reasoning-budget

Benchmark results for MindControl, a sampler-level guided reasoning budget for llama.cpp, show multi-stage signaling reduces token consumption without accuracy loss, and sometimes improves it on easier tasks.