concept
active
concept:manifold-steeringManifold Steering
Central framework: steering neural networks by intervening along the curved manifold where a concept lives, rather than in straight lines through activation space.
Neighborhood — ranked by edge-count
Papers (3)
paper
- The World Inside Neural Networksmentions
Thinkers (1)
thinker
- WurgaftstudiesResearcher whose manifold steering concept is referenced as the same conceptual move underlying self-correcting search.
Frameworks (2)
framework
- Fifteen Properties of Living Structureassociated_withThe set of geometric properties that appear in all living structure: levels of scale, strong centers, boundaries, echoes, gradients, deep interlock and ambiguity, local symmetries, roughness, inner calm, not separateness, and others.
- Geometry-Aware Steering FrameworkimplementsThe overarching theoretical framework proposed in the paper, asserting that steering interventions should be aligned with the geometric structure of the model's representation manifold.
Communities (1)
community
- Neural Geometryassociated_withmembers_of
Claims (1)
claim
- Strong interpretive assertion linking discovery and control: neural computation is fundamentally manifold-structured.
Methods (6)
method
- linear steeringcontradictsextendsTypical approach that adds a scaled steering vector to representations; the paper argues this is mismatched with actual representation geometry.
- Self-Correcting Searchassociated_withTechnique using internal model representations as feedback loops to steer diffusion-based materials generation toward target properties.
- Method to fit a manifold M_h to neural representations in activation space.
- Method to fit a manifold M_y to output probability distributions.
- The procedure of fitting a one-dimensional manifold (path) to clusters in activation or behavior space to capture the geometric structure of a concept.
- The general experimental approach of intervening along geometrically-defined paths rather than single-point or linear-direction interventions
Concepts (18)
concept
- Representation SteeringextendsParent concept; the practice of controlling neural network outputs by manipulating internal representations.
- behavior manifoldimplementsOne-dimensional curved surface in output probability space; the paper shows this mirrors representation manifold structure.
- One-dimensional curved surface in internal activation space; the paper demonstrates alignment with behavior manifold.
- activation manifold M_himplementsManifold fitted to representations in activation space.
- Neural Representation GeometryimplementsThe broader conceptual framework that neural activations exhibit non-Euclidean geometric structure causally linked to behavior.
- Activation ManifoldimplementsThe low-dimensional geometric structure discovered in neural activation space; contrasted with linear/Euclidean geometry.
- Neural Geometriesassociated_with
- Language Modelsassociated_withPrimary substrate for manifold steering experiments; demonstrates method on reasoning and in-context tasks.
- Interpretability-driven steeringassociated_withGeneral approach of using interpretability feedback to steer model generation.
- Video World Modelassociated_withSecondary substrate demonstrating cross-modal applicability of manifold steering.
- Activation-Behavior Manifold Isometryassociated_with
- geometry-based steeringimplementsParadigm of finding the right geometry (manifold) for principled control.
- The goal of mechanistically-grounded, reliable control of neural network behavior via activation interventions
- Gap: linear steering assumes Euclidean geometry and does not account for the actual curved geometry of activation manifoldsassociated_withThe research gap that motivates manifold steering as an alternative to conventional linear approaches
- Pullback Geometry (Behavior-Aware Metric)associated_with
- steering (intervention on internals)implementsGeneral technique of modifying activations to control model behavior.
- Consciousness-UX Vector 5: Care-Like Featuresassociated_withProgram application; manifold steering applicable if care states have cyclic/sequential structure.
Findings (2)
finding
- Core empirical claim comparing steering approaches on cyclic concepts.
- Days-of-Week Cyclic StructuresupportsKey empirical result: days-of-week appear as identical circular manifold in both Llama-3.1-8B internal activations and output token probability distributions.
Venues (1)
venue
- arXiv:2605.05115introducesPublication venue for Wurgaft et al. manifold steering paper.
Events (1)
event
- Coordinated release of three papers (Geiger manifesto, Feucht Geometric Calculator, Wurgaft Manifold Steering) establishing neural geometry framework.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Internal-state feedback technique for steering language models; same conceptual mechanism applied by Hazra et al. to chemistry.
- The central thesis of the paper, motivating the shift from linear to geometry-aware manifold steering.
- The paper demonstrates the bidirectional geometry-behavior relationship across multiple tasks and modalities (language models and video world models)
- Extension of manifold steering validation to video world models and physical dynamics tasks, demonstrating cross-modal generality
- Cross-modality result from the full paper demonstrating that representation-behavior geometry alignment is not limited to language models.
- A method for modifying model behavior by adding perturbation vectors to activations, used here to try to reduce eval awareness.