Curved manifolds often represent concepts better than linear directions.

Proposes that nonlinear geometric structure is superior to linear feature spaces for capturing semantic content.

Source paper

extracted_from

The World Inside Neural Networks

(2026) · Geiger, Atticus · Lubana, Ekdeep Singh · Fel, Thomas · Merullo, Jack +3

Neighborhood — ranked by edge-count

Findings (1)

finding

In the Mountain Car case study, car position is a 1D manifold; linear interventions cross voids causing incoherence; following the 1D curve produces smooth control.
supports
Empirical demonstration that a semantically meaningful variable is encoded as a curved manifold, and that respecting its geometry is critical for effective intervention.

Communities (3)

community

Manifold-aware concept steering in neural representations
members_of
Explores geometry of activation/behavior manifolds to enable selective, non-destructive concept interventions.
Geometric concept representations in neural networks
members_of
Concepts encoded as curved manifolds and circular structures in LLM activation spaces.
Manifold-aware steering for language models
members_of
Geometric approach to model control that respects learned concept structure, contrasting with linear steering that produces off-manifold artifacts.

Concepts (3)

concept

curved manifold
cites
A smoothly varying lower-dimensional surface in activation space that captures a concept better than a straight linear direction.
concept representation
cites
How a neural network encodes a semantic concept internally, argued to be better captured by manifolds than by atomic features.
linear direction
cites
A straight vector in activation space, traditionally used for concept manipulation; claimed to be insufficient when true concept geometry is curved.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Steering along manifolds provides better control than linear steering when the concept geometry is non-linear.claim0.838
The central thesis of the paper, motivating the shift from linear to geometry-aware manifold steering.
Manifold geometry provides a practical steering blueprint in an image-action model predicting car position on a hill, extending results across modalities.finding0.810
Cross-modality result from the full paper demonstrating that representation-behavior geometry alignment is not limited to language models.
The deep nature of space appears to be so profound that just the geometric unfolding, when done properly, leads to results which make sense from an engineering point of view.claim0.806
The mystery that beautiful geometry often yields good structural behavior is acknowledged but not yet fully explained mathematically.
Manifold geometry provides a practical blueprint for steering model behavior across diverse tasks and modalities.hypothesis0.806
The generalizing predictive claim that manifold steering is a broadly applicable framework beyond the days-of-week case study.
There is a clear bidirectional relationship between the geometry of behavior and representation: steering along representation manifolds follows behavior manifolds, and vice versa.claim0.804
The paper's finding that the alignment holds in both directions — from representation to behavior and from behavior back to representation space.
The geometry of the physical world has the most profound impact on human inner freedom and sense of life.claim0.793
Central thesis of the chapter.
manifold geometry principles extend to months, letters, ages, and in-context learning tasks across modalitiesfinding0.790
Evidence that the weekday cyclic structure is not anomalous but reflects broader principle of concept geometry.
Linear steering produces noisy off-target effects; manifold steering cleanly shifts probability mass between sequential concepts.finding0.787
Core empirical claim comparing steering approaches on cyclic concepts.