Constantinos Karouzos

🇬🇷🇪🇺🇬🇧🇮🇹

PhD student @ The University of Sheffield UKRI CDT Speech and Language Tech.
| x ML Eng. @behaviorsignals | 2020 Alumni ECE NTUA

[CoNLL 2026 @conll_conf]

As a result, 12% of the papers in the review pool were desk rejected. Pulling this off in a timely manner was not easy, and was only possible thanks to our wonderful and dedicated reviewers and ACs!

[Athena RC @athenaRICinfo]

Call for Applications for the MSc in “Language Technology” for the Academic Year 2026 - 2027.

𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗱𝗲𝗮𝗱𝗹𝗶𝗻𝗲: 𝟭 𝗝𝘂𝗻𝗲 𝟮𝟬𝟮𝟲

➡️ Learn more: https://t.co/aaydsnVs8f

#AthenaRC #LanguageTechnology #Education

[Greeks in AI @greeksinai]

📢 Are you a PhD student, postdoc, or early-career researcher working in AI?

✅ Reminder: the submission deadline for #GreeksInAI is in 5 days.

📍Eugenides Foundation, Athens
📅15–17 July 2026

Submit now, via: https://t.co/Ew8r8Ufmx9

[Sheffield NLP @SheffieldNLP]

Thanks @suchirsalhan for your seminar today, we loved hearing about your work on BabyLM!

https://t.co/Bx3eIkwMdA

[eaclmeeting @eaclmeeting]

Attention #NLProc researchers, the EACL 2027 website is officially LIVE: https://t.co/0Q0WZVqWV6! 🎉
🇬🇷 Join us in Athens, Greece (Mar 9-13, 2027) at #EACL2027
📅 ARR submission deadline: Aug 6, 2026. Open to all areas of CL/NLP + related fields. Stay tuned for the detailed CfP!

[Emmanuel Macron @EmmanuelMacron]

You know it.

[Ben Clavié @bclavie]

Very odd decision to me, wasn’t the whole point of ARR to give more freedoms once reviews are in and avoid forcing early venue commitment?

[Nikos Aletras @nikaletras]

We need to talk about diversity collapse in post-training...

[AVB @neural_avb]

Here are some that I am just copy pasting from my chat (just SFT and preference tuning, no RLVR/GRPO here):

Self-Rewarding Language Models
UltraFeedback
Few-Shot Preference Optimization
Skywork-Reward
OpenRLHF
JudgeLM
Alignment-Weighted DPO
Where does output diversity collapse in post-training?
Category-Adaptive Safety Alignment
SFT-GO
Online DPO & Iterative Training
f-PO: Generalized Preference Optimization
GOPO: Policy Optimization using Ranked Rewards
One Adapts to Any: Meta Reward Modeling for Personalized LLM Alignment

Some technical reports like the old Llama 3 and Deepseek was there too

[AVB @neural_avb]

Breakdown here

[Constantinos Karouzos @ckarouzos]

@Xingwei__Tan @nikaletras @SheffieldNLP @neural_avb Check the preprint here: https://t.co/ATBAEPzoHQ

[Constantinos Karouzos @ckarouzos]

Thrilled to see such a fantastic breakdown of our latest paper on output diversity collapse in post-training (with @Xingwei__Tan, @nikaletras, @SheffieldNLP)! 🚀

@neural_avb did a phenomenal job summarizing our work. Check it out below! 👇

[AVB @neural_avb]

I have been studying a lot of post-training paper to prepare for my next video.

This morning I am studying a banger from this week: "Where does output diversity collapse in post-training?"

One of the more educational papers I have read recently.

Basically after post-training LLMs often produce less varied answers than their base versions. This paper runs experiments to figure out where exactly this diversity is lost, the effects of distillation vs instruction-tuning vs RLVR, etc.

Their big hypothesis is that the diversity cannot be recovered at inference through sampling techniques or CoT. The collapse irreversibly happens during training.

[Constantinos Karouzos @ckarouzos]

@_Suresh2 @neural_avb If you use narrow, two-teacher distillation (Think), almost all the diversity collapses right at the SFT stage. But if you use broader, multi-source data for SFT (Instruct), the model retains more variety initially, which actually shifts the sharpest diversity drop to the DPO.

[Constantinos Karouzos @ckarouzos]

@_Suresh2 @neural_avb Great question, this is exactly what we wanted to untangle! The short answer is: it depends entirely on your data composition.

[AVB @neural_avb]

https://t.co/81HlbVYz7Q

[Constantinos Karouzos @ckarouzos]

@Xingwei__Tan @nikaletras @SheffieldNLP @neural_avb Here is the full breakdown: https://t.co/Q8HJBrwAHd

[AVB @neural_avb]

https://t.co/81HlbVYz7Q

[CDT in Speech and Language Technologies @sltcdt]

Join us on 8 June in #Sheffield to hear about #PhD students’ #AI and #ML #speech and #NLProc research!

#Keynotes from @egrefen (@GoogleDeepMind) & @wang_wenwu (@cvssp_research).

Find out more and #register (by 31 May) at https://t.co/mx3PuJhb1Q

[Atsuki Yamaguchi @_gucciiiii]

This work is my favourite one in many ways.

If you are interested in cross-lingual transfer/adaptation, please check this out! Also, if you plan to attend ACL in July, please stop by our poster and let’s have a chat!

[Nikos Aletras @nikaletras]

When we adapt LLMs to other languages using unlabeled target language data, they tend to lose their core instruction-following and reasoning abilities.

🚨 New #ACL2026 paper w/ @_gucciiiii, T. Morishita & @AlineVillav on how to mitigate this catastrophic forgetting 👇

[CDT in Speech and Language Technologies @sltcdt]

Applications for our fully-funded 3.5 year #AI #PhD studentship on “#Metamodelling of #Speech Domains” starting in Sept closes TODAY!

Hurry! Apply now!

More info and to apply: https://t.co/PoNwTtx756