[1910.13461] BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
6
Public whispers
6
Contributors
2026-07-16 10:05:46
First whispered
Public whispers on this page
Text Highlight2026-07-16 16:29:46
Original Highlight Excerpt
"arXivLabs: experimental projects with community collaborators"
Whisper Note
Cool, so like a playground for weird new ideas?
Text Highlight2026-07-16 13:26:46
Original Highlight Excerpt
"corrupting text with an arbitrary noising function"
Whisper Note
So basically a fancy autoencoder, but it works better than expected.
Text Highlight2026-07-16 13:17:46
Original Highlight Excerpt
"corrupting text with an arbitrary noising function"
Whisper Note
Interesting how they just toss in any noise and still get good results.
Text Highlight2026-07-16 10:23:46
Original Highlight Excerpt
"denoising autoencoder for pretraining sequence-to-sequence models"
Whisper Note
I wonder how much the sentence shuffling actually matters vs. the infilling.
Text Highlight2026-07-16 10:14:46
Original Highlight Excerpt
"denoising autoencoder for pretraining sequence-to-sequence models"
Whisper Note
The mask-token infilling trick is what caught my eye; that's a clever corruption.
Text Highlight2026-07-16 10:05:46
Original Highlight Excerpt
"denoising autoencoder for pretraining sequence-to-sequence models"
Whisper Note
So it's like BERT and GPT had a baby, but for seq2seq tasks. Makes sense.
Share this page's whispers
Short link
https://domwhisper.com/s/87f5786bc5beEmbed snippet
<iframe src="https://domwhisper.com/embed/87f5786bc5be" width="100%" height="480" style="border:0;border-radius:16px" loading="lazy"></iframe>See what people are discussing on arxiv.org
Install DomWhisper to view live whispers as you browse, and join the discussion.
Get the extension