arXiv:2605.27970v1 Announce Type: new Abstract: While large language models (LLMs) are trained purely on textual data, prior work has shown that their internal representations can exhibit rich geometric structure in embedding space. Building on this line of work, we investigate whether such structure is similar to human perceptual organisation across different domains (e.g., color, pitch, emotion, and taste). Specifically, we study the layer-wise emergence of intrinsic geometrical structure corresponding to perceptual modalities within the residual streams of multiple open-weight transformer a
This research builds on recent advancements in understanding LLM internal representations, leveraging new analytical techniques to probe their emergent structures.
Understanding how LLMs develop human-like perceptual organization could unlock new avenues for more intuitive, robust, and generalizable AI, moving beyond purely text-based reasoning.
This research suggests LLMs are not just statistical text processors but may be spontaneously developing internal models of the world, including perceptual domains, akin to human cognition.
- · AI researchers
- · LLM developers
- · Cognitive science
- · Purely symbolic AI approaches
- · Those underestimating LLM capabilities
This finding deepens our understanding of emergent properties in large language models and their potential for general intelligence.
Improved understanding of LLM internal geometry could lead to new interpretability techniques and more effective model alignment strategies.
The revelation of human-like perceptual structures could contribute to a re-evaluation of consciousness and intelligence in artificial systems.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI