🥇Top ML Papers of the Week

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

This issue highlights the top ML Papers of the Week (Feb 20 - Feb 26).

---

1). LLaMA - a 65B parameter foundation model released by Meta AI; relies on publicly available data and outperforms GPT-3 on most benchmarks despite being 10x smaller. (paper)

Guillaume Lample @GuillaumeLample

Today we release LLaMA, 4 foundation models ranging from 7B to 65B parameters. LLaMA-13B outperforms OPT and GPT-3 175B on most benchmarks. LLaMA-65B is competitive with Chinchilla 70B and PaLM 540B. The weights for all models are open and available at research.facebook.com/publications/l… 1/n

2) Composer - a 5B parameter creative and controllable diffusion model trained on billions (text, image) pairs. (paper)

AK @_akhaliq

Composer is a large (5 billion parameters) controllable diffusion model trained on billions of (text, image) pairs github: github.com/damo-vilab/com… paper: arxiv.org/abs/2302.09778 project page: damo-vilab.github.io/composer-page/

3) Hindsight Instruction Relabeling - an alternative algorithm to train LLMs from feedback; the feedback is converted to instruction by relabeling the original one and training the model, in a supervised way, for better alignment. (paper)

Tianjun Zhang @NeurIPS 2022 @tianjun_zhang

Can we use purely supervised learning for RLHF using large language models? We introduce HIR (Hindsight Instruction Relabeling), which achieves impressive results using FLAN-T5 on hard BigBench tasks!

4). Active-Prompt - a prompting technique to adapt LLMs to different task-specific example prompts (annotated with human-designed chain-of-thought reasoning); this process involves finding where the LLM is most uncertain and annotating those. (paper)

John Nay @johnjnay

Active Prompting for LLMs -Most Chain-of-Thought examples are pulled from a fixed set -Instead, to adapt to diff tasks 1) Find where LLM is most uncertain 2) Annotate those -State-of- the-art on complex reasoning tasks Paper arxiv.org/abs/2302.12246 Code github.com/shizhediao/act…

5). Modular Deep Learning - a survey offering a unified view of the building blocks of modular neural networks; it also includes a discussion about modularity in the context of scaling LMs, causal inference, and other key topics in ML. (paper)

Sebastian Ruder @seb_ruder

In our new survey “Modular Deep Learning”, we provide a unified taxonomy of the building blocks of modular neural nets and connect disparate threads of research. 📄 arxiv.org/abs/2302.11529 📢 ruder.io/modular-deep-l… 🌐 modulardeeplearning.com w/ @PfeiffJo @licwu @PontiEdoardo

6). Recitation-Augmented LMs - an approach that recites passages from the LLM’s own memory to produce final answers; shows high performance on knowledge-intensive tasks. (paper)

Zhiqing Sun @EdwardSun0909

How can LLMs such as GPT-3 and ChatGPT achieve greater factual accuracy without relying on an external retrieval search engine? Our #ICLR2023 paper shows that recitation can help - like humans! Recitation-Augmented Language Models arxiv.org/abs/2210.01296 1/N

7). LLMs to Optimize Code - an approach that uses LLMs to suggest functionally correct, performance-improving code edits. (paper)

Jay Hack @mathemagic1an

AI systems can optimize their own code (!) "Learning Performance-Improving Code Edits" arxiv.org/pdf/2302.07867… Introduces a dataset of (before, after) code optimizations + describes methods for building code optimizing LLMs My takeaways 👇

8). Prompt Injection Threats - a comprehensive analysis of novel prompt injection threats to application-integrated LLMs. (paper)

elvis @omarsar0

If you are building with LLMs, it's good to know about novel adversarial prompting techniques. This paper presents an analysis of the topic. It also discusses attack vectors when augmenting LLMs with retrieval and API calling abilities. arxiv.org/abs/2302.12173 https://t.co/wqcW02Ccoe

9). Aligning Text-to-Image Models using Human Feedback - proposes a fine-tuning method to align generative models using human feedback. (paper)

Kimin @kimin_le2

📄Can "learning from human feedback" improve text-to-image models? I'm excited to share "Aligning Text-to-Image Models using Human Feedback" 📝 arxiv.org/abs/2302.12192 1/N

10). MERF - a memory-efficient radiance field representation for real-time view synthesis of large scenes in a browser. (paper)

AK @_akhaliq

MERF: Memory-Efficient Radiance Fields for Real-time View Synthesis in Unbounded Scenes abs: arxiv.org/abs/2302.12249 project page: merf42.github.io

---

See you next week for another round of awesome ML papers!