[19]

Is automated topic model evaluation broken?: The incoherence of coherence.arXiv preprint arXiv:2107.02173, 2021

Alexander Hoyle et al. Is automated topic model evaluation broken?: The incoherence of coherence.arXiv preprint arXiv:2107.02173, 2021

work page arXiv 2021
[20]

Hu et al

Z. Hu et al. Adaptive grounding in large language models.arXiv, 2025

2025
[21]

Consensus clusterings

Hongfu Liu et al. Consensus clusterings. InICDM, pages 607–612. IEEE, 2017

2017
[22]

Liu et al

Nelson F. Liu et al. Lost in the middle: How language models use long contexts.Transactions of the Association for Computational Linguistics, 12:157–173, 2024

2024
[23]

Pearson, London, 2016

Rose Luckin et al.Intelligence Unleashed: An Argument for AI in Education. Pearson, London, 2016

2016
[24]

Marzano.The Art and Science of Teaching

Robert J. Marzano.The Art and Science of Teaching. ASCD, Alexandria, VA, 2007

2007
[25]

UMAP: Uniform manifold approximation and projection for dimension reduction.arXiv, 2018

Leland McInnes et al. UMAP: Uniform manifold approximation and projection for dimension reduction.arXiv, 2018

2018
[26]

arXiv preprint arXiv:2305.14251 , year=

Sewon Min et al. FActScore: Fine grained atomic evaluation.arXiv preprint arXiv:2305.14251, 2023

work page arXiv 2023
[27]

Consensus clusterings

Nam Nguyen and Rich Caruana. Consensus clusterings. InICDM, pages 607–612. IEEE, 2007

2007
[28]

Parry et al

David A. Parry et al. Named entity recognition from unstructured EMS narratives.Journal of the American Medical Informatics Association, 28(8):1732–1741, 2021

2021
[29]

Poctik et al

J. Poctik et al. Semantic abstraction in hierarchical cognitive architectures. 2025

2025
[30]

Sentence-BERT:SentenceembeddingsusingsiameseBERT- networks

NilsReimersandIryna Gurevych. Sentence-BERT:SentenceembeddingsusingsiameseBERT- networks. InEMNLP-IJCNLP, 2019

2019
[31]

Roberts et al

Margaret E. Roberts et al. Structural topic models for open-ended survey responses.American Journal of Political Science, 58(4):1064–1082, 2014

2014
[32]

Improving students’ help-seeking skills using metacognitive feedback in an intelligent tutoring system.Learning and Instruction, 21(2):267–280, 2011

Ido Roll et al. Improving students’ help-seeking skills using metacognitive feedback in an intelligent tutoring system.Learning and Instruction, 21(2):267–280, 2011

2011
[33]

Conceptual framework for autonomous cognitive entities.arXiv, 2023

David Shapiro et al. Conceptual framework for autonomous cognitive entities.arXiv, 2023

2023
[34]

Valerie J. Shute. Stealth assessment in computer-based games to support learning.Computer Games and Instruction, 55(2):503–524, 2011

2011
[35]

Takii et al

Y. Takii et al. OKLM: Ontology based knowledge and learning model. InProc. Int. Conf. Educational Data Mining, 2024

2024
[36]

Wang et al

X. Wang et al. Dynamic cognitive diagnosis with interpretable latent structure.Computers and Education, 2023. 24

2023
[37]

Philip H. Winne. Self-regulated learning viewed from models of information processing. In B. J. Zimmerman & D. H. Schunk (Eds.),Self-Regulated Learning and Academic Achievement: Theoretical Perspectives, pages 153–189. Routledge, 2001

2001
[38]

Philip H. Winne. Improving measurements of self-regulated learning.Educational Psychologist, 45(4):267–276, 2010

2010
[39]

A unified structured framework for AGI: Bridging cognition and neuromorphic computing.LNCS, 2023

Min Xu et al. A unified structured framework for AGI: Bridging cognition and neuromorphic computing.LNCS, 2023

2023
[40]

Yang et al

Y. Yang et al. Marzano’s new taxonomy in educational technology research.Educational Technology Research and Development, 2023

2023
[41]

Systematic review of research on AI applications.Int

Olaf Zawacki-Richter et al. Systematic review of research on AI applications.Int. Journal of Educational Technology, 16(1):39, 2019

2019
[42]

Ragas: Automated evaluation of retrieval augmented generation

Shahul Es et al. Ragas: Automated evaluation of retrieval augmented generation. InProceed- ings of EACL, 2024

2024
[43]

Optimizing semantic coherence in topic models

David Mimno et al. Optimizing semantic coherence in topic models. InProceedings of EMNLP, pages 262–272, 2011

2011
[44]

Exploring the space of topic coherence measures

Michael Röder et al. Exploring the space of topic coherence measures. InProceedings of WSDM, pages 399–408, 2015

2015
[45]

Borghi et al

Anna M. Borghi et al. Words as social tools: Language, body, and abstract concepts.Frontiers in Psychology, 8:1350, 2017

2017
[46]

Vallacher and Daniel M

Robin R. Vallacher and Daniel M. Wegner. What do people think they’re doing? Action identification and human behavior.Psychological Review, 94(1):3–15, 1987

1987
[47]

Why we need new evaluation metrics for NLG

Jekaterina Novikova et al. Why we need new evaluation metrics for NLG. InProceedings of EMNLP, pages 2241–2252, 2017

2017
[48]

Human-heuristic-level evaluation of LLM-generated text.arXiv preprint arXiv:2304.14376, 2023

Maurice Jakesch et al. Human-heuristic-level evaluation of LLM-generated text.arXiv preprint arXiv:2304.14376, 2023

work page arXiv 2023
[49]

Developing Eternal Learning Model Trainer System to Support Continuous Knowledge Integration and Cognitive Growth

Muhammad Irfan Luthfi and Wu-Yuin Hwang. Developing Eternal Learning Model Trainer System to Support Continuous Knowledge Integration and Cognitive Growth. In2024 21st In- ternational Joint Conference on Computer Science and Software Engineering (JCSSE), pages 1–6. IEEE, 2024. A Detailed PTM Architecture This appendix provides the full architectural diagr...

2024
[50]

WHAT : The main decision , activity , reaction , or problem - solving process
[51]

Day , HH : MM - HH : MM

WHEN : The time of the instance , sp ec if ie d as the day and hour in " Day , HH : MM - HH : MM " format ( e . g . , Friday , 10:00 -12:30) . If only the start time is available , output HH : MM and specify the duration in hours
[52]

WHERE : The specific location of the instance
[53]

WHO : Other i n d i v i d u a l s involved ; if only the user is involved , return an empty string ("")
[54]

WHY : The reason for the instance , the c o n s i d e r a t i o n s between choices , or the u n d e r l y i n g m o t i v a t i o n
[55]

WHAT ": str ,

HOW : The method or process used . Assign an indexed ID to each instance and arrange them in c h r o n o l o g i c a l order . Output the ex tr ac te d i n f o r m a t i o n using this JSON schema : Info = {{" WHAT ": str , " WHEN ": str , " WHERE ": str , " WHO ": str , " WHY ": str , " HOW ": str }} Return = {{ " i n f o r m a t i o n s ": Array < Info ...
[56]

" title ": A concise , d e s c r i p t i v e title for the pattern (3 -7 words )
[57]

content

" content ": A c o m p r e h e n s i v e pa ra gr ap h d e s c r i b i n g the pattern
[58]

s o u r c e _ i n s t a n c e s

" s o u r c e _ i n s t a n c e s ": A JSON array of the original instance IDs that support this specific pattern . - Refer to the subject as " The user ". - Do not include markdown f o r m a t t i n g or any other text in your response . - Do not mention " s o u r c e _ i n s t a n c e s " within the content ; ensure they are only listed in the " s o u r...
[59]

Analyze the provided Layer { l a y e r _ n u m b e r } patterns ( which are Layer 1 patterns )
[60]

Generate exactly { n u m _ d i m e n s i o n s } d i m e n s i o n s for each of the three layers ( L2 , L3 , L4 )
[61]

title " and a

Each di me ns io n must include a " title " and a " d e s c r i p t i o n "
[62]

Habit Analysis

The title of each di me ns io n should be general and directly reflect the focus of its layer , i n c o r p o r a t i n g relevant keywords ( e . g . , " Habit Analysis " , " Goal P r i o r i t i z a t i o n " , " Core Value I d e n t i f i c a t i o n ") . The d e s c r i p t i o n should explain how this general lens applies s p e c i f i c a l l y to t...
[63]

L2 " ,

The output must be a single , valid JSON object with the keys " L2 " , " L3 " , and " L4 "
[64]

L2 ": [ {{

Do not include markdown formatting , headers , or any c o n v e r s a t i o n a l text . Output Schema : {{ " L2 ": [ {{ " title ": " string " , " d e s c r i p t i o n ": " string " }} , {{ " title ": " string " , " d e s c r i p t i o n ": " string " }} ] , " L3 ": [ {{ " title ": " string " , " d e s c r i p t i o n ": " string " }} , {{ " title ": " s...
[65]

Review the list of source nodes provided below
[66]

Group these nodes into { n u m _ c l u s t e r s } distinct clusters based on their r e l a t i o n s h i p to the a n a l y t i c a l di me ns io n
[67]

Triggers

Every node within a cluster must share a specific c o m m o n a l i t y re ga rd in g the di m en si on . For example , if the d im en si on is " Triggers " , clusters could be " Social Triggers " and " Academic Triggers ". 29
[68]

Ensure each cluster contains at least two nodes
[69]

clusters

A node may belong to multiple clusters if applicable , but aim for distinct gr o up in gs . Output Schema : The output must be a single JSON object c o n t a i n i n g a list of clusters : {{ " clusters ": [ {{ " c l u s t e r _ l a b e l ": " string ( A short d e s c r i p t i v e label for this group ) " , " n o d e _ i n d i c e s ": [1 , 5 , 8] ( The ...
[70]

Cluster Patterns

Analyze the provided " Cluster Patterns "
[71]

S y n t h e s i z e one to three (1 -3) distinct , profound insights that explain why these patterns are grouped together under this d im en si on
[72]

If they are simple , one insight is s u f f i c i e n t

If the patterns are complex , split them into distinct insights . If they are simple , one insight is s u f f i c i e n t
[73]

The user relies on external o r g a n i z a t i o n a l support to mitigate anxiety

Each insight should be abstract and model - level . For example , use " The user relies on external o r g a n i z a t i o n a l support to mitigate anxiety " instead of " The user uses a calendar "
[74]

title ":

Both the title and content must be clear and use a c c e s s i b l e language . Avoid complex or te ch ni ca l v o c a b u l a r y ; ensure the output is easy to read and u n d e r s t a n d . Output Schema : The output must be a single JSON array of insight objects : [ {{ " title ": " string (3 -7 words , abstract and p r o f e s s i o n a l ) " , " cont...
[75]

What does the user ty pi ca ll y do on weekday mornings ?

Generate qu es ti on s that cover a diverse range of topics : 31 - Routines and Habits : e . g . , " What does the user ty pi ca ll y do on weekday mornings ?" - P r e f e r e n c e s : e . g . , " What kind of food does the user prefer in the evening while at school ?" - P r i o r i t i e s : e . g . , " When the user has both homework and a midterm exam...
[76]

The ground truth answer must be directly su pp or te d by the provided journal text
[77]

query ":

The output must be a single , valid JSON array . Do not include markdown formatting , headers , or any c o n v e r s a t i o n a l text . Output Schema ( A JSON Array ) : [ {{ " query ": " string ( The question about the user ) " , " g r o u n d _ t r u t h ": " string ( The factual answer from the text ) " }} ] USER ’ S JOURNAL ENTRIES : { j o u r n a l ...
[78]

Context

Source Material Only : Answer the user ’ s question using only the i n f o r m a t i o n provided in the " Context " section below . Do not use your own internal knowledge , even if you believe the i n f o r m a t i o n is correct
[79]

I cannot answer this based on the provided context

No H a l l u c i n a t i o n s : If the answer cannot be found within the provided context , state : " I cannot answer this based on the provided context ." Do not invent facts or attempt to guess . Context : { I N S E R T _ R E T R I E V E D _ C O N T E X T _ H E R E } --- User Query : { I N S E R T _ U S E R _ Q U E S T I O N _ H E R E } Listing 8: Prom...
[80]

U n d e r s t a n d the Context : - Ca re fu ll y read the query to u n d e r s t a n d the context and re le va nc e of the answers

Showing first 80 references.