{"id":"arxiv:2605.05115","title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","date":"2026-05-06","section":"Representation alignment","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/","source_url":"https://arxiv.org/abs/2605.05115","text_available":true,"authors":["Daniel Wurgaft","Can Rager","Matthew Kowal","Vasudev Shyam"],"outgoing":[{"source":"arxiv:2605.05115","target":"arxiv:2309.08600","evidence":"title","id":"arxiv:2309.08600","title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/"},{"source":"arxiv:2605.05115","target":"arxiv:2502.00873","evidence":"id","id":"arxiv:2502.00873","title":"Language Models Use Trigonometry to Do Addition","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7766b2640c6b5ab53a295e8d60920cb2419bfbbd812810d9b35d4b7c323922ed/"}],"incoming":[],"text_parts":[{"start":0,"end":15730,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html"},{"start":15730,"end":31707,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html"},{"start":31707,"end":46977,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html"},{"start":46977,"end":62763,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html"},{"start":62763,"end":77466,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html"},{"start":77466,"end":82318,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-6.html"}],"contents":[{"title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"Abstract","offset":95,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"1 Introduction","offset":1922,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"2 The Geometry of Representation and Behavior","offset":7058,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"2.1 Setup","offset":7108,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"Running example.","offset":7123,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"Concept geometry.","offset":7637,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"Notation.","offset":9064,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"Experimental tasks.","offset":10324,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"2.2 Fitting the Manifolds","offset":10766,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"2.3 Conceptual Structure Appears in Behavior and Activation Space","offset":12253,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"3 Connecting Representation and Behavior via Intervention","offset":15668,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-1.html#text"},{"title":"3.1 Steering Intervention Notation","offset":16250,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"3.2 Steering Along the Activation Manifold Follows the Behavior Manifold","offset":18484,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"An Energy-based View of Naturalness.","offset":19838,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"3.3 Behavior Space Geometry Recovers the Activation Manifold","offset":23140,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"3.4 Unifying Steering Strategies Through Geometry","offset":26129,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"The Geometry of Steering:","offset":26863,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"Definition 1 (Geometries of Steering) .","offset":27708,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"4 Manifold Steering Yields Factored Control in Multi-Dimensional Spaces","offset":31631,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-2.html#text"},{"title":"In-context learning tasks with synthetic conceptual spaces.","offset":32315,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Manifold fitting.","offset":33349,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Isometry results.","offset":34329,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Manifold vs. Linear steering.","offset":35331,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"5 Manifold Steering on a Visual World Model: Mountain Car Task","offset":36745,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Environment and model architecture.","offset":37437,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Activation and behavior manifold fitting.","offset":38808,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Geometry-aware steering in activation space.","offset":40611,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"The geometry assumed by linear steering is not faithful to the conceptual ordering.","offset":41847,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"6 Related Work","offset":43051,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Activation Steering and the Linear Representation Hypothesis.","offset":43070,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Activation Geometry and its Origins.","offset":45518,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"Causal Analysis of Neural Networks.","offset":46935,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-3.html#text"},{"title":"7 Discussion","offset":48433,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Geometry-aware steering reveals the shared structure of behavior and representation.","offset":48450,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Where does the shared geometry of behavior and representation come from?","offset":49943,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Intrinsic coordinates of representation manifolds as units of causal analysis.","offset":51167,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"8 Future Work and Limitations","offset":52088,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Acknowledgments","offset":56379,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Appendix A Experimental Details for Language Tasks","offset":56566,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"A.1 Tasks and Datasets","offset":56818,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"Natural domain tasks.","offset":56846,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"In-context learning of representations.","offset":57479,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"A.2 Model, Intervention Site, and Output Distribution","offset":58067,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"A.3 Fitting the Activation Manifold ℳ h \\mathcal{M}_{h}","offset":59011,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"A.4 Fitting the Behavior Manifold ℳ y \\mathcal{M}_{y}","offset":61417,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-4.html#text"},{"title":"A.5 Geodesic Distances and the Isometry Test","offset":63818,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"A.6 Steering Interventions","offset":66110,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"A.7 Naturalness Metric","offset":67851,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"A.8 Pullback Optimization","offset":68798,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Path Parameterization.","offset":69367,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Loss and optimizer.","offset":70113,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"A.9 Pullback Recovery R 2 R^{2}","offset":71495,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Appendix B Experimental Details for the Vision Task","offset":72637,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"B.1 Mountain Car.","offset":72764,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Data collection.","offset":72787,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Manifold fitting.","offset":73721,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Appendix C Additional Results","offset":76096,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"C.1 In-Context Learning of Representations.","offset":76130,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"C.2 Manifold steering allows manipulation of uncertainty without loss of structure.","offset":76682,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"C.3 Mountain Car","offset":77358,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Structural Correspondence between ℳ h \\mathcal{M}_{h} and ℳ y \\mathcal{M}_{y} .","offset":77380,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-5.html#text"},{"title":"Pullback: Behavior space steering.","offset":78835,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-6.html#text"},{"title":"Instructions for reporting errors","offset":81354,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/part-6.html#text"}],"text_code_points":82318,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/source.md"}