🥇Top ML Papers of the Week

The top ML Papers of the Week (Mar 20 - Mar 26)

1). Sparks of AGI - a comprehensive investigation of an early version of GPT-4 when it was still in active development by OpenAI. (paper)

Sebastien Bubeck @SebastienBubeck

At @MSFTResearch we had early access to the marvelous #GPT4 from @OpenAI for our work on @bing. We took this opportunity to document our experience. We're so excited to share our findings. In short: time to face it, the sparks of #AGI have been ignited. arxiv.org/abs/2303.12712

---

2). Reflexion - proposes an agent with dynamic memory and self-reflection capabilities to enhance its existing reasoning trace and task-specific action choice abilities. (paper)

Siqi Chen @blader

🤯 this paper demonstrates you can improve gpt4 performance an astounding 30% by asking gpt4 to reflect on “why were you wrong?”, and generate a new prompt for itself taking that reason into account until it is correct. this is how humans learn! arxiv.org/pdf/2303.11366… 👇

---

3). GPT-4 for Medical Challenge Problems - shows that GPT-4 exceeds the passing score on USMLE by over 20 points and outperforms GPT-3.5 as well as models specifically fine-tuned on medical knowledge (Med-PaLM, a prompt-tuned version of Flan-PaLM 540B). (paper)

DAIR.AI @dair_ai

Capabilities of GPT-4 on Medical Challenge Problems Shows that GPT-4 exceeds the passing score on USMLE by over 20 points and outperforms GPT-3.5 as well as models specifically fine-tuned on medical knowledge (Med-PaLM, a prompt-tuned version of Flan-PaLM 540B).

---

4). GPTs are GPTs - investigates the potential implications of GPT models and related systems on the US labor market. (paper)

JB Rubinovitz @rubinovitz

The "Will GPT automate all the jobs?" paper is out With participation from @OpenAI, OpenResearch and @Penn 🧵 1/9

---

5). CoLT5 - a long-input Transformer model that employs conditional computation, devoting more resources to important tokens in both feedforward and attention layers. (paper)

Aran Komatsuzaki @arankomatsuzaki

CoLT5: Faster Long-Range Transformers with Conditional Computation Achieves: - stronger performance than LongT5 with much faster training and inference - SOTA on the SCROLLS benchmark - strong gains up to 64k input length arxiv.org/abs/2303.09752

---

6). Artificial muses - compares human-generated ideas with those generated by generative AI chatbots like ChatGPT and YouChat; reports that 9.4% of humans were more creative than GPT-4 and that GAIs are valuable assistants in the creative process. (paper)

Ethan Mollick @emollick

GPT-4 was more creative than all but 9.4% of humans tested in this new paper It gave the Alternative Uses Test (a measure of creativity where you need to come up with unique uses of everyday objects) to AI & 100 people. GPT-4 got high ratings from judges. arxiv.org/abs/2303.12003 https://t.co/T6mUOs3zxT

---

7). Analysis of GPT-3 and GPT-3.5 - a comprehensive capability analysis of GPT series models; evaluates performance on 9 natural language understanding tasks using 21 datasets. (paper)

Aran Komatsuzaki @arankomatsuzaki

A Comprehensive Capability Analysis of GPT-3 and GPT-3.5 Series Models arxiv.org/abs/2303.10420

---

8). Context-faithful Prompting for LLMs - presents a prompting technique that aims to improve LLMs' faithfulness using strategies such as opinion-based prompts and counterfactual demonstrations. (paper)

elvis @omarsar0

Context-faithful Prompting for Large Language Models Presents a neat prompting technique that aims to improve LLMs' faithfulness using strategies such as opinion-based prompts and counterfactual demonstrations. arxiv.org/abs/2303.11315

---

9). Text2Room - a method for extracting room-scale textured 3D meshes from 2D text-to-image models. (paper)

AK @_akhaliq

Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models abs: arxiv.org/abs/2303.11989 project page: lukashoel.github.io/text-to-room/

---

10). PanGu-Σ - a trillion parameter language model with sparse heterogeneous computing. (paper)

Aran Komatsuzaki @arankomatsuzaki

PanGu-Σ: Towards Trillion Parameter Language Model with Sparse Heterogeneous Computing Presents a sparse LM with 1T parameters trained over 329B tokens arxiv.org/abs/2303.10845