Joking aside DeepSeek is really impressive, showing that scaling is **not** all you need

Findings:
❓ Privacy-preservation at inference-time is really underexplored!
🔍 LMs struggle to prevent PII leakage in their summaries.
👩‍⚖️ Human evaluations reveal privacy risks that metrics may overlook.

Paper w/ @naletras.bsky.social and Ning Ma
Cc. @sltcdt.bsky.social

*word

Having a large number of short ARR cycles a year doesn’t make sense to me. Since there is no arxiv anonymity period anymore, we can move to less rushed cycles, more engagement during discussion period and better review/metareview quality given the extra time

Fun fact: If I remember correctly we got desk rejected by Plos One because they couldn't find reviewers with appropriate expertise. I don't think that we even tried *ACL because we didn't have a "novel" END-TO-END model. PeerJ CS got some extremely high quality reviewers though!

The paper was published eight years ago:
peerj.com/articles/cs-...

and apart from inspiring further research in NLP and #legaltech, it also resulted in the creation of the NLLP Workshop and its amazing community 5/5

The initial idea was conceived at a coffee shop in Sheffield (the Couch, still operating) in 2014, trying to explain NLP and text classification to Dimitris 4/n

Looking back, it still amazes me that this work was just a side project, not specifically funded by a grant and published in a less prestigious outlet (although a UCL press release helped enormously) 3/n

For the first time, we showed that it is possible to just use the fact descriptions of legal cases to train classifiers - SVMs (ehm what?!) acting as 👩‍⚖️- for predicting judicial decisions. This sparked huge interest (and 🔥debates) in the use AI in the legal domain 2/n

🙋🏻‍♂️