Skip to main content

Resultados relacionados

Resultados totales (Incluyendo duplicados): 3
Encontrada(s) 1 página(s)

UPCommons. Portal del coneixement obert de la UPC
Publication Artículo científico (JournalArticle)

Intentional policy graphs: A pipeline for explaining agent behavior through intentions

  • Giménez Ábalos, Víctor
  • Álvarez Napagao, Sergio|||0000-0001-9946-9703
  • Tormos Llorente, Adrián
  • Montese, Sara
  • Cortés García, Claudio Ulises|||0000-0003-0192-3096
  • Vázquez Salceda, Javier|||0000-0003-1732-9446
Agents increasingly operate in complex environments, where coherent behavior often emerges from opaque decision-making processes. While such systems can be highly effective, this lack of transparency limits trust, auditing, and meaningful human understanding. We introduce intentional policy graphs, a post hoc, model- agnostic framework that explains agent behavior in terms of intentions: probabilistic commitments to desired outcomes inferred from partial observations. By extending policy graphs with a formal notion of intention, we move beyond action-level descriptions toward telic explanations of why agents pursue particular trajectories. The framework provides a complete construction pipeline, design principles, and quantitative metrics that explicitly characterize the trade-off between interpretability and reliability. Intentions support structured answers to what, how, and why questions, enabling both local and global explanations of behavior. We demonstrate the approach in a cooperative multi-agent game and on real-world human driving data, highlighting its generality and explanatory power without access to internal reasoning models., This work has been partially supported by the AIXPERT (grant agreement ID: 101214389), AI4CCAM (grant agreement ID: 101076911), and HumanE-AI- Net (grant agreement ID: 952026) European projects and fellowships to V.G.-A., A.T., and S.M. within the ‘‘Generacio´n D’’ initiative, Red.es, MTDFP, for talent attraction (C005/24-ED CV1). This work was partially funded by the European Union NextGenerationEU funds, through PRTR., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

Boldly propose, carefully verify: LLMs in agents should be used only for abduction

  • Sabaté Iturgaiz, Eneko
  • Giménez Ábalos, Víctor
  • Tormos Llorente, Adrián
  • Miró López-Feliu, Oriol
  • Álvarez Napagao, Sergio|||0000-0001-9946-9703
The appearance of Large Language Models (LLMs) has transformed the field of agents with new architectures and inference mechanisms. These systems, however, struggle with capabilities such as rationality or coherent behaviour. Stand-alone LLM-based architectures are not capable of reliably performing planning to produce rational behaviour. On the other hand, traditional agents have longstanding knowledge to tackle these capabilities, but struggle with tasks requiring autonomous abstractions and knowledge abduction, making some domains unattainable: these agents do not have full reasoning capabilities due to failing at abduction. While abduction is computationally difficult, Large Language Models (LLMs) excel at producing probable hypotheses without many requirements. Moreover, verifying these hypotheses via deductive knowledge and interaction with the surroundings is within the capacity of traditional agents, and the feedback can establish a virtuous loop with abduction mechanisms. In this paper, we propose to work toward hybrid systems integrating two stand-alone LLM-based abductive modules: an Abductive Reasoner Module to generate hypotheses based on detected discrepancies between agent beliefs and environment behaviour, and an Experiment Designer Module to generate goals testing these hypotheses. These modules could be integrated with cognitive architectures using means-ends reasoning to fulfil goals (tests and system-goals). We illustrate this perspective through a practical case study in Baba Is You, a complex rule-learning environment, empirically showing that abductive-deductive separation is viable even with minimal agent design., This work has been partially supported by the AIXPERT (Grant agreement ID: 101214389) Horizon Europe project and V. Giménez Abalos and A. Tormos fellowships within the “Generación D” initiative, Red.es, MTDFP, for talent attraction (C005/24-ED CV1). Funded by the European Union NextGenerationEU funds, through PRTR., Peer Reviewed
Recercat. Dipósit de la Recerca de Catalunya
Publication Artículo científico (JournalArticle)

Intentional policy graphs: A pipeline for explaining agent behavior through intentions

  • Giménez Ábalos, Víctor
  • Álvarez Napagao, Sergio
  • Tormos Llorente, Adrián
  • Montese, Sara
  • Cortés García, Claudio Ulises
  • Vázquez Salceda, Javier
Agents increasingly operate in complex environments, where coherent behavior often emerges from opaque decision-making processes. While such systems can be highly effective, this lack of transparency limits trust, auditing, and meaningful human understanding. We introduce intentional policy graphs, a post hoc, model- agnostic framework that explains agent behavior in terms of intentions: probabilistic commitments to desired outcomes inferred from partial observations. By extending policy graphs with a formal notion of intention, we move beyond action-level descriptions toward telic explanations of why agents pursue particular trajectories. The framework provides a complete construction pipeline, design principles, and quantitative metrics that explicitly characterize the trade-off between interpretability and reliability. Intentions support structured answers to what, how, and why questions, enabling both local and global explanations of behavior. We demonstrate the approach in a cooperative multi-agent game and on real-world human driving data, highlighting its generality and explanatory power without access to internal reasoning models., This work has been partially supported by the AIXPERT (grant agreement ID: 101214389), AI4CCAM (grant agreement ID: 101076911), and HumanE-AI- Net (grant agreement ID: 952026) European projects and fellowships to V.G.-A., A.T., and S.M. within the ‘‘Generacio´n D’’ initiative, Red.es, MTDFP, for talent attraction (C005/24-ED CV1). This work was partially funded by the European Union NextGenerationEU funds, through PRTR., Peer Reviewed, Postprint (published version)