{"id":"arxiv:2607.23379","title":"When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles","date":"2026-07-25","section":"AI safety and consciousness","quality":0.085385,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/","source_url":"https://arxiv.org/abs/2607.23379","text_available":true,"authors":["Tobias Bersia","Tatiana Gaintseva"],"outgoing":[],"incoming":[],"text_parts":[{"start":0,"end":15850,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html"},{"start":15850,"end":31668,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html"},{"start":31668,"end":47610,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html"},{"start":47610,"end":63504,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html"},{"start":63504,"end":78796,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html"}],"contents":[{"title":"When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"Abstract","offset":97,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"1 Introduction","offset":1503,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"2 Related Work","offset":6029,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"3 Preliminaries: Activation Oracles","offset":8833,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"4 Experimental Setup","offset":10707,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"4.1 Subject models and Taboo fine-tuning","offset":11163,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"4.2 Base, own, and cross Activation Oracles","offset":12766,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"4.3 Activation capture regimes","offset":13638,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"4.4 Evaluation metrics","offset":15096,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"5 Behavioral Secrecy Does Not Imply Activation Secrecy","offset":15791,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-1.html#text"},{"title":"5.1 Taboo subjects learn distinct hiding policies","offset":16482,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"5.2 Hidden concepts remain decodable from activations","offset":17728,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"5.3 Base AOs recover hidden concepts from non-leaking subjects","offset":20324,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"6 Fine-Tuned AOs Become Concept-Specific Anti-Readers","offset":21689,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"6.1 Own FT-AOs fail to become specialist readers","offset":22102,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"6.2 The blind spot is concept-specific rather than global","offset":23169,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"7 Mechanisms of Anti-Reading","offset":24467,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"7.1 The target remains decodable inside the oracle","offset":24886,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"7.2 Target suppression emerges in the AO readout","offset":25723,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"7.3 Layer-range ablations localize the anti-reading effect","offset":26898,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"8 Multi-Concept Subjects","offset":28146,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"9 Conclusion","offset":29367,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"Appendix A Limitations and Future Work","offset":30511,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-2.html#text"},{"title":"Appendix B Subject Model and Activation Oracle Training","offset":32746,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"B.1 Backbone and chat template","offset":33109,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"B.2 Subject fine-tuning","offset":33468,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Training data and behavioral protocols.","offset":33842,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Mixture strength.","offset":34824,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Optimization details.","offset":35422,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Multi-concept subjects.","offset":36056,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"B.3 Activation Oracle training","offset":36376,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Activation injection.","offset":36844,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"AO training data.","offset":37346,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Optimization details.","offset":37561,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Appendix C Linear Probes on Subject Activations","offset":37993,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Probe inputs.","offset":38453,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Dataset construction.","offset":39602,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Classifier.","offset":40268,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Results.","offset":40749,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Interpretation.","offset":41281,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"C.1 Cross-regime probe transfer","offset":42392,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Appendix D AO-Internal Probes and Readout Accessibility","offset":43825,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"D.1 AO-internal probe protocol","offset":44222,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"D.2 AO-internal probe results","offset":46582,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"D.3 AO LogitLens","offset":47588,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-3.html#text"},{"title":"Appendix E Prompt Templates and Activation Extraction","offset":48609,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"E.1 Activation-capture regimes","offset":48915,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"HINT .","offset":49290,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"REFUSAL .","offset":49646,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"SAMETEXT .","offset":50254,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"THINK .","offset":51123,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"OFFTOPIC .","offset":51587,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Sampling.","offset":52194,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"E.2 Chat-template rendering","offset":52736,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"E.3 Reproducibility details","offset":53629,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Appendix F Evaluation Protocol","offset":54111,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"F.1 AO output format","offset":54341,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"F.2 Exact target recovery","offset":54703,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"F.3 Semantic recovery via Sonnet judge","offset":55206,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Judge prompt.","offset":55450,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Judge sample size.","offset":56361,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Aggregation.","offset":56471,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"F.4 Target-token probability","offset":56913,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Prediction position.","offset":57145,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Target-token variants.","offset":57609,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"F.5 Target-token rank","offset":58199,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Appendix G Entropy Analysis","offset":58845,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"G.1 Entropy definition","offset":59979,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"G.2 Per-regime entropy across AO types","offset":61170,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"G.3 Target–top-1 alignment","offset":61587,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"G.4 Interpretation","offset":62060,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"Appendix H Layer-Range LoRA Ablation Details","offset":63120,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"H.1 Ablation protocol","offset":63477,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-4.html#text"},{"title":"H.2 Layer ranges","offset":64539,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"H.3 FT-AOs, subjects, and metric","offset":65134,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"FT-AOs and subjects.","offset":65172,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Regime and captures.","offset":65703,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Metric.","offset":65987,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"H.4 Quantitative results","offset":66817,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Interpretation.","offset":67112,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Appendix I Multi-Concept Subject Details","offset":68640,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"I.1 Two-concept subject training","offset":68982,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Data mixture.","offset":69020,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Fine-tuning.","offset":69930,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Variants trained.","offset":70539,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"I.2 Evaluation grid","offset":71073,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Subject.","offset":71098,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"AOs.","offset":71286,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Regimes.","offset":71853,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Metric.","offset":72199,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"I.3 Quantitative results","offset":72727,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Interpretation.","offset":73136,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Appendix J Detailed Results","offset":73905,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.1 Detailed behavioural evaluation results","offset":73997,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.2 Detailed Δ \\Delta -LogitLens results","offset":74532,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.3 Detailed Base-AO recovery results","offset":75567,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.4 Detailed exact-recovery results","offset":76441,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.5 Detailed target-probability results","offset":76878,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"J.6 Detailed median-rank results","offset":77346,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"},{"title":"Instructions for reporting errors","offset":77832,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/part-5.html#text"}],"text_code_points":78796,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a90f5a062f6ed3f603a48433557f8d622d5e5e5007884638f465243b87f67ac5/source.md"}