AutoThinkSQL Dynamically Bypasses Reasoning to Reduce Latency

AutoThinkSQL uses SFT and DPO to enable models to dynamically toggle between simple lookups and complex Chain-of-Thought reasoning. On Qwen3-Coder-30B-A3B, the method reduced average latency by up to 17.1% and output tokens by 24.6% on Spider and BIRD benchmarks.