{"id":"arxiv:2309.08600","title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","date":"2023-09-15","section":"AI safety and consciousness","quality":0.096043,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/","source_url":"https://arxiv.org/abs/2309.08600","text_available":true,"authors":["Hoagy Cunningham","Aidan Ewart","Logan Riggs","Robert Huben"],"outgoing":[{"source":"arxiv:2309.08600","target":"arxiv:1803.03635","evidence":"id","id":"arxiv:1803.03635","title":"The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a9d7f0235205afcc709e3e74f28cc7707831bcb1d46239c418cc7e4b624d23bd/"}],"incoming":[{"source":"arxiv:2407.15017","target":"arxiv:2309.08600","evidence":"title","id":"arxiv:2407.15017","title":"Knowledge Mechanisms in Large Language Models: A Survey and Perspective","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8900dc2821aa61e59c5b1219cf9eb36dfaabf303c17d0064deaca653b5ec9a77/"},{"source":"arxiv:2502.00873","target":"arxiv:2309.08600","evidence":"title","id":"arxiv:2502.00873","title":"Language Models Use Trigonometry to Do Addition","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7766b2640c6b5ab53a295e8d60920cb2419bfbbd812810d9b35d4b7c323922ed/"},{"source":"arxiv:2604.11962","target":"arxiv:2309.08600","evidence":"title","id":"arxiv:2604.11962","title":"The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8428fd0cce9753855f4d0f801609b0205a0be52a459d53ae44e7860d00348485/"},{"source":"arxiv:2605.05115","target":"arxiv:2309.08600","evidence":"title","id":"arxiv:2605.05115","title":"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d9659f959b746810c945d693de7b0e4d1a73aeaa7ec183aaa7b8f68549e3d50/"}],"text_parts":[{"start":0,"end":15478,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html"},{"start":15478,"end":31456,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html"},{"start":31456,"end":40916,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html"}],"contents":[{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"Abstract","offset":77,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"1 Introduction","offset":1539,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"2 Taking Features out of Superposition with Sparse Dictionary Learning","offset":4851,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"3 Interpreting Dictionary Features","offset":9417,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"3.1 Interpretability at Scale","offset":9456,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"3.2 Sparse dictionary features are more interpretable than baselines","offset":10677,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"4 Identifying Causally-Important Dictionary Features for Indirect Object Identification","offset":13236,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"4.1 Adapting activation patching to dictionary features","offset":14323,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-1.html#text"},{"title":"4.2 Precise Localisation of IOI Dictionary Features","offset":16737,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"5 Case Studies","offset":17427,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"5.1 Input: Dictionary Features are Highly Monosemantic","offset":18040,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"5.2 Output: Dictionary Features have Intuitive Effects on the Logits","offset":19480,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"5.3 Intermediate Features: Dictionary Features Allow Automatic Circuit Detection","offset":20475,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"6 Discussion","offset":21974,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"6.1 Related Work","offset":21991,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"6.2 Limitations and Future Work","offset":23136,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"6.3 Conclusion","offset":25532,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"Acknowledgments","offset":26634,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"Appendix A Autointerpretation Protocol","offset":27483,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"Appendix B Sparse Autoencoder Training and Hyperparameter Selection","offset":29504,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"Appendix C Further Autointerpretation Results","offset":30753,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"C.1 Interpretability is Consistent across Dictionary Sizes","offset":30803,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"C.2 High Interpretability Scores Are Not an Artefact of Top Scoring","offset":31383,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-2.html#text"},{"title":"C.3 Interpreting the MLP Sublayer","offset":33559,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"C.4 Interpretability Scores Correlate with Kurtosis and Skew of Activation","offset":34957,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"Appendix D Qualitative Feature Analysis","offset":36031,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"D.1 Residual Stream Basis","offset":36075,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"D.2 Examples of Learned Features","offset":36455,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"D.3 Feature Search Details","offset":36554,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"D.4 Failed Interpretability Methods","offset":37417,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"Appendix E Number of Active Features","offset":37887,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"Appendix F Editing IOI Behaviour on other Layers","offset":38738,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"Appendix G Top K Comparisons","offset":38892,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"},{"title":"Instructions for reporting errors","offset":39952,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/part-3.html#text"}],"text_code_points":40916,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/311cc8e986e42947a9cd36e64cb7dad93d7352e1f5ad495664fce8d0f6fca5da/source.md"}