Happy Halloween! Be careful out there. Look what I found inside some trick-or-treat candy.

(Credit to Arthur Spirling for the meme idea)

Detecting Pretraining Data from Large Language Models

We propose Min-K% Prob, a simple and effective method that can detect whether if a LLM was pretrained on the provided text without knowing the pretraining data.

proj: swj0419.github.io/detect-pretrai…
abs: arxiv.org/abs/2310.16789

High-quality data is a key to successful pretrain/finetuning in the GPT era, but manual data curation is expensive💸 We tackle data quality challenges involving large models and datasets with ScAlable Meta leArning (SAMA) #NeurIPS2023 💫

Arxiv: arxiv.org/abs/2310.05674
🧵 (1/n)

Introduce In-Context Pretraining🖇️: train LMs on contexts of related documents. Improving 7B LM by simply reordering pretrain docs
📈In-context learning +8%
📈Faithful +16%
📈Reading comprehension +15%
📈Retrieval augmentation +9%
📈Long-context reason +5%
arxiv.org/abs/2301.12652

Introducing RAD, a cheap and efficient method for using an auxiliary reward model for controlling text generation that can match the performance of methods that update the LM.

📝arxiv.org/abs/2310.09520
💾github.com/haikangdeng/RAD
🧵⬇️

1/

So, I wrote about my experience as a Senior Area Chair for EMNLP, hoping to offer some insights on the processes around conference paper decisions.

The Impossible Task of Conference SACs/PCs
or
How I lost 3 Nights of Sleep
gist.github.com/antonisa/2158d…

comments/feedback welcome!

It was so interesting to see participants' perceptions of 'paradigm shifts' paralleled across eras and cycles of NLP, and at the same time nothing until recent years had reached quite the level of *47%* of ACL papers in 2021 citing BERT

We all know that “recently large language models have”, “large language models are”, and “large language models can.” But why LLMs? How did we get here? (where is “here”?) What forces are shaping NLP, and how recent are they, actually?

To appear at EMNLP: arxiv.org/abs/2310.07715

Today I want to take a break from sharing research to share a personal story instead. It’s a story about my name, why I once decided to quit academia, why I came back, what I learnt from it, and why I’m grateful to have an audience here on Twitter.

Save the Date! The LTI will hold an info-session for applicants to our PhD and MLT programs on Wednesday, November 8 from 12pm -1pm, ET. Faculty and students will answer questions about the application process. More information will be posted soon!

Why does RLHF make outputs longer?

arxiv.org/pdf/2310.03716… w/ Tanya Goyal Jiacheng Xu Greg Durrett

On 3 “helpfulness” settings
- Reward models correlate strongly with length
- RLHF makes outputs longer
- *only* optimizing for length reproduces most RLHF gains

🧵 below:

Retrieval-augmented LMs are not robust to irrelevant context. Retrieving entirely irrelevant context can throw off the model, even when the answer is encoded in its parameters!

In our new work, we make RALMs more robust to irrelevant context.

arxiv.org/abs/2310.01558

🧵[1/7]

Nucleus and top-k sampling are ubiquitous, but why do they work?
John Hewitt, Alexander Koller, Swabha Swayamdipta, Ashish Sabharwal and I explain the theory and give a new method to address model errors at their source (the softmax bottleneck)!
📄 arxiv.org/abs/2310.01693
🧑‍💻 github.com/mattf1n/basis-…