Hunting for Eureka moments Research engineer with @mint_nlp_mbzuai & @riken_aip • formerly @nyuniversity Intelligence really is super
Joined May 2024
- Tweets605
- Following92
- Followers138
- Likes3.1K
(Sam) Munachiso Samuel Nwadike retweeted
Friend of mine saw my reading list of architectures papers, and said I should post it in case anyone finds it handy
app.notion.com/p/Architectur…
A key highlight is the really recent progression from parity -> modular arithmetic -> non-commutative state tracking for SSMs
For example, in a paper uploaded to Arxiv just ~2 weeks ago, a Japanese team showed how Mamba-3 can be made to perform (length-generalising) non-commutative state tracking (S5) with some simple adjustments to the A-matrix (NPLR)
Shoutout @zmkzmkz (said friend; he invented multi-layer key value sharing, which was a super nice contribution)
Couple great papers in there also from @hhsun1 (looped transformers), @lambdaviking (expressivity analyses) as well
Key words for the algorithm: neural network architectures
(Sam) Munachiso Samuel Nwadike retweeted
Wrote little blogpost of what I've been working on in kernels and found small but interesting thing about it on MI210s.
erland366.github.io/blog/fli…
(Sam) Munachiso Samuel Nwadike retweeted
good reading list to get a jumpstart in the many lines of work in architectures!
Friend of mine saw my reading list of architectures papers, and said I should post it in case anyone finds it handy
app.notion.com/p/Architectur…
A key highlight is the really recent progression from parity -> modular arithmetic -> non-commutative state tracking for SSMs
For example, in a paper uploaded to Arxiv just ~2 weeks ago, a Japanese team showed how Mamba-3 can be made to perform (length-generalising) non-commutative state tracking (S5) with some simple adjustments to the A-matrix (NPLR)
Shoutout @zmkzmkz (said friend; he invented multi-layer key value sharing, which was a super nice contribution)
Couple great papers in there also from @hhsun1 (looped transformers), @lambdaviking (expressivity analyses) as well
Key words for the algorithm: neural network architectures
(Sam) Munachiso Samuel Nwadike retweeted
Finally out 🎉
Introducing HyperBrowseComp: A multilingual, multimodal stress-test benchmark for search agents!
Paper: arxiv.org/abs/2610.03574
Web: hyperbrowsecomp.github.io/
Currently covers 13 languages with handcrafted, 👿diabolical questions.
(See this image as an example)
(Sam) Munachiso Samuel Nwadike retweeted
so are we/the agents truly just bottlenecked by having to do physical experiments for progress in other fields of science? how significant are the theoretical physics and logics results in this batch compared to the math ones? I’m just trying to gauge how far they’ve really come
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
(Sam) Munachiso Samuel Nwadike retweeted
Skeptical or bullish about looped transformers? Want a rigorous, controlled setup to understand their real potential?
Our key finding — looping enables systematic generalization and depth extrapolation: It lets models combine knowledge not seen in any composition, and reason over more hops than they ever see during training.
Come chat with us (@yuekun_yao @hkohli14) at #COLM2026 and join our poster session on Wednesday, 11AM–1PM!
@osunlp has many students attending the conference, including @hanane_mnsr @YifeiLiPKU @ziruch__ @Zhehao_Zhang123 @yuting_ning grab them to have a chat!
Also—@OhioStateCSE is hiring! Our department has multiple faculty openings across AI foundations, responsible AI, and agentic AI. Happy to answer questions and make connections!
(Sam) Munachiso Samuel Nwadike retweeted
we had such a blast making this benchmark, the questions I made in german are among the hardest, I’m quite proud of them lol
if you think deep browsing tip-of-my-tongue-type questions are a good measure of agent capability, you should test your favorite LLMs/harnesses on this!
Finally out 🎉
Introducing HyperBrowseComp: A multilingual, multimodal stress-test benchmark for search agents!
Paper: arxiv.org/abs/2610.03574
Web: hyperbrowsecomp.github.io/
Currently covers 13 languages with handcrafted, 👿diabolical questions.
(See this image as an example)
(Sam) Munachiso Samuel Nwadike retweeted
our lab is doing this super cool program where we try to help people out with their graduate applications! if you need some advice on your proposals, CVs, emails, etc. especially for those who are under-represented, do consider signing up for this
🎓 Graduate Program Application Mentorship
Applying to grad school doesn't have to be overwhelming!
Get personalized 1:1 support from students who’ve successfully navigated the process.
🔗 Apply now: forms.gle/FyjEuLmjcdREF3Ba6
🗓️ Deadline: Nov 1, 2026
👇 Here is what you get:
(Sam) Munachiso Samuel Nwadike retweeted
This wonderful survey is finally out!
I'm honored to have played a small part in it 🙇
Thanks a lot for taking the lead, @marco_computers !!
(Sam) Munachiso Samuel Nwadike retweeted
Architecture research is important for the advancement of intelligence
(Sam) Munachiso Samuel Nwadike retweeted
I feel like this type of 'research' not only rewards them, but also punishes the careful ones.
The fast slop gets traction and citations, but the careful work gets dismissed with, "Oh, it has already been done before".
(Sam) Munachiso Samuel Nwadike retweeted
Now what? Do we just employ JEV for accept/reject decisions?
TIMELINE:
Sept 15th JEV came out
Sept 18th papers on JEV submit abstracts
Sept 22nd paper on JEV is already on arxiv
What do you call this, because it's not research....
(Sam) Munachiso Samuel Nwadike retweeted
this is will smith all over again.
see you guys in 6 months this time lol.
(Sam) Munachiso Samuel Nwadike retweeted
nice thread!
(Sam) Munachiso Samuel Nwadike retweeted
My very first co-first-authored paper was accepted to EMNLP 2026! 🥳
TL;DR: We observed a serial position effect in Mamba and took a peek at how it manifests internally in the model.
See you in Budapest! (Hopefully 😅)🇭🇺
(Sam) Munachiso Samuel Nwadike retweeted
Our work was published by Nature Astronomy @NatureAstronomy! See details here: nature.com/articles/s41550-0… Congrats to the lead students Lucas Carrit Delgado Pinheiro and @ziruch__ @osunlp and the entire team,
Bruno Caixeta Piazza, @ShroffNess, Yingbin Liang, and @TingAstro!
(Sam) Munachiso Samuel Nwadike retweeted
Congratulations to Zhilin Yang, founder and CEO of @Kimi_Moonshot, on the latest Kimi release. What a huge win for the open-source community!
It feels like just yesterday Zhilin was graduating from my lab at CMU, jointly co-advised with William Cohen. Not only did he complete his Ph.D. in just four years, but he also made truly fundamental contributions to ML during his time at CMU.
What a spectacular career! Congrats again Zhilin, and thank you and the entire Kimi team for everything you're doing for the open-source community.
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.com/blog/kimi-k3
