Pinned Tweet
Join our PhD programme in Designing Responsible Natural Language Processing at the UKRI AI Centre for Doctoral Training, University of Edinburgh. Applications are now re-opened for Home fee status candidates (past candidates need not re-apply). responsiblenlp.org
0
6
15
6,342
EdinburghNLP retweeted
New paper❗
Does understanding how AI models work actually make them safer?
This assumption often underlines interpretability research, but different subfields operationalise this intuition in very different ways. Our survey of mechanistic interpretability (MI) for safety guide to the field of MI, detailing conceptual approaches and techniques under this broad banner and how different types of work relates to the mission of making AI model safer. 🧵
Does understanding how AI models work actually make them safer?
This assumption often underlines interpretability research, but different subfields operationalise this intuition in very different ways. Our survey of mechanistic interpretability (MI) for safety guide to the field of MI, detailing conceptual approaches and techniques under this broad banner and how different types of work relates to the mission of making AI model safer. 🧵
9
63
315
26,638
EdinburghNLP retweeted
Massive congrats to the whole team! 🎉
We also explored distilling MCTS traces into LLMs to improve reasoning with @PMinervini. Exciting to see these ideas connect across domains!
x.com/Lorenzo_molf/s…
We also explored distilling MCTS traces into LLMs to improve reasoning with @PMinervini. Exciting to see these ideas connect across domains!
x.com/Lorenzo_molf/s…
Can self-play on board games teach an LLM to reason? ♟️🧠
Self-Play Search Distillation (SPSD) turns the search of MuZero experts, trained by self-play on board games, into reasoning traces for LLMs.
The reasoning transfers to unseen games and to math (24.1 → 36.6) 🧵
0
4
7
556
EdinburghNLP retweeted
Back from @emnlpmeeting in Budapest a couple of days ago! Nice to catch up with old friends and meet some new people. Also great to have quite a few people from our group there, we even managed to get some of us together for a photo! @EdinburghNLP @AmsterdamNLP
0
1
33
1,631
EdinburghNLP retweeted
Join us for sharing the cool demos on your hands about world models!
⏰ Deadline: Nov 13, 2026, AoE
🌐 Website: robowm-ws.github.io
📝Submission google form: forms.gle/BypA9ZdLTCss8q…
Organizers: Homanga Bharadhwaj @mangahomanga , Joycelyn @JingJoycelyn , Hiroki Furuta @frt03_ , @kuanghueilee , @ruoshi_liu , @Liu_Zeyi_ , @yifuqiu98 , Wenhao Yu
0
3
12
1,239
EdinburghNLP retweeted
my amazing collaborators will be presenting a few papers at EMNLP/NeurIPS/LoG -- if you're attending any of these conferences, touch base with them; they're great! 🚀🚀🚀
blog post: neuralnoise.com/2026/emnlp-neu…
blog post: neuralnoise.com/2026/emnlp-neu…
0
4
42
1,521
EdinburghNLP retweeted
VLMs benefit from tool calls as they allow for "thinking with images".
But can a VLM simply internalise the effects of calling correct tools in the latent space and not need to explicitly call them during inference? 🧵
Presenting tomorrow at 11 AM @COLM_conf !
But can a VLM simply internalise the effects of calling correct tools in the latent space and not need to explicitly call them during inference? 🧵
Presenting tomorrow at 11 AM @COLM_conf !
6
3
42
128,123
EdinburghNLP retweeted
Did you know training on your own incorrect samples can beat distilling from a much larger teacher?
Traditional RFT only reinforces a narrow set of approaches. Strategically sampling for approach-level diversity gives better pass@k, test-time scaling and RL initialization!
alexgurung.me/diversity/
🧵
Traditional RFT only reinforces a narrow set of approaches. Strategically sampling for approach-level diversity gives better pass@k, test-time scaling and RL initialization!
alexgurung.me/diversity/
🧵
5
12
146
12,404
EdinburghNLP retweeted
Very happy to see @j0hngou’s work accepted as a NeurIPS oral! I particularly like the broader idea: let models interact richly during training while pressuring them to reduce that interaction. This gives a denser learning signal and leads to more effective one-shot collaboration
Happy to share that our paper "Clarification as Supervision" was accepted as an Oral at #NeurIPS2026 (top ~0.3%). When a text-only reasoner answers from a VLM's caption and gets it wrong, we usually blame the reasoning. Often the caption left out the details the reasoner needed🧵
0
2
16
1,510
EdinburghNLP retweeted
Self-play on a massive number of synthetic environments as an endless source of super-human post-training data 🚀🚀🚀 Really happy this is now out — it may be how we achieve AGI/ASI 🙂
📄 arxiv.org/abs/2609.30936
🌐 lormolf.github.io/publications/s…
Work done during my visiting period at the University of Edinburgh with @PMinervini , together with the UniboNLP group
Kudos to all my coauthors!
1
4
48
5,662
EdinburghNLP retweeted
Latent reasoning (i.e. recursively decomposing goals into sub-goals in latent space, and extracting symbolic knowledge by decoding goal and sub-goal representations) waaay before it was cool, straight outta ICML’20
Very cool work from UCL+FAIR-London on neural architectures for reasoning that learn to generate and select rules.
0
2
44
2,072
EdinburghNLP retweeted
Back in 2021 we managed to get an ICLR Outstanding Paper Award (iclr-conf.medium.com/announcing-icl…) with @_kire_kara_’s amazing MSc thesis work, today this looks a bit more tricky to achieve 😬
ICLR 2027 has received more submissions than all previous years (2013–2026) combined
2
2
36
4,740
EdinburghNLP retweeted
Edoardo is amazing, catch this opportunity! 🚀🚀🚀🚀🚀
I am looking for 1 postdoc and 2 fully funded PhDs at Imperial @ICComputing to join my @ERC_Research project AToM ⚛
We aim to build models that learn to compress what they perceive and remember, for permanent memory, longer horizons, and >10x speedups
ducdauge.github.io/atom/
0
2
9
1,415
EdinburghNLP retweeted
I am looking for 1 postdoc and 2 fully funded PhDs at Imperial @ICComputing to join my @ERC_Research project AToM ⚛
We aim to build models that learn to compress what they perceive and remember, for permanent memory, longer horizons, and >10x speedups
ducdauge.github.io/atom/
We aim to build models that learn to compress what they perceive and remember, for permanent memory, longer horizons, and >10x speedups
ducdauge.github.io/atom/
5
54
203
14,922
EdinburghNLP retweeted
Fazl and his group are stellar; consider applying!!! 🚀🚀🚀
Hiring!
Our Technical Safety and Governance (TSG) Lab at Oxford is hiring a Senior Researcher to work on interpretability, evaluations and AI safety for continually learning systems.
Please apply by Sep 22 and share with anyone who might be interested.
DM/email with questions!
0
1
6
965
EdinburghNLP retweeted
Whether a model is faithful (and, e.g., whether it's relying on a cue) depends heavily on environmental factors, and faithfulness/mechinterp/safety research should take that into account! 🚀 Amazing work by @aryopg et al.
CoT monitoring assumes reasoning traces record what shapes an answer.
Yet, we found that models are less likely to verbalize a cue in its CoT when the cue comes from a tool return instead of the user message! 👀
🧵
0
3
16
1,380
EdinburghNLP retweeted
Developing next-generation accelerators requires benchmarking them on realistic and agentic AI workloads.
With @ARIA_research's Scaling Compute programme, we are releasing TEAS 🍵, serving 5 models (4B to ~1T total) on 9 accelerators for 6 workloads.
teasbench.com
With @ARIA_research's Scaling Compute programme, we are releasing TEAS 🍵, serving 5 models (4B to ~1T total) on 9 accelerators for 6 workloads.
teasbench.com
1
7
51
4,188
EdinburghNLP retweeted
CoT monitoring assumes reasoning traces record what shapes an answer.
Yet, we found that models are less likely to verbalize a cue in its CoT when the cue comes from a tool return instead of the user message! 👀
🧵
Yet, we found that models are less likely to verbalize a cue in its CoT when the cue comes from a tool return instead of the user message! 👀
🧵
2
17
113
270,109
EdinburghNLP retweeted
Scotland is pretty amazing AI-wise! For example, @EdinburghUni folks have been doing AI research since the 60s (informatics.ed.ac.uk/aiai/history-o…), and they still have a pretty massive impact (impactrank.org, informatics.ed.ac.uk/news/latest-ne…, etc.).
Didn’t know Scotland had so many AI labs
1
6
18
2,587
EdinburghNLP retweeted
We are pleased to announce the winner of the ACL Computational Linguistics Doctoral Dissertation Award (ACL Dissertation Award)! Congratulations Verna Dankers!
@vernadankers #NLproc #CLJournal
@vernadankers #NLproc #CLJournal
0
4
53
3,590