🤖 AI

Geometry of Human Perceptual Domains Emerges Transiently in LLM Representations

This paper demonstrates that human-like perceptual geometries across domains such as color and emotion transiently emerge within the intermediate layers of large language models, following a distinct developmental trajectory despite the absence of direct perceptual training.

Original authors: Simardeep Singh, Paras Chopra

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Simardeep Singh, Paras Chopra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a Large Language Model (LLM) as a massive, multi-story library where every book is a word, and every floor represents a different stage of "thinking." Usually, we think of these models as just being really good at predicting the next word in a sentence. They are trained only on text—no eyes to see colors, no ears to hear music, and no tongue to taste food.

However, this paper discovers something surprising: even though these models have never experienced the world with their senses, the "geometry" (the shape and arrangement) of how they organize concepts inside their brain looks a lot like how humans organize them.

Here is a breakdown of their findings using simple analogies:

1. The "Ghost" of Human Perception

Think of the model's internal brain as a giant, invisible map. The researchers asked: If we draw a map of how the model sees "colors," "emotions," "musical notes," or "tastes," does it look like a human map?

They found that yes, it does.

Colors: Humans see colors in a circle (a color wheel) where red blends into orange, which blends into yellow. The model, despite only reading text, also arranges colors in a perfect circle in its internal map.
Emotions: Humans often map emotions on a grid based on how "happy" vs. "sad" and how "calm" vs. "excited" we feel. The model creates this same grid.
Pitch: Humans hear musical notes as a smooth slide from low to high. The model arranges these notes in a smooth arc, just like we do.

2. The "Flashlight" Effect (Transient Emergence)

This is the most fascinating part. You might think that if the model has a "color map," it would keep that map perfectly clear from the bottom floor of the library to the top.

But the paper found that the model's "perception" is like a flashlight that only shines brightly in the middle of the room.

Early Layers (The Bottom Floor): The map is blurry and messy. The model is just looking at the raw letters and basic words. It hasn't organized the concepts yet.
Middle Layers (The Middle Floor): This is where the magic happens. The flashlight turns on full brightness. The messy dots suddenly snap into a perfect circle (for colors) or a smooth curve (for pitch). The model's internal structure suddenly looks very human-like here.
Late Layers (The Top Floor): As the model moves to the very top to prepare its final answer, the flashlight dims again. The perfect geometric shapes start to blur or break apart. The model stops caring about the "shape" of the concept and starts focusing on the specific task (like answering a question or finishing a sentence).

The Analogy: Imagine a sculptor working on a statue.

Early: They have a rough block of stone (messy data).
Middle: They carve the perfect, smooth shape of the face (the geometric structure emerges).
Late: They start adding tiny details for a specific purpose, and the perfect smooth curve of the face gets slightly altered or obscured by those final touches.

3. Different Speeds for Different Senses

Not all senses "wake up" at the same time.

Taste: The model figures out the shape of tastes (sweet, salty, sour) very early on, but it gets messy quickly. It's like a quick sketch that doesn't last.
Color: The model takes a bit longer to organize colors, but the shape is very clear and stable in the middle layers.
Emotion: This one is the most persistent. Once the model organizes emotions, it holds onto that structure even as it moves to the later layers. It's like a sturdy sculpture that doesn't fade away.

4. Why This Matters (According to the Paper)

The paper doesn't say the robot is "feeling" happy or "seeing" red. It doesn't claim the model is conscious.

Instead, it suggests that language itself contains the blueprint for human perception. Because humans talk about colors, emotions, and tastes in specific ways that relate to each other, the model learns to arrange these words in a shape that mimics how our brains arrange the actual sensations.

In summary: The paper shows that even without eyes or ears, an AI trained only on text builds a temporary, internal "map" of the world that looks surprisingly similar to a human's map. However, this map is fleeting—it appears clearly in the middle of the model's processing and then fades away as the model gets ready to speak.

Technical Summary: Geometry of Human Perceptual Domains Emerges Transiently in LLM Representations

Problem Statement

While Large Language Models (LLMs) are trained exclusively on textual data, prior research has established that their internal embedding spaces exhibit rich geometric structures corresponding to conceptual domains (e.g., cyclical manifolds for days of the week or 1D manifolds for time). However, it remains unclear whether these models, lacking direct sensory grounding, develop internal representations that align with human perceptual organization across sensory and affective modalities such as color, pitch, taste, and emotion. Furthermore, existing studies have largely focused on model outputs or task-specific probing, leaving a gap in understanding how and where such perceptual structures emerge and evolve across the intermediate layers of transformer architectures.

Methodology

The authors propose a fully intrinsic, layer-wise analysis framework to investigate the emergence of perceptual geometry without additional training or probing.

Stimuli and Extraction: A set of stimuli (e.g., specific colors, emotions, pitch frequencies, taste compounds) is processed through pretrained transformer models (including LLaMA-3, Qwen, and Gemma variants) using minimal, structured prompts to extract hidden state activations ( $h_i^{(l)}$ ) at every layer $l$ .
Geometric Construction:
- Dissimilarity Computation: Pairwise cosine dissimilarities are calculated between embeddings to form a dissimilarity matrix $D$ , isolating directional structure while remaining invariant to scale.
- Dimensionality Reduction: Multidimensional Scaling (MDS) is applied to $D$ to generate low-dimensional (2D) geometric maps that preserve pairwise dissimilarities, creating a "qualia map" for each layer.
Alignment Metrics: The model-generated geometric maps are quantitatively compared against human perceptual baselines (derived from established psychological datasets like ANEW and Marjieh et al., 2023) using two metrics:
- Representational Similarity Analysis (RSA): Spearman rank correlation between the vectorized upper-triangular entries of the model and human dissimilarity matrices.
- Generalized Procrustes Analysis (GPA): Measures geometric alignment by finding the optimal orthogonal transformation between the model and human maps, yielding a score where higher values indicate stronger alignment.
Profiling: This pipeline is applied across all layers to trace the evolution of perceptual structure, identifying where alignment peaks and how it changes with depth.

Key Contributions

The paper presents three primary contributions:

Layer-wise Emergence of Perceptual Geometry: The authors demonstrate that LLMs spontaneously develop structured geometric representations for multiple perceptual domains (color, pitch, taste, emotion) within their residual streams. These structures quantitatively align with human perceptual baselines despite the absence of direct sensory supervision.
Modality-Specific Emergence Profiles: The study reveals that different perceptual domains follow distinct trajectories. While all exhibit a rise-and-fall pattern, the timing of peak alignment and the stability of the structure vary by domain (e.g., taste peaks earlier than emotion) and by model architecture.
Transient Perceptual Encoding: The paper establishes a consistent representational trajectory across models: perceptual geometry is weak or diffuse in early layers, becomes progressively organized in intermediate layers, and attenuates in later layers. This suggests that perceptual geometry is encoded transiently as part of the model's internal transformation pipeline rather than being uniformly preserved.

Results

The analysis across four modalities yields the following specific findings:

Color: Models recover a circular manifold resembling the human color wheel, peaking in early-to-intermediate layers. This structure is transient, weakening in deeper layers as representations specialize for task-specific features. This "rise–peak–fall" profile is consistent across LLaMA-3-8B, LLaMA-3.2-3B, and Gemma-7B, though Qwen-3-4B shows a brief late-layer rebound before final degradation.
Emotion: The models recover a well-organized affective manifold aligned with the human valence–arousal structure. Unlike color, emotional geometry remains comparatively stable across later layers, indicating more persistent retention of this structure.
Pitch: A smooth, arc-like organization consistent with the continuous and ordinal nature of pitch perception emerges in intermediate layers. The structure deforms in deeper layers, suggesting pitch is represented relationally rather than categorically.
Taste: Models recover a taste manifold matching human perceptual arrangements (e.g., ordering of primary tastes). However, taste exhibits lower RSA scores and noisier layer-wise profiles compared to other domains, with geometry emerging early but degrading rapidly.

Significance and Claims

The paper claims to provide new insight into the mechanistic nature of LLM representations, specifically how human-like perceptual geometry arises from purely linguistic training. By demonstrating that these structures emerge transiently in intermediate layers, the authors offer a principled pathway for mechanistic analysis of internal representations.

The authors explicitly note the limitations of their work:

The analysis is descriptive; it identifies where and how geometry emerges but does not explain the underlying mechanisms causing these manifolds.
The study is limited to four perceptual domains and specific architectures.
Geometric alignment with human baselines is interpreted as evidence of structurally similar representational organization, not as proof of human-like subjective perception.

The work suggests that language models implicitly encode perceptual structures through the statistical regularities of language, organizing these structures in a transient, depth-dependent manner that mirrors, yet differs from, human perceptual stability.