benchmarkin-context memorypublished Oct 1, 2026 · arXiv:2609.38866When Context Changes: Understanding Update Failures in LLMs
unembedding geometryBayesian priorspublished Sep 4, 2026 · arXiv:2609.02959The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors
model compressiontransfer learningpublished Aug 6, 2026 · arXiv:2608.02829Wiring Beats Blending: What Transfers Between Transformer Sizes -- and What Doesn't