finding

active

finding:manifold-steering-produces-clean-probability-shifts-along-natural-behavior-structure-linear-steering-cuts-across-manifold-and-produces-off-target-noisy-effects

manifold steering produces clean probability shifts along natural behavior structure; linear steering cuts across manifold and produces off-target noisy effects

Empirical demonstration on Llama-3.1-8B that steering along representation manifold aligns outputs with behavior manifold, whereas linear steering does not.

Source paper

extracted_from

Steering Along Manifolds to Control Neural Networks

Neighborhood — ranked by edge-count

Papers (1)

paper

Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior
supports

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.

Methods (1)

method

linear steering
contradicts
Typical approach that adds a scaled steering vector to representations; the paper argues this is mismatched with actual representation geometry.

Findings (1)

finding

Linear steering produces noisy off-target effects; manifold steering cleanly shifts probability mass between sequential concepts.
restates
Core empirical claim comparing steering approaches on cyclic concepts.

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.

Manifold-respecting steering produces smooth natural behavioral trajectories while linear steering teleports between non-adjacent concepts.claim0.864
Linear steering cuts through off-manifold regions and hence produces unnatural outputs.claim0.864
Attribution of failure to Euclidean assumption.
Linear steering is often mismatched with a model's internal representation geometry, producing noisy, off-target effects.claim0.849
The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
Steering along manifolds provides better control than linear steering when the concept geometry is non-linear.claim0.839
The central thesis of the paper, motivating the shift from linear to geometry-aware manifold steering.
Steering Llama-3.1 8B along the circular representation manifold produces outputs that follow the natural circle of the behavior manifold, cleanly shifting probability mass from Monday through successive days.finding0.835
Core empirical result demonstrating that manifold steering produces on-target, behavior-aligned outputs.
Manifold steering demonstrates bidirectional geometry-behavior link in a video world model on tasks with geometry corresponding to physical dynamicsfinding0.815
Extension of manifold steering validation to video world models and physical dynamics tasks, demonstrating cross-modal generality
Linear steering on Llama-3.1 8B for the days-of-week task cuts across the behavior manifold, producing noisy off-target effects where predicted tokens are not even days of the week.finding0.811
Empirical result demonstrating the failure mode of linear steering when concept geometry is cyclic.
Steering along M_h yields behavioral trajectories that follow M_y, producing more natural outputs than linear steeringfinding0.809
Core empirical result demonstrating the superiority of manifold steering over linear steering

Restated by (1)

cosine ≥ 0.90

Other entities that say roughly the same thing. May be merge candidates or independent restatements across papers.

finding
Linear steering produces noisy off-target effects; manifold steering cleanly shifts probability mass between sequential concepts.