John Wieting
@johnwieting2
Followers
549
Following
2K
Media
2
Statuses
105
Senior Research Scientist @GoogleDeepMind 🧠. PhD @LTIatCMU.
Joined December 2015
Work led by @YapeiChang with @kalpeshk2011, @houmansadr, myself, and @MohitIyyer. Work done at @manningcics, @UMassNLP with collaboraters from @GoogleAI, @GoogleDeepMind. #NLProc.
0
0
4
Interested in robust watermarks for detecting LLM-generated text that do not require logit access to the underlying models? Check out our new paper to be presented at @emnlpmeeting! We propose PostMark, a black-box watermarking scheme that is more robust to paraphrasing attacks.
Is it possible to have a watermark that reliably detects LLM-generated text, is robust to paraphrasing attacks, preserves quality, and can be applied to any LLM without access to logits? Check out PostMark, a method with all these properties!. . 🧵below:
1
2
12
0
0
0
Check out our new paper1 We propose a watermarking scheme for LLMs that works with black-box access, is distortion-free, has provable guarentees, and can be chained with secret keys. Able to outperform white-box methods in comprehensive experiments.
1
11
52
Congratulations to UMD!.
.@UofMaryland's CS dept is welcoming eight new tenure-track faculty members! With expertise in cutting-edge areas, they'll boost our academic & research capabilities. Welcome aboard! 🐢📚. Read more:.
0
0
2
RT @YapeiChang: Is it possible to have a watermark that reliably detects LLM-generated text, is robust to paraphrasing attacks, preserves q….
0
28
0
RT @nandan__thakur: Excited to share that SWIM-IR has been accepted at #NAACL2024! 🍻. I'm quite delighted with this work as it was complete….
0
3
0
There are so many accounts (bots?) on X/Twitter posting ChatGPTish responses or the same irrelevant responses across posts. It seems fairly obvious how these could be filtered out. I wonder why nothing is done. I thought it'd improve over time, but it really hasn't.
0
0
4
RT @aclmeeting: ACL announcement: ."The ACL Executive Committee has voted to significantly change ACL's approach to protecting anonymous pe….
0
186
0
This was work done both at @GoogleAI and @UMassAmherst with @kalpeshk2011, @yixiao_song, @mar_kar_, @MohitIyyer, and myself. #NLProc.
0
0
1
Arxiv: Models (Hugging Face versions strongly recommended): . Code:
github.com
Official repository for our NeurIPS 2023 paper "Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense" (https://arxiv.org/abs/2303.13408). - ma...
1
0
1
Come by our poster this evening! @kalpeshk2011, @yixiao_song, and I will be presenting. Links to paper and models below 👇.
1
0
1
3. A state-of-the-art paraphrasing model that paraphrases long-form text, can use context to paraphrase text within delimiters, controls output output lexical and syntactic diversity, and is validated by copious human evaluations . .
1
0
2
Today at #NeurIPS2023, If you want to lean more about:. 1. Robustness of detectors and watermarks to paraphrase attacks (spoiler alert: needs improvement). 2. An alternative detection approach using simple retrieval methods. and . .
To detect text written by LMs like #ChatGPT, many methods have recently emerged: DetectGPT, watermarks, GPTZero. We present a paraphrasing attack that can drop their detection rates to <10%. To defend against it, we propose detection with retrieval. 🧵👇
1
4
20
Our work on cross-lingual and multilingual attribution will be presented at #EMNLP2023 in Singapore!. We have also released our dataset of ~10k 3-way annotations over 5 typologically diverse languages. Arxiv: Dataset:
arxiv.org
Trustworthy answer content is abundant in many high-resource languages and is instantly accessible through question answering systems, yet this content can be hard to access for those that do not...
We all want accurate responses from our QA systems, and this need becomes especially vital when interacting with text in languages unfamiliar to us, rendering answer verification reliant on translation. This challenge is particularly felt by speakers of low-resource languages.
1
8
24
RT @ben_mlr: Excited to be presenting our work on.**Evaluating and Modeling Attribution for Cross-Lingual Question Answering** at #EMNLP202….
arxiv.org
Trustworthy answer content is abundant in many high-resource languages and is instantly accessible through question answering systems, yet this content can be hard to access for those that do not...
0
5
0
Later today:. Sam Altman: "I'm pleased to announce I'm looking to raise venture capital for my new startup: ClosedAI.".
2
0
6
RT @sunjiao123sun_: Can LLMs generate exact 5 words? No.How about 5 sentences? No.How about 5 paragraphs? No.🤷🏻♀️.In .
arxiv.org
While recent studies have looked into the abilities of large language models in various benchmark tasks, including question generation, reading comprehension, multilingual and etc, there have been...
0
76
0