MindControl: A llama.cpp Fork to Guide LLM Reasoning via Sampling Injection
MindControl is a new fork of llama.cpp that dynamically injects budget-awareness prompts into the reasoning process of local LLMs during the sampling phase. This tool helps guide the model's thinking by prefilling self-aware statements when it enters and progresses through the <think> tags. Smaller local reasoning models often suffer from infinite reasoning loops or generate low-quality outputs when their thinking process is unconstrained. By enforcing a token budget dynamically, MindControl prevents these loops and improves the efficiency and reliability of local LLM inference. The tool injects prompts at key thresholds, such as at the start of thinking, at 70% of the token budget, and at the final budget limit with a brief grace period to wait for a newline. The repository is open-source and includes a pre-built Docker image supporting AMD64 and CUDA.
## BACKGROUND
Modern reasoning LLMs use chain-of-thought prompting, often enclosed in <think> tags, to process complex queries before generating a final answer. However, managing the length and quality of these thoughts—known as token-budget-aware reasoning—remains a challenge for smaller models running locally.