{"total":20,"rows":[{"id":"2510.04340","title":"Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time","created":"2025-10-05","cited_by_count":43,"area":null},{"id":"2605.02087","title":"Model Spec Midtraining: Improving How Alignment Training Generalizes","created":"2026-05-03","cited_by_count":20,"area":null},{"id":"2601.21571","title":"Shaping capabilities with token-level data filtering","created":"2026-01-29","cited_by_count":13,"area":null},{"id":"2501.12391","title":"Physics of Skill Learning","created":"2025-01-21","cited_by_count":8,"area":null},{"id":"2507.09937","title":"Memorization Sinks: Isolating Memorization during LLM Training","created":"2025-07-14","cited_by_count":7,"area":null},{"id":"2505.15811","title":"On the creation of narrow AI: hierarchy and nonlocality of neural network skills","created":"2025-05-21","cited_by_count":4,"area":null},{"id":"2505.09341","title":"Access Controls Will Solve the Dual-Use Dilemma","created":"2025-05-14","cited_by_count":3,"area":null},{"id":"2606.06533","title":"Position: Don't Just \"Fix it in Post\": A Science of AI Must Study Training Dynamics","created":"2026-06-03","cited_by_count":3,"area":null},{"id":"2607.08077","title":"Modular Pretraining Enables Access Control","created":"2026-07-09","cited_by_count":3,"area":null},{"id":"2502.02470","title":"Studying Cross-cluster Modularity in Neural Networks","created":"2025-02-04","cited_by_count":2,"area":null},{"id":"2602.05164","title":"Position: Capability Control Should be a Separate Goal From Alignment","created":"2026-02-05","cited_by_count":2,"area":null},{"id":"2606.12360","title":"Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal","created":"2026-06-10","cited_by_count":2,"area":null},{"id":"2607.00510","title":"Prototype Language Models","created":"2026-07-01","cited_by_count":2,"area":null},{"id":"2607.20379","title":"Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations","created":"2026-07-22","cited_by_count":2,"area":null},{"id":"2510.06527","title":"Wide Neural Networks as a Baseline for the Computational No-Coincidence Conjecture","created":"2025-10-08","cited_by_count":1,"area":null},{"id":"2505.00509","title":"Self-Ablating Transformers: More Interpretability, Less Sparsity","created":"2025-05-01","cited_by_count":0,"area":null},{"id":"2509.01874","title":"Preserving Bilinear Weight Spectra with a Signed and Shrunk Quadratic Activation Function","created":"2025-09-02","cited_by_count":0,"area":null},{"id":"2512.12663","title":"PerNodeDrop: A Method Balancing Specialized Subnets and Regularization in Deep Neural Networks","created":"2025-12-14","cited_by_count":0,"area":null},{"id":"2604.00324","title":"The Persistent Vulnerability of Aligned AI Systems","created":"2026-03-31","cited_by_count":0,"area":null},{"id":"2608.30978","title":"Sparse Competition during Training For the Emergence of Specialized Modules","created":"2026-08-31","cited_by_count":0,"area":null}],"next_cursor":null}