←
Sheffield NLP944 posts
Sheffield NLP avatar

Sheffield NLP

@SheffieldNLP

Established 1993, the University of Sheffield's #NLProc Group is one of the UK's largest natural language processing research centres.

◷ Joined November 2009▧ 58 Photos and videos
Sheffield NLP avatar
Sheffield NLP@SheffieldNLP·Oct 22

With #EMNLP2026 just around the corner, a reminder of the EMNLP papers from Sheffield NLP! If you're attending, do check out our papers and say hi to the authors! Looking forward to seeing everyone at EMNLP!

Post attachment

We’ve been locked out of this account for a while😅 but we’re back, with some good news! 🎉 We have a bunch of new Sheffield NLP papers at @emnlpmeeting @aaclmeeting CIKM and INLG 📚 Congrats to everyone involved! 👏 More details below 👇

Post attachment
Sheffield NLP avatar
Sheffield NLP@SheffieldNLP·Sep 24

We’ve been locked out of this account for a while😅 but we’re back, with some good news! 🎉 We have a bunch of new Sheffield NLP papers at @emnlpmeeting @aaclmeeting CIKM and INLG 📚 Congrats to everyone involved! 👏 More details below 👇

Post attachment
Sam Lewis-Lim avatar
⟳ Sheffield NLP retweeted
Sam Lewis-Lim@samlewislim·Aug 6

Very happy to be presenting this paper at #COLM2026! 🎉 We found that monitor models struggle to detect which modality the model was relying on just from the reasoning trace. Especially with longer reasoning. A small thread on some extra findings below 🧵

1/ 🎉 Excited to share that our paper "Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models" has been accepted to #COLM2026! 📄 Final version: arxiv.org/pdf/2604.14888 A thread 🧵👇

Nikos Aletras avatar
⟳ Sheffield NLP retweeted
Nikos Aletras@nikaletras·Aug 17

Let your agents communicate via "hidden" states with StateBridge. Downstream performance 🚀 Coming to @COLM_conf 2026 👇

AI agents are getting a new communication channel – hidden states A method from @sheffielduni – StateBridge – lets them skip the text conversion and pass information in a continuous form This cuts out an entire step in the communication pipeline. Here is how StateBridge works: - It takes the final-layer hidden states corresponding to the last 64 tokens of an agent’s message - Aligns them with the receiving model’s embedding space - Calibrates their scale - And finally, inserts them as a continuous prefix. This method needs no training or model modification. So in the end, we go from "hidden states → text/tokens → embeddings → hidden states" to "hidden states → aligned states → receiver" The results show that: - StateBridge beat or tied baselines on 22/26 tests. - Qwen3-32B hit 64.1% on GPQA vs. 58.3% with text. - It also recovered details missing from the visible text. But there is one limitation. StateBridge was tested with identical model weights across agents, and it's unclear how good cross-model communication happens. Most interestingly, StateBridge arrives in the same week many researchers showed that hidden reasoning blobs can leak sensitive data across users and API sessions. Internal states may indeed preserve more information. But the more communication moves beyond text, the less of it we can actually observe, debug, and govern. So what's better to optimize for?

Post attachment
Nikos Aletras avatar
⟳ Sheffield NLP retweeted
Nikos Aletras@nikaletras·Aug 6

Shedding more light on the reasoning faithfulness of language models, now on multimodal settings! Paper w/ @danaesavi @samlewislim @delliott to be presented at COLM 👇

1/ 🎉 Excited to share that our paper "Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models" has been accepted to #COLM2026! 📄 Final version: arxiv.org/pdf/2604.14888 A thread 🧵👇

Nikos Aletras avatar
⟳ Sheffield NLP retweeted
Nikos Aletras@nikaletras·Jun 29

Here is MultiHashFormer, our new framework for vocabulary efficient language modelling. Work w/ @HuiyinXue and @_gucciiiii Inspired by chaotic dynamic memory systems with distributed state spaces, we replace the LM embedding matrix with a modular hashing interface 1/N

Post attachment
Sheffield NLP avatar
Sheffield NLP@SheffieldNLP·Jun 22

Brilliant seminar today from @amuuueller on using interpretability as an engineering tool for NLP. Can't wait for your paper to hit arXiv soon about using ambiguity representations to improve confidence estimation!

Post attachment
Nikos Aletras avatar
⟳ Sheffield NLP retweeted
Nikos Aletras@nikaletras·Jun 3

New #ICML2026 paper: "On Effectiveness and Efficiency of Agentic Tool-calling and RL Training". ⚠️The methodology for measuring and improving tool-calling capabilities is fragile. We break down our analysis along two critical axes: (1) effectiveness; and (2) efficiency. 👇

Post attachment
Constantinos Karouzos avatar
⟳ Sheffield NLP retweeted
Constantinos Karouzos@ckarouzos·Apr 23

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! 👇

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.

Post attachment