Karolina Naranjo

Ph.D. Candidate @uvadatascience | Law+AI ⚖️ | NLP 🤖| Views are mine | She/her/ella

[Karolina Naranjo @karolinaranjo]

Thrilled to share our #EMNLP2025 paper!

𝐌𝐎𝐃𝐄𝐋𝐂𝐈𝐓𝐈𝐙𝐄𝐍𝐒: Community-informed toxicity detection 6.8K posts, 40K diverse annotations show toxicity isn't universal—it's shaped by norms, identity & context.

📄https://t.co/hVjMO2iTfz

[Ashima Suvarna🌻 @suvarna_ashima]

1/ 🧵
New #EMNLP2025 Paper !!
Toxicity detection is subjective; shaped by norms, identity, & context. Existing models and dataset overlook this nuance.
Enter MODELCITIZENS: a new dataset designed to address this.
✔️ 6.8K posts, 40K annotations across diverse groups
✔️ Context-augmented scenarios
✔️ New finetuned models (LLAMACITIZEN-8B, GEMMACITIZEN-12B) beat GPT-4o-mini by 5.5%

(🚨 Offensive Content Warning)

[Joël Niklaus @joelniklaus]

Great to see @nllpworkshop already celebrating its 8th occurrence, this time in Budapest collocated with @emnlpmeeting!

Go submit your legal AI papers there 🔥

Kudos to the organizing committee for making this happen year after year!

@nikaletras (University of Sheffield), Leslie Barrett (Bloomberg), @KiddoThe2B (University of Copenhagen), @CatalinaGoanta (Utrecht University), Daniel Preotiuc-Pietro (Bloomberg), @gerasimoss (Maastricht University), @shanshan_xu3 (University of Copenhagen)

[Karolina Naranjo @karolinaranjo]

@WiMLworkshop Can you make it public?

[alex zhang @a1zhang]

Much like the switch in 2025 from language models to reasoning models, we think 2026 will be all about the switch to Recursive Language Models (RLMs).

It turns out that models can be far more powerful if you allow them to treat *their own prompts* as an object in an external environment, which they understand and manipulate by writing code that invokes LLMs!

Our full paper on RLMs is now available—with much more expansive experiments compared to our initial blogpost from October 2025!

https://t.co/x47pIfIkTb

[MikaStars★✈️ICML @MikaStars39]

Stop using LoRA for RLVR!!!
New paper released👉Evaluating Parameter Efficient Methods for RLVR
📖Alphaxiv: https://t.co/VCNTG4TZvP
💻Github: https://t.co/bUkRjjzBvc

Is standard LoRA truly the optimal choice for Reinforcement Learning?. We present the first large-scale evaluation of over 12 PEFT methodologies using the DeepSeek-R1-Distill family on complex mathematical reasoning benchmarks.

Key Finding: Standard LoRA is suboptimal. Structural variants such as DoRA, AdaLoRA, and MiSS consistently outperform standard LoRA. Notably, DoRA (46.6% avg. accuracy) even surpasses full-parameter fine-tuning (44.9%) across multiple benchmarks.

The failure of SVD-based initialization. 
Strategies like PiSSA and MiLORA experience significant performance degradation or total training collapse. This is due to a fundamental "spectral misalignment": these methods force updates on principal components, while RLVR intrinsically operates in the off-principal regime.

The Expressivity Floor. 
While RLVR can tolerate moderate parameter reduction, extreme compression (e.g., VeRA, IA³, or Rank-1 adapters) creates an information bottleneck. Reasoning tasks require a minimum threshold of trainable capacity to successfully reorient policy circuits.

Recommendations for the community:
a. Move beyond the default adoption of standard LoRA.
b. Prioritize geometry-aware adapters like DoRA that decouple magnitude and direction.
c. Avoid SVD-informed initializations for RL tasks.

[Sebastian Raschka @rasbt]

Efficiency and performance tweaks in the transformer architecture usually focus(ed) on the normalization, attention, and FFN modules.

Well, here is a holiday gift from DeepSeek (https://t.co/ow1RpEG2Bv).
Finally some improvements of the residual path as well.

[Karolina Naranjo @karolinaranjo]

@Yoshua_Bengio @yaringal @uiuc_aisecure @acidflask @evijit This event wouldn't be possible without our incredible organizing team: @_cagarwal, @jiaqiMa, Sarah Tan, @hima_lakkaraju, Junwei Deng, @PingbangHu, Eileanor LaRocco, and @ShichangZhang

See you in Room 1AB on Dec 7th!

#RegulatableML #NeurIPS2025 #NeurIPSanDiego

[Karolina Naranjo @karolinaranjo]

Join us at the #RegulatableML workshop at #NeurIPS2025 #NeurIPSanDiego to bridge the gap between AI regulation and real-world practice!

🗓 Date: Dec 7, 2025
⏰ Time: 8:45 am - 5:30 pm (PST)
📍 Location: Upper Level Room 1AB, San Diego, CA

More info: https://t.co/NtaC9wV0ed

[Karolina Naranjo @karolinaranjo]

@Yoshua_Bengio @yaringal @uiuc_aisecure 𝐏𝐚𝐧𝐞𝐥: Get ready for "The AI Wars: Regulation, Open Science, and the Race for Global Power", featuring Rich Caruana, Jiahao Chen @acidflask , Ian Eisenberg, Avijit Ghosh @evijit, and Bo Li @uiuc_aisecure .

[Karolina Naranjo @karolinaranjo]

@Yoshua_Bengio @yaringal @uiuc_aisecure @acidflask @evijit 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡: We’re also hosting 8 oral presentations and a poster session with 53 accepted papers.

[Karolina Naranjo @karolinaranjo]

#neurips2025 #NeurIPS Call for AI researchers!

Attend the 𝟑𝐫𝐝 𝐑𝐞𝐠𝐮𝐥𝐚𝐭𝐚𝐛𝐥𝐞 𝐌𝐋 𝐖𝐨𝐫𝐤𝐬𝐡𝐨𝐩, Dec 7, San Diego, CA. Bridging ML safety research with regulatory frameworks.

Join us! 🔗https://t.co/NtaC9wV0ed

[UVA School of Data Science @uvadatascience]

Today, the UVA School of Data Science was packed with new and familiar faces for our annual Datapalooza event. A special thank you to our keynote speaker Danielle Citron and all of our presenters for Datapalooza 2025: Truth and Accountability in the Age of AI.

[Ryan Calo @rcalo]

First time holding my book 🥹 Thanks @UCLA!

[Saadia Gabriel @GabrielSaadia]

Also not to be missed: Sophie will be presenting our poster for ModelCitizens the day before (session 2, 11-12:30pm).

[Tom Hartvigsen @tom_hartvigsen]

📢Please retweet: We are hiring a **Postdoc** at UVA to work on Continually Monitoring and Updating Multi-modal Medical AI Models!

Great opportunity to design impactful methods alongside great collaborators @_ahmedmalaa and @RoxanaDaneshjou

More info: https://t.co/a3Qs5X0pgy

[Karolina Naranjo @karolinaranjo]

@PingbangHu Yep!

[Joël Niklaus @joelniklaus]

⏰✨ Good news for all the fashionably late submitters out there! We’ve extended the deadline for the First International Workshop on Data Mining and AI for Law (DMAIL 2025) 🎉.

Instead of September 1, you now have until September 10 to polish, perfect, and procrastinate (just a little bit longer 😉).

So if your paper was 90% ready, congrats—it’s now magically 100% on time. 🚀

🔗 Workshop details: https://t.co/mP44wDzBWD

[Karolina Naranjo @karolinaranjo]

Big shoutout to the dream team! 🤩@suvarna_ashima for her outstanding leadership, our fantastic collaborators @christinachanc @hamidpalangi & Sophie, and huge thanks to our professors @GabrielSaadia & @tom_hartvigsen for their invaluable guidance!

[Karolina Naranjo @karolinaranjo]

Thrilled to share our #EMNLP2025 paper!

𝐌𝐎𝐃𝐄𝐋𝐂𝐈𝐓𝐈𝐙𝐄𝐍𝐒: Community-informed toxicity detection 6.8K posts, 40K diverse annotations show toxicity isn't universal—it's shaped by norms, identity & context.

📄https://t.co/hVjMO2iTfz

[Ashima Suvarna🌻 @suvarna_ashima]

1/ 🧵
New #EMNLP2025 Paper !!
Toxicity detection is subjective; shaped by norms, identity, & context. Existing models and dataset overlook this nuance.
Enter MODELCITIZENS: a new dataset designed to address this.
✔️ 6.8K posts, 40K annotations across diverse groups
✔️ Context-augmented scenarios
✔️ New finetuned models (LLAMACITIZEN-8B, GEMMACITIZEN-12B) beat GPT-4o-mini by 5.5%

(🚨 Offensive Content Warning)