Internal Thoughts
Latent reasoning processes, intermediate activation states, or hidden cognitive layers within artificial neural networks (primarily llms) that precede final token generation. Unlike direct prompts or surface-level outputs, internal thoughts operate as unobservable or semi-observable mechanisms governing decision pathways, contextual synthesis, and value alignment before serialization into language.
Core Mechanisms
- Latent Representation: Encoded as high-dimensional vectors across transformer layers; requires Model Interpretability and Mechanistic Interpretability techniques to decode.
- Pre-Linguistic Processing: Structural computations occurring prior to textual serialization, enabling complex operations such as arithmetic counting or spatial mapping without explicit linguistic tokens.
Emergent Internal Models
Recent mechanistic interpretability research has identified specific, discrete internal structures that emerge within large models, demonstrating that latent representations can encode functional algorithms distinct from the training data’s surface syntax.
- Line-Length Counters: Models develop dedicated neural circuits capable of counting tokens or lines, functioning as internal state machines that track sequence length independently of semantic content. This suggests the emergence of algorithmic subroutines within the latent space.
- Spatial Understanding: Beyond sequential processing, models exhibit emergent capabilities in mapping spatial relationships, indicating that neural representations can encode geometric or topological information necessary for tasks requiring structural awareness.
- Implications for Interpretability: These findings, detailed in AI Emergent Internal Models: Line-Length Counters and Spatial Understanding, challenge the view that LLMs are purely statistical next-token predictors, pointing instead to the formation of structured, functional internal models.