Postdoc at MIT CSAIL. Working on combinatorial optimization with neural nets https://teapawt.net/t.co/bf4UWg2BUQ
Cambridge, Massachusetts
Joined March 2012
- Tweets388
- Following1K
- Followers414
- Likes11K
Nikos Karalias retweeted
i think it’s only fair that Grigori Perelman should IPO for a trillion dollars
Nikos Karalias retweeted
Imagine giving Claude Opus 5 ten questions it reliably answers individually. Put them in one prompt, and it starts getting some wrong.
Excited to share our new work that investigates why. Our findings point to “implicit budgeting” in LLMs: as more questions share a prompt, reasoning per question shrinks, following an approximate power law (even with plenty of context space left). Thus, questions that need longer reasoning are hit hardest.
Understanding this also helps us batch questions more efficiently. For Opus 5 on GPQA-Diamond, that saves ~24% of tokens with roughly unchanged accuracy.
Huge credit to @chenpan_li for the thoughtful experiments and hard work behind this paper!
Nikos Karalias retweeted
📄 Paper: arxiv.org/abs/2610.27190
🌐 Project: chenpanli.github.io/fragilit…
💻 Code: github.com/chenpanli/fragili…
🤗 discussion: huggingface.co/papers/2610.2…
Nikos Karalias retweeted
LLMs can solve tasks one at a time, yet fail when the same tasks are combined into one prompt. Why?
In our paper, we find that LLMs operate under an implicit thinking budget, with the reasoning effort applied to each task shrinking according to a power law.
Nikos Karalias retweeted
Anyway, as always my advice is to log off. I will never take this advice myself.
Nikos Karalias retweeted
📢 New position paper: “𝗡𝗲𝘂𝗿𝗮𝗹 𝗔𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝗶𝗰 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗥𝗲𝗾𝘂𝗶𝗿𝗲𝘀 𝗮 𝗖𝗹𝗲𝗮𝗿 𝗦𝗰𝗼𝗽𝗲, 𝗧𝗵𝗲𝗼𝗿𝗲𝘁𝗶𝗰𝗮𝗹 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀, 𝗮𝗻𝗱 𝗘𝗺𝗽𝗶𝗿𝗶𝗰𝗮𝗹 𝗥𝗶𝗴𝗼𝗿”
When and where should we use Neural Algorithmic Reasoning (NAR)? NAR is an exciting emerging area at the intersection of 𝙘𝙡𝙖𝙨𝙨𝙞𝙘𝙖𝙡 𝙖𝙡𝙜𝙤𝙧𝙞𝙩𝙝𝙢 𝙙𝙚𝙨𝙞𝙜𝙣 and 𝙣𝙚𝙪𝙧𝙖𝙡 𝙘𝙤𝙢𝙥𝙪𝙩𝙖𝙩𝙞𝙤𝙣 — but its foundations are still underdeveloped.
In this paper, we argue that the field needs:
✨ a sharper 𝙙𝙚𝙛𝙞𝙣𝙞𝙩𝙞𝙤𝙣 of what NAR is (and is not), including clearer distinctions from neighboring paradigms,
📐 stronger 𝙩𝙝𝙚𝙤𝙧𝙮 for expressivity and generalization,
🧪 more rigorous 𝙗𝙚𝙣𝙘𝙝𝙢𝙖𝙧𝙠𝙨 and connection to high-impact 𝙖𝙥𝙥𝙡𝙞𝙘𝙖𝙩𝙞𝙤𝙣𝙨.
We have extended the deadline :)
Can LLMs reason about TSP, SAT, or routing?
Join DiffCoALG @NeurIPSConf 2025 — where discrete algorithms meet differentiable ML.
📅 Deadline: *extended* to Aug 31 (AOE)!
🔗 sites.google.com/view/diffco…
2
468
Nikos Karalias retweeted
Join DiffCoALG @NeurIPSConf 2025 — a workshop on neural methods + classic algorithms (think: TSP, SAT, MIP, and more).
🧠 Theory meets practice.
The deadline is only 3 days away!
📆 Submit by Aug 22
🔗 sites.google.com/view/diffco…
Nikos Karalias retweeted
🔥 New workshop at @NeurIPSConf!
DiffCoALG bridges the gap between classic algorithms & differentiable learning.
Think: LLM reasoning, routing, SAT, MIP — neurally optimized end-to-end.
📌 Submit by Aug 22!
🔗 sites.google.com/view/diffco…
#NeurIPS2025
Nikos Karalias retweeted
It’s not advisable to vote for someone who tells you they will do something in the hope that they are lying to you and won’t do it.
Nikos Karalias retweeted
Mr. President, Ukraine did not “start” this war. Russia launched an unprovoked and brutal invasion claiming hundreds of thousands of lives. The Road to Peace must be built on the Truth.🇺🇸🇺🇦
“Russia Invades Ukraine in Largest European Attack Since WWII” @FoxNews (February 24, 2022) foxnews.com/world/russian-in…
Accepted at NeurIPS 2024 as a spotlight :)
We'll post an updated version on Arxiv next week!
I'll be presenting our work on optimal approximation algorithms for combinatorial problems with graph neural nets at the Mathematics of Modern Machine Learning (M3L) workshop. Saturday 4pm, room 242!
Arxiv: arxiv.org/abs/2310.00526
I'll be presenting our work on optimal approximation algorithms for combinatorial problems with graph neural nets at the Mathematics of Modern Machine Learning (M3L) workshop. Saturday 4pm, room 242!
Arxiv: arxiv.org/abs/2310.00526
Nikos Karalias retweeted
New reading group recording on YouTube: youtu.be/02I7EtgIQec
@AspectStalence presents his paper "Neural Set Function Extensions: Learning with Discrete Functions in High Dimensions"!
You can join the reading group yourself here: m2d2.io/talks/log2/about/
I'll be presenting our work on continuous extensions to the graph reading group tomorrow!
Alright 2023 let's go - tomorrow we have our first paper presentation with "Neural Set Function Extensions: Learning with Discrete Functions in High Dimensions" arxiv.org/abs/2208.04055!
Join us on zoom at 11am EST / 5pm CET on Monday via this link: m2d2.io/talks/log2/about/
This Wednesday 2PM (PST) we will present our poster on continuous extensions of discrete functions (hall J #305 - poster session 4). I'll be here at #NeurIPS2022 all week so if you want to chat about comb opt/ML dm me :)
1
Our paper on neural extensions of set functions arxiv.org/abs/2208.04055 has been accepted at #NeurIPS2022.
Was a pleasure to work with @Josh_d_robinson, @loukasa_tweet, and @StefanieJegelka on this one :)
Thread/repo/camera-ready coming soon.
Nikos Karalias retweeted
One quite famous (and moving) example (CW: suicide):
ALT The first author must state that his coauthor and close friend, Tom Trobaugh, quite intelligent, singularly original, and inordinately gener- ous, killed himself consequent to endogenous depression. Ninety-four days later, in my dream, Tom's simulacrum remarked, "The direct limit char- acterization of perfect complexes shows that they extend, just as one extends a coherent sheaf." Awaking with a start, I knew this idea had to be wrong, since some perfect complexes have a non-vanishing Ko ob. struction to extension. I had worked on this problem for 3 years, and saw this approach to be hopeless. But Tom's simulacrum had been so insistent, I knew he wouldn't let me sleep undisturbed until I had worked out the argument and could point to the gap. This work quickly led to the key results of this paper. To Tom, I could have explained why he must be listed as a coauthor. During his lifetime, Tom also pointed out the interesting comparison of the careers of Grothendieck and Newton.
