Mathilde Papillon🦋 mathildepapillon .bsky .social

418 posts

Mathilde Papillon🦋 mathildepapillon .bsky .social banner
Mathilde Papillon🦋 mathildepapillon .bsky .social

Mathilde Papillon🦋 mathildepapillon .bsky .social

@mathildepapillo

Non-Euclidean + Interpretable AI PhD-ing @Geometric_Intel @UCSBPhysics Fellow @GoodfireAI Prev. @Dodgers⚾️ @McGillUPhysics BlueSky: @mathildepapillon

Katılım Temmuz 2022
347 Takip Edilen2.5K Takipçiler
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Goodfire
Goodfire@GoodfireAI·
> replicate J-space on GLM 5.2 > train a reward model and run RL to reduce hallucinations > show me how this model makes cancer predictions Using our platform Silico is like having a team of AI researchers ready to run experiments like these. Private beta is open now. 🧵 (1/6)
English
8
71
703
99.2K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Melanie Weber
Melanie Weber@mweber_PU·
How does neural feature geometry evolve during training? Modeling feature spaces as geometric graphs, we show that nonlinear activations drive transformations resembling discrete Ricci flow - revealing how class structure emerges and suggesting geometry-informed training principles. Led by Moritz Hehl. Details here: arxiv.org/abs/2509.22362
Melanie Weber tweet media
English
0
47
319
26.5K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Matthew Kowal
Matthew Kowal@MatthewKowal9·
You might have noticed that the world around us has a ton of multidimensional geometric structure; unsurprisingly, so do concepts in large vision models! Check out how to find them🔬 Super fun project to be a part of and very excited to apply BSFs to other domains! 🤖📚🩺🎵
Goodfire@GoodfireAI

If models think in shapes, our tools should too. Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations - using multidimensional “blocks” instead of single directions. (1/9)

English
1
6
37
3K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Tom McGrath
Tom McGrath@banburismus_·
one example from the blog that isn't in the thread: we looked into a robotics model with the BSF and found that the robot arm is actually represented by a little arm in the same position in activation space! this is the first evidence for a topographic representation (en.wikipedia.org/wiki/Topograph…) that I've seen in a model
Tom McGrath tweet media
Goodfire@GoodfireAI

The broader point: a featurizer is a hypothesis about how a model’s representations are structured. We think this is a step forward in tools reflecting that. Paper, code, and full post: goodfire.ai/research/bsf-v…

English
7
20
130
16.4K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Tom McGrath
Tom McGrath@banburismus_·
we've found a way to discover neural geometry, completely unsupervised, and the results are incredible! this is a really major step forward in interp in my opinion. there's so much structure in there
Goodfire@GoodfireAI

If models think in shapes, our tools should too. Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations - using multidimensional “blocks” instead of single directions. (1/9)

English
1
19
181
14.7K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Thomas Fel
Thomas Fel@thomas_fel_·
Our work on Block-Sparse Featurizer is out 🧊 :) We revive an old idea from the structured sparsity literature and use it to carve activation space into meaningful regions. It's a first concrete answer to the question our concept manifolds work left open ! :)
Thomas Fel tweet media
Goodfire@GoodfireAI

If models think in shapes, our tools should too. Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations - using multidimensional “blocks” instead of single directions. (1/9)

English
4
46
250
23.7K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Goodfire
Goodfire@GoodfireAI·
If models think in shapes, our tools should too. Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations - using multidimensional “blocks” instead of single directions. (1/9)
Goodfire@GoodfireAI

Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵

English
25
159
967
255.1K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Fatih Dinc
Fatih Dinc@fatihdin4en·
Thank you SIAM for this life changing award for my PhD work! You can read more about me and my work in SIAM's Spotlight here: #Dinc" target="_blank" rel="nofollow noopener">siam.org/publications/s…
SIAM@TheSIAMNews

Join us in congratulating the 17 SIAM prize recipients who will be honored at #SIAMAN26, July 6–10 in Cleveland! 👏 🔗 Read more about each recipient: siam.org/publications/s… 📅 View the conference program: meetings.siam.org/program.cfm?CO… #SIAMED26 #SIAMLS26 #SIAMMPE26

English
2
7
23
2.8K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Cohere
Cohere@cohere·
Happy Canada Day!! 🇨🇦 We're so proud to be founded and growing in Canada. Let's keep building💡
English
16
18
360
13.4K
Mathilde Papillon🦋 mathildepapillon .bsky .social
Grateful to be spending my summer @GoodfireAI ! 🌁 Interpretability is the next frontier, I think, deciding how much and in what capacity we will trust AI. I can’t wait to learn from the brilliant folks working here. DM me if you’re in SF, would love to chat☕️
Mathilde Papillon🦋 mathildepapillon .bsky .social tweet media
English
2
6
79
3.7K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Goodfire
Goodfire@GoodfireAI·
Stories have shapes: a comedy rises toward joy; a tragedy falls into loss. Inside an LLM, that’s visible more literally: as an LLM reads a story, its internal activations trace a wandering path that reflects the model’s sense of what kind of story it is reading. (1/5)
Goodfire@GoodfireAI

Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵

English
21
122
866
103.6K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
UCSB ECE
UCSB ECE@ucsbece·
🎉 Congratulations to @fatihdin4en, postdoc in UC Santa Barbara #ECE and the Kavli Institute for Theoretical Physics, on winning the Richard C. DiPrima Prize from the Society for Industrial and Applied Mathematics (SIAM). 🏆 🔗 Read the full story: lnkd.in/gd27SUAA
UCSB ECE tweet media
English
2
7
35
8.4K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Nina Miolane 🦋 @ninamiolane.bsky.social
🧠Can geometry give us a theory of neural computation? @claire_i_webb and I discussed how #geometry, the hidden shapes & structure inside high-dimensional neural data, can be a powerful tool to understand how the brain computes. @longnow @berggruenInst @geometric_intel
Long Now Foundation@longnow

How are scientists mapping consciousness? Could geometry hold the key? Claire Isabel Webb, director of the Future Humans program at the @berggruenInst, asks @NinaMiolane of the UCSB Geometric Intelligence Lab. Watch the full episode here: na2.hubs.ly/H068tRX0

English
9
18
54
6.6K
Mathilde Papillon🦋 mathildepapillon .bsky .social
We are 58 days away from the 2026 TDL Challenge deadline! And the live leaderboard is looking spicy🌶️ geometric-intelligence.github.io/topobench/lead… Submit your implementation of a graph or topological neural network today, and run the chance of winning up to 1000$ USD! Many other cool prizes too💰✈️
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo

🏆 The 2026 Topological Deep Learning Challenge is officially live, now in its 4th edition! 🏆 This year’s theme is “Bridging the Gap” between the GNN and TDL worlds. Win incredible prizes including up to $1000 in cash 💸 and AI research internships! Submission deadline: Aug. 1

English
2
5
17
2.7K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Goodfire
Goodfire@GoodfireAI·
Have you debugged your training data? You might not like what you find. Introducing predictive data debugging: reveal and shape what your model will learn before training. In DPO datasets, we found broken guardrails, hallucinations, and fish fart fan fiction (seriously). (1/9)
English
26
109
906
182.1K
Mathilde Papillon🦋 mathildepapillon .bsky .social retweetledi
Erica Busch
Erica Busch@ericalbusch·
Our new paper is out this week in @NatureNeuro! We built a BCI that works with the brain's natural geometry — and we found that people could learn to play a video game with their brains in <1 hr of training. This efficiency is groundbreaking & here's why: nature.com/articles/s4159…
English
2
51
233
18.8K