📢 Introducing USB: A Unified Summarization Benchmark Across Tasks and Domains! 🧵

🌟 Our benchmark consists of 8 diverse tasks, including evidence extraction, factuality and controllable summarization tasks.

🤔Can we reduce the cost of reasoning in large language models while maintaining accuracy? Introducing our new paper: 'Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning with LLMs' a dynamic alternative to Self-Consistency!

🌐: sample-step-by-step.info

🧵

📢Meet Self-Refine: Iterative Refinement with Self-Feedback, our latest research on LLMs!

🤖Self-Refine enables LLMs to refine their outputs via self-generated feedback, without human intervention. Explore the potential of self-refinement in AI!

selfrefine.info

LLMs Can Iteratively Self-Refine

-LLM creates draft
-Provides its own feedback
-Iteratively refines

On all 7 eval tasks
(review & code rewriting
toxicity removal
responses
acronyms
stories
etc.)
outputs are preferred by humans & by automated metrics

arxiv.org/abs/2303.17651

In a new preprint, we investigate how summarization models are affected by the presence of noise in the input, and propose a way to improve it without prior knowledge of the type of noise that can be present (as it often happens in real-world scenarios)
arxiv.org/abs/2212.09928