Hannah Lawrence

PhD @ MIT CSAIL. Finding/exploiting patterns in data, especially symmetries. AI4Science. Cofounder @bostonsymmetry.

[Hannah Lawrence @HLawrenceCS]

@bostonsymmetry is hosting a poster session (4-5:30pm) + social on Tuesday, June 9th, at Northeastern! All are welcome to come chat about geometry, symmetries, + AI!

Location: Raytheon Amphitheater, https://t.co/uYjUlqwxzC

Register here: https://t.co/pR7FyySz8I

[Julian Suk @sukjulian]

Results for the GRaM Competition are online. Congratulations to the winner and runners-up 🎉

🥇 https://t.co/Q76n3UBlaa with SmoothSplatNet
🥈 https://t.co/gw4YJmK4ek with CDFDoubleGridNet
🥉 https://t.co/CV0lU0uW7S with VRTEnsemble

https://t.co/S5KaCpGtcU

[GRaM Workshop at ICLR 2026 @GRaM_org_]

@wellingmax @mmbronstein @AlexanderTong7 @erikjbekkers A memorable final quote from @AlexanderTong7 (echoed by several of the panelists): "geometry is about understanding the world, and it is here to stay."

[GRaM Workshop at ICLR 2026 @GRaM_org_]

Next up: @mayabechlerspei is sharing her insights on billion-parameter graph foundation models!

[GRaM Workshop at ICLR 2026 @GRaM_org_]

Unmissable panel starting now! @wellingmax, @mmbronstein, Kathlén Kohn, Gabriel Loaiza-Ganem, and @AlexanderTong7 discuss "scale, simplicity, and geometry", moderated by @erikjbekkers

The first question: is scale really all you need? What is the role of inductive bias? 🌶️

[Maya Bechler-Speicher @mayabechlerspei]

Will be speaking today at the GRaM workshop at ICLR @iclr_conf (https://t.co/hsMRCrK6s9 ) about how we are building efficient and effective billion-scale Graph Foundation Models at Meta!!! 🤩
(GraphBFF 👉 https://t.co/MPdRT0qU8r).
Anddddd we also have an awesome paper at the workshop! "Improving LLM Predictions via Inter-Layer Structural Encoders", poster session A :) (https://t.co/xyiUtKnO2S)
#iclr2026

[GRaM Workshop at ICLR 2026 @GRaM_org_]

Gabriel Loaiza-Ganem is speaking NOW on "A Geometric View of Deep Generative Models"!

The claim: generative models can be divided into (data) manifold-aware and manifold-unaware methods. Manifold-aware methods work better!

[GRaM Workshop at ICLR 2026 @GRaM_org_]

GRaM 2026 is kicking off NOW in 101A with @erikjbekkers!

Full schedule at https://t.co/xYVH0RPNjk #doitfortheGRaM

[GRaM Workshop at ICLR 2026 @GRaM_org_]

We're excited to kick off the GRaM Workshop tomorrow (Sunday) at 9am in Room 101A! Check out the paper search tool on our website, where you can find posters at the intersection of your interests :)

Full schedule + speakers at https://t.co/nSV29xzPt2

#doitfortheGRaM

[Hannah Lawrence @HLawrenceCS]

Was the whole blogpost an excuse to make this meme? Perhaps.

[Julia Balla @julballa]

We must choose both the tokens AND their order.
A joint distribution can be factored into a product of conditionals in any order, but this choice determines the available context at each step. Thus, some orders are easier to model in practice (https://t.co/gYAHZRi62h)
(2/7)

[Hannah Lawrence @HLawrenceCS]

Come chat about tokenization and trends in the treatment of non-sequential data!

P4-5313

[Julia Balla @julballa]

New blogpost on tokenizing non-sequential data!

Language has sequential structure, which gave rise to the next-token prediction paradigm of LLMs. But we increasingly use LLMs for data without inherent order (e.g. images, molecules, sets). What does “next token” mean here?
(1/7)

[Samuel Stanton @samuel_stanton_]

You have a safe model you've tested, and you have a new post-trained model. How far can you trust a new model before it becomes unsafe? The answer goes to the heart of statistical decision-making. To be safe, the agent must be self-aware. Read @DrewPrinster's thread for more.

[Hannah Lawrence @HLawrenceCS]

It was a pleasure working on this with co-first author @ElyssaHofgard and great collaborators @vportilheiro, Yuxuan Chen, @tesssmidt, and @RobinSFWalters. Come chat with Elyssa, Vasco, and I this afternoon!

Paper: https://t.co/5UQX19zdKk
Blog post: https://t.co/IFq7D5n8mo

[Hannah Lawrence @HLawrenceCS]

TL;DR: poster today at 3:15pm, P3-#1109!

Have you ever benchmarked your method on QM9, MD17, OC20, or ModelNet? It turns out that the 3D orientations of point clouds in commonly used datasets are highly non-random. How did we prove this and why should you care? 🧵

[Hannah Lawrence @HLawrenceCS]

So: where does this leave equivariant methods? Here’s a handy flowchart :)

(N.B. we’re lumping together equivariant architectures and augmentation here. They are of course different, and our experiments used both. But both provably discard input-level rotational biases!)