Fundamental rethink of neural network interpretability that could reshape mechanistic interpretability research approaches.
New research challenges the superposition theory of polysemanticity, suggesting neural networks are more interpretable than previously believed through recursive generalization patterns.
Researchers argue that LLM user personas may deserve moral consideration alongside assistant personas, based on philosophical implications of simulator frameworks.
Preface introduces computational agent theory framework for analyzing ASI alignment risks and existential threats.
Introduces a theoretical framework for artificial superintelligence alignment while acknowledging ASI's existential risks as argued by Yudkowsky and others.
Researchers introduce Latent Compaction, a test-time training method that compresses KV caches 32× while maintaining 96.7% accuracy on retrieval tasks.
Researchers introduce a new generative representation learning approach for scalable latent tokenization that achieves compute-optimal improvements over direct diffusion methods.
