AI Daily Digest · 2026-11-12
🔬 New AI Craft
1. Learning query-aware routing for vector retrieval
Apple researchers train small networks offline for a fixed vector database to route maximum-inner-product queries toward promising regions using the expected query distribution. The approach trades upfront training and a fixed-key setting for retrieval efficiency; gains depend on the workload and index.
https://machinelearning.apple.com/research/faster-retrieval
2. LiteLLM migrates routing core to Rust with differential tests
LiteLLM describes refactoring its Python routing interface and side effects before moving the core to Rust, using differential tests to compare behavior at each step while keeping the Python API. The team presents this as a compatibility-focused migration and does not claim an immediate major speedup for existing users.
https://docs.litellm.ai/blog/router-rust
3. Using linear programming to reproduce a loop-suppression weight edit
Nathan Barry reformulates a selection problem in a prior Qwen3.5-0.8B weight edit as a linear program and compares a smaller edit with the original approach. The experiment targets reasoning loops in that model under specified evaluations, so it does not establish a general capability improvement across models.
https://nathan.rs/posts/weight-space-lp
🛠️ Tools & Tips
(no items this day)
⭐ Open Source Highlights
(no items this day)
📰 Industry News
(no items this day)
🚀 Major Releases
(no items this day)
