pith. sign in

arxiv: 2605.30632 · v1 · pith:NTISFUJZnew · submitted 2026-05-28 · 💻 cs.HC · cs.AI· cs.LG

Rationalize: Shared Semantic Reasoning for Human-AI Alignment

Aritra Dasgupta , Naga Datha Saikiran Battula , Avina Nakarmi , Sohom Sen , Subhodeep Ghosh , Xun Song This is my paper

Pith reviewed 2026-06-29 05:10 UTC · model grok-4.3

classification 💻 cs.HC cs.AIcs.LG
keywords human-AI alignmentrole pairsshared semantic reasoningdata-driven sensemakingbidirectional alignmentrationalizationcritical thinking
0
0 comments X

The pith

Rationalize structures human-AI collaboration as complementary role pairs in a shared reasoning space to align on intent and action rather than outputs alone.

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

The paper presents Rationalize as a framework that treats human-AI interaction in data-driven sensemaking as a set of four role pairs: Explorer-Guide, Investigator-Informant, Teacher-Student, and Judge-Advocate. These pairs work inside one shared space where both humans and AI models must state their purposes, questions, assumptions, evidence, inferences, and implications. The goal is alignment at the level of how each side rationalizes its intent and next actions. A reader would care because most current alignment methods match final outputs without addressing whether the underlying reasoning steps are mutually understood.

Core claim

Rationalize conceptualizes human-AI interaction as complementary role pairs operating in a shared reasoning space. In this space humans and AI models make purposes, questions, assumptions, evidence, inferences, and implications explicit, enabling alignment not only at the output level but at the level of rationalization of intent and action by each side. The framework relates these pairs to the bidirectional human-AI alignment problem and outlines element-level and role-specific methods for design and assessment.

What carries the argument

The Rationalize role-pair framework, which defines four complementary pairs that operate inside one shared semantic reasoning space to force explicit articulation of reasoning elements.

If this is right

  • Alignment becomes bidirectional and role-dependent rather than one-directional.
  • Assessment of alignment can shift from final outputs to element-level checks on reasoning components.
  • Design of human-AI systems can incorporate role-specific protocols for making reasoning explicit.
  • A research agenda emerges that combines role-pair definitions with measurement of shared rationalization.

Where Pith is reading between the lines

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

  • The same role-pair structure could be tested in domains outside data analysis, such as collaborative planning or education.
  • Explicit role assignment might reduce cases where an AI produces correct answers for the wrong reasons that a human cannot detect.
  • The framework suggests a practical way to operationalize critical-thinking steps inside interactive AI tools without adding new model training.

Load-bearing premise

That defining human-AI interaction through these specific role pairs and requiring both sides to state their reasoning elements will produce alignment on intent and action beyond what output matching alone achieves.

What would settle it

A controlled comparison in a data sensemaking task where one group uses the role-pair structure and the other does not, measuring whether the groups show higher mutual agreement on stated purposes, assumptions, and inferences.

read the original abstract

We introduce Rationalize, a role-pair framework for shared semantic reasoning between humans and AI models in data-driven sensemaking. Building on ideas in human-machine teaming and critical thinking, we conceptualize human-AI interaction as a series of complementary role pairs (Explorer-Guide, Investigator-Informant, Teacher-Student, Judge-Advocate) operating in a shared reasoning space. In this space, human analysts and AI models (such as LLMs) make purposes, questions, assumptions, evidence, inferences, and implications explicit, facilitating alignment not only at the output level but at the level of rationalization of intent and action by each side. We relate these role pairs to the bidirectional human-AI alignment framework, illustrating how "aligning AI to humans" and "aligning humans to AI" differ by role, and sketch a collaborative research agenda for alignment design and assessment using element-level and role-specific approaches.

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.

Referee Report

2 major / 1 minor

Summary. The paper introduces 'Rationalize', a role-pair framework for shared semantic reasoning between humans and AI models in data-driven sensemaking. It conceptualizes interaction via four complementary role pairs (Explorer-Guide, Investigator-Informant, Teacher-Student, Judge-Advocate) operating in a shared reasoning space where purposes, questions, assumptions, evidence, inferences, and implications are made explicit. This is claimed to enable alignment at the level of rationalization of intent and action (beyond outputs), with relations drawn to bidirectional human-AI alignment and a sketched research agenda for design and assessment.

Significance. If the proposed mechanism can be shown to yield measurable alignment improvements, the framework could usefully structure human-AI teaming research by integrating critical-thinking elements with role complementarity. The explicit distinction between 'aligning AI to humans' and 'aligning humans to AI' by role is a clear conceptual contribution. The manuscript provides no empirical data, formal derivation, or worked examples, so significance remains prospective.

major comments (2)
  1. [Abstract] Abstract: the central claim that explicit role-pair reasoning in the shared space 'facilitating alignment not only at the output level but at the level of rationalization of intent and action by each side' is presented without any argument, concrete example, or mechanism demonstrating why or how the six elements produce this outcome over baseline interaction.
  2. [Bidirectional alignment discussion] The bidirectional alignment section: the statement that the role pairs 'illustrat[e] how "aligning AI to humans" and "aligning humans to AI" differ by role' is asserted without any mapping, table, or worked scenario showing role-specific differences in alignment direction or outcomes.
minor comments (1)
  1. The six elements of the shared reasoning space (purposes, questions, assumptions, evidence, inferences, implications) are listed but not defined or operationalized, which could lead to inconsistent application across role pairs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments correctly identify that the central claims would benefit from more concrete support. We address each major comment below and have revised the manuscript to incorporate illustrative examples and mappings.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that explicit role-pair reasoning in the shared space 'facilitating alignment not only at the output level but at the level of rationalization of intent and action by each side' is presented without any argument, concrete example, or mechanism demonstrating why or how the six elements produce this outcome over baseline interaction.

    Authors: We agree that the abstract states this outcome without elaboration or demonstration. The body describes the shared reasoning space and the six elements, but does not include a concrete example contrasting the process with baseline interaction. In the revised manuscript, we add a short worked example using the Explorer-Guide role pair to show how explicit rationalization of the six elements produces alignment at the level of intent, distinct from output-only exchange. revision: yes

  2. Referee: [Bidirectional alignment discussion] The bidirectional alignment section: the statement that the role pairs 'illustrat[e] how "aligning AI to humans" and "aligning humans to AI" differ by role' is asserted without any mapping, table, or worked scenario showing role-specific differences in alignment direction or outcomes.

    Authors: We acknowledge that the differences were asserted narratively without an explicit mapping or scenario. The revised version adds a table that maps each role pair to the two alignment directions, specifying the distinct mechanisms and outcomes for 'aligning AI to humans' versus 'aligning humans to AI'. A brief scenario is also included to illustrate role-specific differences. revision: yes

Circularity Check

0 steps flagged

No circularity: conceptual framework introduced independently without self-referential reduction

full rationale

The paper proposes Rationalize as a new role-pair framework (Explorer-Guide etc.) for explicit shared reasoning in human-AI sensemaking, building on external ideas in human-machine teaming and critical thinking. No equations, fitted parameters, predictions, or uniqueness theorems appear; the alignment claim is presented as a direct descriptive consequence of making purposes/questions/assumptions explicit rather than derived by construction from the framework itself. No self-citations are load-bearing, and the proposal remains self-contained as an independent conceptual contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests entirely on newly introduced conceptual structures with no supporting derivations, data, or external benchmarks; the role pairs and shared reasoning space are postulated without independent evidence.

axioms (1)
  • domain assumption Human-AI interaction in sensemaking can be usefully modeled as complementary role pairs operating in a shared reasoning space.
    This premise underpins the entire framework and is invoked in the abstract's description of the role pairs.
invented entities (2)
  • Rationalize framework no independent evidence
    purpose: To structure shared semantic reasoning for human-AI alignment
    Newly defined in the paper as the core contribution.
  • Role pairs (Explorer-Guide, Investigator-Informant, Teacher-Student, Judge-Advocate) no independent evidence
    purpose: To operationalize complementary interactions in the shared reasoning space
    Invented entities introduced to define the framework.

pith-pipeline@v0.9.1-grok · 5711 in / 1382 out tokens · 32377 ms · 2026-06-29T05:10:02.273341+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 9 canonical work pages

  1. [1]

    Gerile Aodeng, Guozheng Li, Yunshan Feng, Qiyang Chen, Yu Zhang, and Chi Harold Liu. 2025. InReAcTable: LLM-powered Interactive Visual Data Story Construction from Tabular Data. InProceedings of the 38th Annual ACM Symposium on User Interface Software and Technology (UIST ’25). Associa- tion for Computing Machinery, New York, NY, USA, Article 139, 16 page...

    work page doi:10.1145/3746059.3747719 2025

  2. [2]

    In Kwon Choi, Nirmal Kumar Raveendranath, Jared Westerfield, and Khairi Reda. 2019. Visual (dis)Confirmation: Validating Models and Hypotheses with Visualizations. In2019 23rd International Conference Information Visualisation (IV). IEEE, 116–121. doi:10.1109/IV.2019.00028

    work page doi:10.1109/iv.2019.00028 2019

  3. [3]

    Yi Guo, Danqing Shi, Mingjuan Guo, Yanqiu Wu, Nan Cao, and Qing Chen. 2024. Talk2Data: A Natural Language Interface for Exploratory Visual Analysis via Question Decomposition.ACM Trans. Interact. Intell. Syst.14, 2, Article 8 (April 2024), 24 pages. doi:10.1145/3643894

    work page doi:10.1145/3643894 2024

  4. [4]

    Gary Klein, Jennifer K Phillips, Erica L Rall, and Deborah A Peluso. 2007. A data–frame theory of sensemaking. InExpertise out of context. Psychology Press, 118–160

    2007

  5. [5]

    Doris Jung-Lin Lee, Abdul Quamar, Eser Kandogan, and Fatma Özcan. 2021. Boomerang: Proactive Insight-Based Recommendations for Guiding Conver- sational Data Analysis. InProceedings of the 2021 International Conference on Management of Data(Virtual Event, China)(SIGMOD ’21). Association for Com- puting Machinery, New York, NY, USA, 2750–2754. doi:10.1145/3...

    work page doi:10.1145/3448016.3452748 2021

  6. [6]

    Michael Xieyang Liu, Tongshuang Wu, Tianying Chen, Franklin Mingzhe Li, Aniket Kittur, and Brad A Myers. 2024. Selenite: Scaffolding Online Sensemak- ing with Comprehensive Overviews Elicited from Large Language Models. In CHI ’26 BiAlign Workshop, April 15, 2026, Barcelona, Spain Dasgupta et al. Proceedings of the 2024 CHI Conference on Human Factors in ...

    work page doi:10.1145/3613904.3642149 2024

  7. [7]

    Cassandra Overney, Belén Saldías, Dimitra Dimitrakopoulou, and Deb Roy. 2024. SenseMate: An Accessible and Beginner-Friendly Human-AI Platform for Quali- tative Data Analysis. InProceedings of the 29th International Conference on Intelli- gent User Interfaces(Greenville, SC, USA)(IUI ’24). Association for Computing Machinery, New York, NY, USA, 922–939. d...

    work page doi:10.1145/3640543.3645194 2024

  8. [8]

    2019.The Miniature Guide to Critical Thinking Concepts and Tools

    Richard Paul and Linda Elder. 2019.The Miniature Guide to Critical Thinking Concepts and Tools. Rowman & Littlefield

    2019

  9. [9]

    Peter Pirolli and Stuart Card. 2005. The sensemaking process and leverage points for analyst technology as identified through cognitive task analysis. In Proceedings of international conference on intelligence analysis, Vol. 5. McLean, VA, USA, 2–4

    2005

  10. [10]

    Hua Shen, Tiffany Knearem, et al. 2024. Towards Bidirectional Human-AI Align- ment: A Systematic Review for Clarifications, Framework, and Future Directions. arXiv preprint arXiv:2406.09264(2024)

    work page arXiv 2024

  11. [11]

    Arjun Srinivasan and Vidya Setlur. 2021. Snowy: Recommending Utterances for Conversational Visual Analysis. InThe 34th Annual ACM Symposium on User Interface Software and Technology(Virtual Event, USA)(UIST ’21). Association for Computing Machinery, New York, NY, USA, 864–880. doi:10.1145/3472749. 3474792

    work page doi:10.1145/3472749 2021

  12. [12]

    Thomas and Kristin A

    James J. Thomas and Kristin A. Cook. 2006. A Visual Analytics Agenda.IEEE Computer Graphics and Applications26, 1 (2006), 10–13

    2006

  13. [13]

    University of Louisville. 2009. Paul-Elder Critical Thinking Framework. https: //louisville.edu/ideastoaction/about/criticalthinking/framework. Accessed: 2026- 02-18

    2009

  14. [14]

    Lei Wang, Songheng Zhang, Yun Wang, Ee-Peng Lim, and Yong Wang. 2023. LLM4Vis: Explainable Visualization Recommendation using ChatGPT. InProceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, Mingxuan Wang and Imed Zitouni (Eds.). Association for Compu- tational Linguistics, Singapore, 675–692. doi:10.186...

    work page doi:10.18653/v1/2023.emnlp-industry.64 2023

  15. [15]

    John Wenskovitch, Corey Fallon, Kate Miller, and Aritra Dasgupta. 2021. Beyond Visual Analytics: Human-Machine Teaming for AI-Driven Data Sensemaking. In 2021 IEEE Workshop on TRust and EXpertise in Visual Analytics (TREX). IEEE

    2021

  16. [16]

    Wickens, Benjamin A

    Christopher D. Wickens, Benjamin A. Clegg, Angela Z. Vieane, and Angelique L. Sebok. 2015. Complacency and Automation Bias in the Use of Imperfect Automa- tion.Human Factors57, 5 (2015), 728–739

    2015