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 👇

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 🧵👇

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?

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 🧵👇

I'm really excited to be delivering a seminar at @SheffieldNLP on Tuesday about my research on Small and Human-Scale Language Modeling! Thanks @_joestacey_, @_gucciiiii, Marco Valentino and @nikaletras for the invite!

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.

#SIGIR2026 accepted paper list has been released on sigir2026.org/en-AU/pages/pr… Vibe coded the analysis again based on csrankings.org (again, by no means accurate), found that this year, Renmin University continues to be the top-1 institute, and China and US are still the top-2 countries with most publications.