claim

active

claim:the-core-problem-of-steering-should-be-recast-from-finding-the-right-direction-to-finding-the-right-geometry

The core problem of steering should be recast from finding the right direction to finding the right geometry

The paper's programmatic conclusion about how the field should reconceptualize neural network steering

Source paper

extracted_from

(2026) · Daniel Wurgaft · Can Rager · Matthew Kowal · Vasudev Shyam +12

paper

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.
Geometric goal-alignment in multi-scale systems
members_of
Framework treating teleonomic behavior and goal-directedness as geometric alignment problems across anatomical, physiological, and representational spaces; emphasizes intervention via internal geometry rather than external direction.

claim

question

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.

Linear steering is often mismatched with a model's internal representation geometry, producing noisy, off-target effects.claim0.765
The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
Optimally steering model behavior requires isolating concept geometry and defining operators to navigate it.claim0.754
Euclidean geometry assumption in steeringconcept0.753
Linear steering implicitly assumes a flat, Euclidean activation space, leading to off-manifold excursions.
steering (intervention on internals)concept0.752
General technique of modifying activations to control model behavior.
direction-based steeringconcept0.747
Paradigm of finding the right direction in activation space (e.g., linear steering).
Steering along manifolds provides better control than linear steering when the concept geometry is non-linear.claim0.747
The central thesis of the paper, motivating the shift from linear to geometry-aware manifold steering.
geometry-based steeringconcept0.745
Paradigm of finding the right geometry (manifold) for principled control.
Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.claim0.741
Supported by the instruction discovery experiments comparing steering vs. embedding baselines.