From fc146824c8cc1bcc9e3e52d05a5ddc3adc4d0b93 Mon Sep 17 00:00:00 2001 From: Ted Sanders Date: Thu, 25 May 2023 12:08:04 -0700 Subject: [PATCH] adds two more interesting papers --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index 5d4ceac8..f59e6ad7 100644 --- a/README.md +++ b/README.md @@ -105,11 +105,13 @@ People are writing great tools and papers for improving outputs from GPT. Here a - [Tree of Thoughts: Deliberate Problem Solving with Large Language Models (2023)](https://arxiv.org/abs/2305.10601): Searching over trees of step by step reasoning helps even more than voting over chains of thought. It lifts `GPT-4`'s scores on creative writing and crosswords. - [Language Models are Zero-Shot Reasoners (2022)](https://arxiv.org/abs/2205.11916): Telling instruction-following models to think step by step improves their reasoning. It lifts `text-davinci-002`'s score on math word problems (GSM8K) from 13% to 41%. - [Large Language Models Are Human-Level Prompt Engineers (2023)](https://arxiv.org/abs/2211.01910): Automated searching over possible prompts found a prompt that lifts scores on math word problems (GSM8K) to 43%, 2 percentage points above the human-written prompt in Language Models are Zero-Shot Reasoners. +- [Reprompting: Automated Chain-of-Thought Prompt Inference Through Gibbs Sampling (2023)](https://arxiv.org/abs/2305.09993): Automated searching over possible chain-of-thought prompts improved ChatGPT's scores on a few benchmarks by 0–20 percentage points. - [Faithful Reasoning Using Large Language Models (2022)](https://arxiv.org/abs/2208.14271): Reasoning can be improved by a system that combines: chains of thought generated by alternative selection and inference prompts, a halter model that chooses when to halt selection-inference loops, a value function to search over multiple reasoning paths, and sentence labels that help avoid hallucination. - [STaR: Bootstrapping Reasoning With Reasoning (2022)](https://arxiv.org/abs/2203.14465): Chain of thought reasoning can be baked into models via fine-tuning. For tasks with an answer key, example chains of thoughts can be generated by language models. - [ReAct: Synergizing Reasoning and Acting in Language Models (2023)](https://arxiv.org/abs/2210.03629): For tasks with tools or an environment, chain of thought works better you prescriptively alternate between **Re**asoning steps (thinking about what to do) and **Act**ing (getting information from a tool or environment). - [Reflexion: an autonomous agent with dynamic memory and self-reflection (2023)](https://arxiv.org/abs/2303.11366): Retrying tasks with memory of prior failures improves subsequent performance. - [Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP (2023)](https://arxiv.org/abs/2212.14024): Models augmented with knowledge via a "retrieve-then-read" can be improved with multi-hop chains of searches. +- [Improving Factuality and Reasoning in Language Models through Multiagent Debate (2023)](https://arxiv.org/abs/2305.14325): Generating debates between a few ChatGPT agents over a few rounds improves scores on various benchmarks. Math word problem scores rise from 77% to 85%. ## Contributing