Representations
The structures through which models encode concepts: features, directions, geometry, and the compression schemes that pack a world model into a residual stream.
A transformer's residual stream is a high-dimensional vector space, and the working hypothesis of this programme (the linear representation hypothesis) is that models encode meaningful concepts as directions in that space. The hypothesis is productive precisely because it is testable: probes can find directions, interventions can move activations along them, and both can fail in informative ways.
Models represent more concepts than they have dimensions. That forces superposition: features share capacity, interfere with one another, and defeat naive neuron-level reading. Sparse autoencoders and their successors (transcoders, crosscoders) attempt to unmix this compression into interpretable units, and much of the field's current empirical activity is testing where and how well that works.
This programme also covers what the representations are of: internal world models, self-relevant features, persona directions, and the stability of all of these across context, fine-tuning, and scale. Everything downstream, circuits, cognition, behaviour, reads from this substrate.
Residual-stream geometry
How activation space is organised: anisotropy, feature directions, and what distance means inside a layer.
Superposition
More features than dimensions: interference patterns, capacity allocation, and when compression defeats interpretation.
Sparse autoencoders
Dictionary learning on activations; feature quality, reconstruction error, and what SAE features do and do not license.
Crosscoders & transcoders
Dictionaries across layers and models: shared features, layer-to-layer transport, and replacement-model tracing.
Representation universality
Which features recur across models, scales, and training runs, and what recurrence implies.
Persona vectors & concept directions
Directions that track personas, traits, and abstract concepts; what steering them changes.
Feature steering
Adding and ablating directions as a causal test of what a representation does.
Representational stability
Whether the same concept keeps the same direction across context, prompts, and fine-tuning.
- Which concepts have linear representations, and which are encoded in ways probes systematically miss?
- How much of an SAE's dictionary reflects the model, and how much reflects the SAE's own inductive bias?
- Do self-relevant features, situation, role, persona, form a coherent subspace, and is it stable under fine-tuning?
- What is the right null model for claiming a direction 'represents' anything at all?
Methods in use or proposed: Linear and nonlinear probes · sparse autoencoders and crosscoders · activation steering · representational similarity across models · synthetic-data controls.
A short orientation list of primary sources, not a survey. The Interpretability Map holds the full dependency graph.
- Olah, C., et al. (2020). Zoom In: An Introduction to Circuits. Distill. distill.pub/2020/circuits/zoom-in/
- Elhage, N., et al. (2022). Toy Models of Superposition. Transformer Circuits Thread. transformer-circuits.pub/2022/toy_model/index.html
- Bricken, T., et al. (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Transformer Circuits Thread. transformer-circuits.pub/2023/monosemantic-features/index.html
- Templeton, A., et al. (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread. transformer-circuits.pub/2024/scaling-monosemanticity/index.html
- Park, K., Choe, Y. J., & Veitch, V. (2023). The Linear Representation Hypothesis and the Geometry of Large Language Models. arxiv.org/abs/2311.03658
- Gurnee, W., & Tegmark, M. (2023). Language Models Represent Space and Time. arxiv.org/abs/2310.02207
- Chen, R., et al. (2025). Persona Vectors: Monitoring and Controlling Character Traits in Language Models. arxiv.org/abs/2507.21509