nowait-reasoning-optimizer

Reduce reasoning latency in DeepSeek-R1 or QwQ by suppressing non-essential reflection tokens like wait and rethink. This shortens generation time and lowers total token usage without compromising correctness in inference-heavy workloads.

Reflection tokens are deliberation markers like wait, rethink, and scratchpad used in R1-style reasoning models. They increase inference latency and token usage by extending the reasoning chain during generation.

Yes, you can optimize token usage in Phi4-Reasoning by suppressing non-essential reflection tokens. This minimizes deliberation steps to achieve faster inferences while maintaining correctness on long reasoning chains.

Suppress wait and rethink tokens in long reasoning chains by applying NOWAIT optimization configurations to your prompts. This requires frontmatter with a name and description to function correctly.

NOWAIT optimization works specifically with R1-style reasoning models such as QwQ, DeepSeek-R1, and Phi4-Reasoning. It targets prompts with long reasoning chains and iterative thinking in inference-heavy workloads.

Avoid suppressing reflection tokens when your prompt requires deep iterative thinking where deliberation steps are critical to correctness. NOWAIT optimization is designed for scenarios prioritizing rapid responses over extended reasoning.