Large Language Models (LLMs) are known for their impressive generative abilities, crafting human-like text with remarkable fluency. But what about their capacity to evaluate their own work? A new research paper, "Direct-Inverse Prompting: Analyzing LLMs' Discriminative Capacity in Self-Improving Generation," delves into this intriguing area of self-assessment. The study explores how LLMs can leverage their 'discriminative capability' to reduce uncertainty in their own outputs, essentially grading their own tests. The core idea is simple yet powerful: Given a set of possible answers generated by the LLM for a given problem, can the LLM itself identify the most likely correct answer? Researchers experimented with various "discriminative prompts" to evaluate this self-assessment ability. They explored 'direct' prompts, which ask the LLM to directly identify correct answers, 'inverse' prompts, which ask it to point out incorrect ones, and a 'combination' approach leveraging both. The results, tested across open-source and closed-source LLMs like GPT-4 and Llama, revealed interesting insights. Top-tier commercial LLMs like GPT-4 demonstrated considerable skill in self-assessment, significantly improving their performance when employing these discriminative prompts. This suggests that, given multiple attempts at an answer, these LLMs can effectively discern right from wrong, boosting their overall accuracy. However, the research also highlighted limitations. Open-source LLMs, especially those not specifically trained for instruction following, struggled with these discriminative prompts, particularly with the 'inverse' approach involving negation. This points towards the importance of continued refinement in training methods to enhance LLMs’ critical reasoning and evaluative abilities. The research opens up a new dimension in understanding LLM capabilities. It underscores the potential of self-assessment to refine LLM outputs and pave the way for more reliable and accurate AI-generated content. However, it also exposes the challenges in imbuing LLMs with robust critical thinking skills, particularly in open-source models, indicating a critical area for future research.