Sabitlenmiş Tweet
Recent advances such as self-consistency and test-time reinforcement learning improve the reliability of LLMs without additional supervision, but do we really know why they work?
We dig into the maths behind it
GIF
English
Paula Cordero
45 posts
Paula Cordero
@pcorderoencinar
PhD student at Imperial College and Oxford university
London, England Katılım Ekim 2023
135 Takip Edilen49 Takipçiler
@ValentinDeBort1 Braseiro around Copacabana, Tacacá do Norte (amazing north Brazilian food), Cultivar for açais and pão de queijo. For live music bars: beco das garrafas and forro da gloria
Português
Flying to Rio for ICLR woop! Hmu if you want to chat about generative models or even better if you have restaurant/bar recommendation
English
Paula Cordero retweetledi
We’re excited to unveil the name of our new London building: Platform 37. 📍
The name honors both the surrounding area’s transport heritage and "Move 37" – the critical moment where our AI system AlphaGo showed it could find novel solutions humans hadn't considered.
English
Paula Cordero retweetledi
Softmax attention giving you O(n²) headaches? Maybe try Weighted Iterative Low-rank Decomposed Coreset Attentions (or WildCat for short), an approximate attention module that runs fast in theory and practice. Super excited to share this rewarding collaboration with @LesterMackey.
Lester Mackey@LesterMackey
@tobias_schrdr and I are excited to share WildCat: Near-Linear Attention in Theory and Practice arxiv.org/abs/2602.10056 By attending over a spectrally-accurate optimally-weighted coreset, WildCat approximates exact attention with super-polynomial error decay in near-linear time
English
I'll be here for the whole week, happy to meet for tacos 🌮 and discuss about generative models, OT, uncertainty quantification, and some recent work that I've done on test-time scaling for LLMs
Paula Cordero@pcorderoencinar
At sunny SD for #NeurIPS2025, ready to present "Sampling by averaging: a multiscale approach to score estimation" tomorrow at 4.30pm poster #602. Stop by if you want to chat about diffusion-based sampling and generative models!
English
At sunny SD for #NeurIPS2025, ready to present "Sampling by averaging: a multiscale approach to score estimation" tomorrow at 4.30pm poster #602. Stop by if you want to chat about diffusion-based sampling and generative models!
English
@BerivanISIK @NeurIPSConf Would love to grab a coffee and chat about LLM evaluation and benchmarking!
English
I’ll be @NeurIPSConf all week and would love to connect on LLM data, evaluation, benchmarking, and scaling laws. If you’re working on related problems, feel free to reach out.
PS: Don’t miss our one-of-a-kind workshop on LLM evaluation: sites.google.com/view/llm-eval-…
English
Paula Cordero retweetledi
Diffusion annealed Langevin dynamics: a theoretical study ift.tt/kTJNy6A
English
English
Thrilled to share that I successfully defended my PhD last Friday at École Polytechnique! It’s been an incredible journey with my advisor @marylougab and all the amazing researchers I met along the way.
Next chapter: Research Fellow at @FlatironInst in New York, starting Jan 2026
English
Had an amazing time working with @canned_urn_wad on this project! If you're interested, please check out our paper, blog, and code!
📄 Paper: arxiv.org/abs/2510.17472…
🌍 Blog: paulaoak.github.io/certified_self…
💻 Code: github.com/paulaoak/certi…
English
In short:
We unify self-consistency, test-time training, and statistical certification in one framework for reliable reasoning in LLMs.
Self-consistency isn’t just a heuristic, it’s provably reliable!
(10/n)
English
Recent advances such as self-consistency and test-time reinforcement learning improve the reliability of LLMs without additional supervision, but do we really know why they work?
We dig into the maths behind it
GIF
English
English
I'm thrilled to announce that my #ERCStG project **Optinfinite : Efficient infinite-dimensional optimization over measures**
has been accepted. Thank you
@ERC_Research !
Many thanks also to @CrestUmr @IP_Paris_ for their support, as well as to my collaborators and friends.
English