End-to-end DAG-Path Prompting (EEDP) is a new technique to boost LLMs' reasoning by "flattening" graphs into text that LLMs can read. It preprocesses graphs into Directed Acyclic Graphs (DAGs), extracting backbone paths for conversion to structured text. This preserves long-distance connections and compresses redundant parts, allowing LLMs to better process graph data. Learn more

Instance-adaptive Zero-shot CoT is a new technique to improve LLM reasoning on unseen data:

1. Collect diverse prompts with different reasoning styles.
2. Use saliency analysis to evaluate how well each prompt handles the question.
3. Find the best reasoning path using IAP-ss (sequential substitution) or IAP-mv (majority vote) strategy.

This allows LLMs to adapt prompts dynamically for better performance on new instances. Learn more

Conditional Prompt Learning (CoCoOp) is a new method to adapt vision-language models like CLIP to downstream datasets. It introduces a Meta-Net that generates dynamic, conditional prompts tailored to each image, allowing better adaptation to new or unseen classes. CoCoOp is ideal for tasks like image classification with changing data distributions, cross-dataset transfer, and fine-grained classification. Learn more