Published :- January, 2026

Download the pdf here

preprint

Abstract

As current systems scale towards more general capabilities, performance gains on benchmarks are yet to translate into robust , long-horizon reasoning, coherent world models , or reliable alignment with human values. This gap suggests that Artificial General Intelligence (AGI) is not merely a question of scaling and subsequent emergent abilities of scaling, but of structure. In this paper, we argue that progress towards a human-centered AGI requires a multi-layered computational substrate that integrates technical design (the technical layer), epistemic reliability (epistemic layer) and human oversight (human layer) as co-evolving components of a single system. We introduce a layered perspective on AGI development with significant emphasis on the technical layer. This layer governs learning dynamics and representations. Here, architectural choices determine representational coherence, learning dynamics and long-term stability. We also analyze the limitations of scaling , the inability of models to continually learn and objectives of long-horizon planning and context. Drawing inspiration from cognitive science , we argue that human cognition should serve as an inductive prior rather than an engineering template that can shape representations, reasoning and accomodates hierarchial abstractions . Beyond technical performance, we situate the proposed framework with epistemic and human social contexts., highlighting risks such as misgeneralization, dependence on reward-centric optimization, etc. We contend that alignment cannot be achieved solely through post-hoc finetuning or preference optimization , but must be embedded structurally through continual human participation . Instead of presenting a single algorithmic solution to AGI, this paper extends foundational research agenda, arguing that human centeredness must be structurally embedded at the level of architectures, objectives, and evaluation metrics. By clarifying the usefulness of human values within a multi-layered framework, we aim to provide a principled foundation for developing AGI systems that are not merely powerful, but coherently aligned with human reasoning, values, and long-term interests.

Citation

Cite the above work.

@misc{nandi_2026_18230989,
  author       = {Nandi, Soumyadip},
  title        = {Towards a Human-Centered AGI},
  month        = jan,
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.18230989},
  url          = {https://doi.org/10.5281/zenodo.18230989},
}