TY - JOUR
T1 - Reasoning Alignment For Agentic Ai
T2 - Argumentation, Belief Revision, And Dialogue
AU - Rienstra, Tjitze
AU - Van der Torre, Leendert
AU - Yu, Liuwen
N1 - Funding Information:
L. van der Torre and L. Yu acknowledge the financial support of the Luxembourg National Research Fund (FNR). L. van der Torre is supported through the projects The Epistemology of AI Systems (EAI) (C22/SC/17111440), DJ4ME \u2013 A DJ for Machine Ethics: the Dialogue Jiminy (O24/18989918/DJ4ME), and Logical Methods for Deontic Explanations (LoDEx) (INTER/DFG/23/17415164/LoDEx). L. van der Torre and L. Yu are supported through the project Symbolic and Explainable Regulatory AI for Finance Innovation (SERAFIN) (C24/IS/19003061/SERAFIN). L. Yu is additionally supported by the University of Luxembourg\u2019s Marie Speyer Excellence Grant (2024) for the project Formal Analysis of Discretionary Reasoning (MSE-DISCREASON).
Publisher Copyright:
© 2025, College Publications. All rights reserved.
PY - 2025/10/1
Y1 - 2025/10/1
N2 - Agentic AI-deployed as technical systems that perceive, decide, and act via tools-faces requirements of safety, accountability, controlled adaptivity, and compositionality. We develop Reasoning Alignment Diagrams (RADs), commutative reasoning representations that align a source specification with an argumentation-based explanation route. As illustrative examples, we first show that full-meet belief base revision admits an exact representation within base argumentation via a restricted-attack construction: the revised base equals the intersection of the premises appearing in all stable extensions of the modified framework. This yields a RAD from input to sanctioned output that doubles as an explanation engine. We then compose the "listen" (revision) and "assert" (inference) RADs to model dialogue among agents, enabling explainable and auditable autonomy. Although our results are entirely symbolic, the RAD template can serve as a specification layer even when other components are opaque or learned. The approach realizes core themes of Gabbay's programme-logic as a toolbox, combining logics, and argumentation as a host formalism-and supports a principle-based analysis of correctness, transparency, and modularity.
AB - Agentic AI-deployed as technical systems that perceive, decide, and act via tools-faces requirements of safety, accountability, controlled adaptivity, and compositionality. We develop Reasoning Alignment Diagrams (RADs), commutative reasoning representations that align a source specification with an argumentation-based explanation route. As illustrative examples, we first show that full-meet belief base revision admits an exact representation within base argumentation via a restricted-attack construction: the revised base equals the intersection of the premises appearing in all stable extensions of the modified framework. This yields a RAD from input to sanctioned output that doubles as an explanation engine. We then compose the "listen" (revision) and "assert" (inference) RADs to model dialogue among agents, enabling explainable and auditable autonomy. Although our results are entirely symbolic, the RAD template can serve as a specification layer even when other components are opaque or learned. The approach realizes core themes of Gabbay's programme-logic as a toolbox, combining logics, and argumentation as a host formalism-and supports a principle-based analysis of correctness, transparency, and modularity.
KW - STRUCTURED ARGUMENTATION
KW - FIBERED SEMANTICS
KW - LOGIC
KW - CONTRACTION
M3 - Article
SN - 2631-9810
VL - 12
SP - 1683
EP - 1712
JO - Journal of Applied Logics
JF - Journal of Applied Logics
IS - 6
ER -