Evaluation same question :
Gemini : https://github.com/Prasann2004/Gemini-3x-Blend/assets/83667133/f960ad1b-c8ea-48d7-bc22-a64068fce0df
Gemini 3x Blend : https://github.com/Prasann2004/Gemini-3x-Blend/assets/83667133/5f605479-4a1b-4c78-ab3e-02f9800c80e8
I implemented the following techniques to improve the output from Gemini :
Corrective Retrieval Augmented Generation integrates a lightweight retrieval evaluator designed to assess the quality of retrieved documents in relation to a given query. This evaluator provides a confidence score that informs subsequent retrieval actions, allowing for adaptive adjustments in knowledge retrieval strategies. Importantly, CRAG extends beyond static and limited corpora by leveraging large-scale web searches to augment retrieval results, thereby expanding the breadth and depth of available information.
LLM-Blender is a novel framework designed to enhance the performance of large language models (LLMs) by combining the strengths of multiple models. It consists of two key modules: PairRanker and GenFuser. PairRanker utilizes a specialized method for pairwise comparison of candidate outputs, leveraging cross-attention encoders to determine the superior one based on input text. Results show that PairRanker correlates highly with ChatGPT-based ranking. GenFuser merges the top-ranked candidates to generate an improved output by capitalizing on their strengths and addressing weaknesses. The framework is evaluated using a benchmark dataset called MixInstruct, showcasing significant performance improvements over individual LLMs and baseline methods across various metrics.
SELF-DISCOVER tries to employ a method which we all use in day to day baisis to solve complex problems .That is to break problems into subparts and solve it step by step .