Five papers worth your time

Five papers from last week that are worth your time, each with a source.

1. 1Embeddings Are Overpriced. In Gemma 3 270M the embedding table is 63% of the parameters. A frozen 16 MB table held up on three of four benchmarks.
2. 2FC-SWE. A 4B model went from 38.9% to 41.7% on SWE-bench Verified by learning from its failed patches.
3. 3DLite. Agent workloads cut speculative decoding acceptance by up to 80%. History-free drafting is the fix they propose.
4. 4PremiseCheck. Checking claims with a cheap model after a rewrite reached frontier-level accuracy at about half the cost.
5. 5OpenMobile-2. A mobile agent that mixes screens and app tools nearly doubled its real-device score.

The thread running through them: the interesting gains come from where the work happens, not from a bigger model. Vocabulary, retries, drafting and checking.

If you have a paper or repo that belongs on this list, send it to me.

Which of the five will change something you build?