What's this about?
Small multimodal agents choose among reusable natural-language reasoning patterns, scoring them in parallel instead of writing a new chain before every action.
What exactly is a shared KV cache?
It reuses internal representations of the common context while evaluating several candidate continuations, avoiding repeated computation.
Why does this matter?
The authors report large latency cuts without lower success on seven search benchmarks. Predefined plans improve efficiency but may restrict behaviour when a genuinely novel strategy is required.