Reasoning Prompting Techniques that no one talks about

As a researcher in the AI ​​Evolution, I saw that proper requirement techniques create superior results. I generally focus on AI and large language models in general. Five years ago, the field of data science, CNN and Transformers emphasized. The request then remained dark. Now it serves as an essential component for context engineering to refine and control LLMs and agents. I have experimented and still play with various prompt styles to sharpen LLM answers. For me, three techniques stand out: chain of thought (cot): I integrate phrases like "let's think step by step". This approach increases the accuracy of complex mathematical problems. It is characterized in multi -stage challenges at companies such as Google Deepmind. Nevertheless, it increases the token three to five times. Self -consistency: This method creates several argumentation paths and uses the majority. It reduces errors in operating systems by scanning five to ten outputs at 0.7 temperature. It delivers 97.3% accuracy for Math-500 with Deepseek R1 models. Despite higher calculation requirements, it proves to be valuable for precision -critical tasks. REACT: It combines argumentation with actions in monument obescycles. This anchors the answers to external data sources. It achieves up to 30% higher accuracy for sequential questions to answer benchmarks. Success is based on robust API integrations, as can be seen in tools at companies like IBM. Now, with 2025 starts, the comparison of these methods is becoming more convincing. Openai presented the open weight model GPT-OS-12M20B in August. Xai followed by open sourcing grok 2.5 weights shortly afterwards. I am really striving to experiment and create workflows in which I use a new open source model locally. Maybe also create a user interface around you. I also find out to examine evaluation approaches, including the accuracy assessment, cost outlets and the latency-focused scorecards. What thoughts do you have about the request of techniques and their evaluation methods? And did you experiment with open source publications?