https://research.yehoyada.com/https://research.yehoyada.com/about/https://research.yehoyada.com/research/learning-from-reasoning-mistakes-with-sparse-updates/https://research.yehoyada.com/research/compression-discovers-generalizing-programs/https://research.yehoyada.com/research/cyclic-consensus-optimization/https://research.yehoyada.com/research/demand-paged-experts-preserve-quality/https://research.yehoyada.com/research/geometric-early-stopping/https://research.yehoyada.com/research/headwise-depth-routing-quality-cost/https://research.yehoyada.com/research/larger-expert-banks-routing-bottleneck/https://research.yehoyada.com/research/neural-fabric-structured-success-language-failure/https://research.yehoyada.com/research/qwen-small-updates-efficient-adaptation/https://research.yehoyada.com/research/small-moe-proxies-fail-rank-validation/https://research.yehoyada.com/research/trajectory-denoising-and-transfer/https://research.yehoyada.com/research/training-neural-networks-to-be-corrected/https://research.yehoyada.com/research/complementary-memories-for-targeted-replay/https://research.yehoyada.com/research/bounded-search-for-targeted-replay/https://research.yehoyada.com/research/finding-damage-is-not-measuring-forgetting/https://research.yehoyada.com/research/when-fixed-updates-beat-a-learned-controller/https://research.yehoyada.com/research/predicting-self-change-without-better-decisions/https://research.yehoyada.com/research/capacity-and-certainty-budgets-in-continual-learning/https://research.yehoyada.com/research/oracle-gradient-inversion-meets-the-transformer/https://research.yehoyada.com/research/shared-learning-directions-mostly-learn-token-frequencies/https://research.yehoyada.com/research/low-shared-learning-rates-and-soft-targeted-replay/