llm reasoning tutorial

Reasoning with (Large) Language Models

---

Presentations

• BHI 2023 Tutorial (Oct 2023)

• ACL 2023 Tutorial (July 2023)

TLDR

• A central theme of both these tutorials is a set of key techniques for prompting large language models. Combining these techniques captures many of the recent works on prompting LLMs.

---

---

Notebooks

• Language Models: An overview of language models, focusing on GPT-2's architecture and operation. Contains interactive demonstrations of how these models estimate word sequence probabilities, predict the next tokens, and generate coherent text.

• Prompting Basics: Demonstrates few-shot and zero-shot prompting. Utilizing the OpenAI API, allows playing with parameters like temperature, stop tokens, and maximum tokens, observing their impact on language model outputs.

• Prompting Key Techniques: Shows a typical workflow for prompting language models, including choosing between prompts (direct and chain-of-thought prompts) and parsing outputs for evaluation. It also includes a demonstration of the memory assisted prompting, where the prompt is updated with the generated text at each step.

Other resources

• My blog on prompting large language models, used for a tutorial at the CMU LTI Seminar on Large Language Models (April 2023). In the blog, I draw an analogy between LLMs as stochastic compilers and prompting as a high-level language for specifying tasks.