Mehran Hosseini

Assistant Professor @OfficialUoM 🐝 • Safe, Verifiable, and Efficient AI 🤖 • Mathematics of AI 💻 • Formal Verification • 📖 PhD @UniofOxford

[Mehran Hosseini @MehransTweet]

Start Date: July 2026

🎓 Supervisors: Dr Kieran Wood (Aerospace) and myself (Computer Science | @csmcr )

Sharing and repost is much appreciated!

[Mehran Hosseini @MehransTweet]

📢 Fully Funded International PhD Scholarship in Safe Robotics & AI at @OfficialUoM

This PhD project develops safe perception, control, and planning for assured autonomy in robotics through redundancy: combining diverse AI & formal approaches.

🔗 Apply: https://t.co/AlDz8gXfit

[The Green Party @TheGreenParty]

Today we launch our campaign for the May elections.

Locally and nationally, Labour is cosying up to developers and failing to provide the secure, affordable homes we need.

We deserve better.

On 7th May vote for council homes, rent controls and an end to leaseholds. Vote Green.

[Alex Shams @alexshams_]

Kharg Island isn't just a military target

It is home to a medieval Portuguese fort, religious tombs, and the ruins of one of the oldest Christian monasteries in the world

It also is home to 8,000 people and flocks of wild gazelles

This is what Trump just bombed:

[Physics In History @PhysInHistory]

The word "algorithm" comes from the name of 9th-century Persian mathematician al-Khwarizmi, whose Latinized name "algorismi" was used to refer to his systematic methods of calculation.

[Mehran Hosseini @MehransTweet]

📢 2-year postdoc in Technical AI Safety is open at KCL. You'll be advancing the reliability & robustness of latent probes for #LLMs, collaborating with Nicola Paoletti, Osvaldo Simeone, @EdoardoPona (@KingsCollegeLon) and myself (@OfficialUoM)!

https://t.co/8MNNYbs2Vl

[Mehran Hosseini @MehransTweet]

@davidad I agree!! Formal safeguards can very much ensure safety.

At the same, I think investigating AI intention isn't totally hopeless. E.g., through neurosymbolic approaches, internal representation analysis, mechanistic interpretability,...

[Mehran Hosseini @MehransTweet]

@janleike I'm pretty sure this counts as jailbreak.

[Mehran Hosseini @MehransTweet]

📢We have 2/2 papers accepted in @AAMASconf 2025 🥳

Both papers on the Safety of AI agents; specifically, on
• "LTL Verification of Memoryful Neural Agents" and
• "Certified Guidance for Planning with Deep Generative Models".

More details soon!
#AAMAS2025 #AAMAS #SafeAI

[Mehran Hosseini @MehransTweet]

📋If you're attending @NeurIPSConf , please visit our poster on Verifiably Robust Conformal Prediction. @linusjeary and @TDKuipers will be presenting our work!

#NeurIPS #NeurIPS2024

[Mehran Hosseini @MehransTweet]

@iclr_conf reviewer:

I don't understand 60% of your paper. But I'm going to reject it, and I'm 100% certain about my assessment.

#ICLR2025 #PeerReview

[Mehran Hosseini @MehransTweet]

✨Just in! Our paper has been accepted in the IEEE/CVF Winter Conference on Applications of Computer Vision (@wacv_official) 2025!

Keep an eye out for more details coming soon!

[Peyman Hosseini @Peyman_Hs]

📢 I am excited to share that our paper (with @MehransTweet) "GeoPos: A Minimal Positional Encoding for Enhanced Fine-Grained Details in Image Synthesis Using Convolutional Neural Networks" has been accepted to be presented at @wacv_official 2025 at Arizona🌵!

Keep an eye out for the PDF and code, which will be released soon.

[Mehran Hosseini @MehransTweet]

@davidad That's a strong come back from Christopher Strachey :))))

[Mehran Hosseini @MehransTweet]

I'm very excited to share that our paper "Verifiably Robust Conformal Prediction" been accepted in NeurIPS 2024!!!

Here is the link to the arXiv version:
https://t.co/RvCJFA9woe

#NeurIPS #NeurIPS24

[Mehran Hosseini @MehransTweet]

📢 Checkout our new paper on Verifiably Robust Conformal Prediction (VRCP)!!!

🔑 Key contributions:
• maintains coverage for any ℓ_p norm-bounded attack
• formal guarantees on the prediction set & coverage
• supports regression & classification
• informative sets vs SotA

[Geoffrey Irving @geoffreyirving]

Yoshua Bengio is looking for theory folk to join him to work on Bayesian approaches to AGI safety. I think this is a great opportunity: I've quite enjoyed the theory discussions I've had with Yoshua so far, and would love more work in this direction.

[Mehran Hosseini @MehransTweet]

@Grad62304977 @Vultr @wrauvultr Keep me posted!

[Mehran Hosseini @MehransTweet]

Checkout our proof-of-concept small LM

Not the most capable LM in town, but pretty good considering its size and # of training tokens!

Parts of the training compute was kindly provided by @Vultr & @wrauvultr. Reach out to them to try their amazing GPUs!

https://t.co/wDptZEW0CA

[Peyman Hosseini @Peyman_Hs]

We trained a small language model using Efficient Attention we introduced recently, and the result is interesting! You can check it out on @huggingface. It's been only trained on 15B tokens and achieves similar loss/training token to models of similar size
https://t.co/Lbujk3bxRu

[Mehran Hosseini @MehransTweet]

@Grad62304977 @Vultr @wrauvultr Slicing :)

[Mehran Hosseini @MehransTweet]

@Grad62304977 @Vultr @wrauvultr Here is a visualisation of GQA from its paper. Not much changes in GQA except using a group of queries for each key-value pair. For this, I don't see why there is a need to bring back Wk and Wv

[Mehran Hosseini @MehransTweet]

@Grad62304977 @Vultr @wrauvultr Good question! If you check the 1st version on arXiv, we were advocating for 1 head, which is indeed related to M/GQA. But in the 2nd version, we shifted to multi-head.

Back to your question, Efficient Attention is orthogonal to M/GQA and can benefit from M/GQA.