Naomi Saphra
banner
nsaphra.bsky.social
Naomi Saphra
@nsaphra.bsky.social
Waiting on a robot body. All opinions are universal and held by both employers and family.

Literally a professor. Recruiting students to start my lab.
ML/NLP/they/she.
Pinned
I wrote something up for AI people who want to get into bluesky and either couldn't assemble an exciting feed or gave up doomscrolling when their Following feed switched to talking politics 24/7.
The AI Researcher's Guide to a Non-Boring Bluesky Feed | Naomi Saphra
How to migrate to bsky without a boring feed.
nsaphra.net
Reposted by Naomi Saphra
Our next paper on comparing dynamical systems (with special interest to artificial and biological neural networks) is out!! Joint work with @annhuang42.bsky.social , as well as @satpreetsingh.bsky.social , @leokoz8.bsky.social , Ila Fiete, and @kanakarajanphd.bsky.social : arxiv.org/pdf/2510.25943
November 10, 2025 at 4:16 PM
Reposted by Naomi Saphra
Through this release, we aim both to support the emerging ecosystem for pretraining research (NanoGPT, NanoChat), explainability (you can literally look at Monad under a microscope) and the tooling orchestration around frontier models.
November 10, 2025 at 5:34 PM
Reposted by Naomi Saphra
Oh cool! Excited this LM + construction paper was SAC-Highlighted! Check it out to see how LM-derived measures of statistical affinity separate out constructions with similar words like "I was so happy I saw you" vs "It was so big it fell over".
November 10, 2025 at 4:27 PM
Reposted by Naomi Saphra
huge news for the worst architects you know
November 9, 2025 at 8:16 PM
The Boston Ballet (on now, recommend) is my 2nd time seeing Balanchine's Jewels and I realized: Emeralds/Rubies/Diamonds are known portrayals of France/US/Russia but also are how you'd see each if you were a child in Petersburg, a youth in Paris, and a stalker who married 4 of his own dancers in NY.
November 8, 2025 at 4:29 PM
Reposted by Naomi Saphra
Delighted Sasha's (first year PhD!) work using mech interp to study complex syntax constructions won an Outstanding Paper Award at EMNLP!

Also delighted the ACL community continues to recognize unabashedly linguistic topics like filler-gaps... and the huge potential for LMs to inform such topics!
November 7, 2025 at 6:22 PM
Reposted by Naomi Saphra
It’s grad school application season, and I wanted to give some public advice.

Caveats:
-*-*-*-*


> These are my opinions, based on my experiences, they are not secret tricks or guarantees

> They are general guidelines, not meant to cover a host of idiosyncrasies and special cases
November 6, 2025 at 2:55 PM
Reposted by Naomi Saphra
Either we now have a truly excellent open-source model or it's benchmark hacked. Soon we will know which world we live in
November 6, 2025 at 8:29 PM
Reposted by Naomi Saphra
This story is getting wilder - not only is it a known solitary spider species living communally, but it's cohabiting with a second species, a type of commensalism(?) not before observed. Would be very interested to see if the species are cooperating with weaving/hunting, and how that works.
@aptshadow.bsky.social Children of Time as prophecy?
This ‘arachnid megacity’ may be largest spider’s web ever found
November 6, 2025 at 2:42 PM
the only kind of Rat Race I'm down for
November 6, 2025 at 2:43 PM
Reposted by Naomi Saphra
Sci-fi short story:
scientist discovers how to create superhuman artificial intelligence (ASI): you need to train your #LLM on deep center embeddings! But scientist is disinterested in this direction.

…So in order to do so, tech bros would need to study #linguistics

Working title: X-Bar X-Risk
Transformer LMs get pretty far by acting like ngram models, so why do they learn syntax? A new paper by sunnytqin.bsky.social, me, and @dmelis.bsky.social illuminates grammar learning in a whirlwind tour of generalization, grokking, training dynamics, memorization, and random variation. #mlsky #nlp
Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization
Language models (LMs), like other neural networks, often favor shortcut heuristics based on surface-level patterns. Although LMs behave like n-gram models early in training, they must eventually learn...
arxiv.org
November 6, 2025 at 12:04 AM
If I were an ascending superpower right now, I would do everything I could to convince the dumbass legacy rival superpower to pour everything into AI, while I spent a large majority of government science funds cementing my dominance in renewable energy.
November 5, 2025 at 11:53 PM
While lead author @sunnytqin.bsky.social sadly couldn't go to her !!!HOMETOWN!!!! of Suzhou due to visa reentry issues, her EMNLP paper with @dmelis.bsky.social and me is still fantastically cool and I will absolutely take advantage of EMNLP week to reshare it.
Transformer LMs get pretty far by acting like ngram models, so why do they learn syntax? A new paper by sunnytqin.bsky.social, me, and @dmelis.bsky.social illuminates grammar learning in a whirlwind tour of generalization, grokking, training dynamics, memorization, and random variation. #mlsky #nlp
Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization
Language models (LMs), like other neural networks, often favor shortcut heuristics based on surface-level patterns. Although LMs behave like n-gram models early in training, they must eventually learn...
arxiv.org
November 5, 2025 at 11:03 PM
Reposted by Naomi Saphra
🆕 from us at #EMNLP: Are LMs better at answering questions about Germany in German than in French? Is national knowledge linguistically contingent?

Interestingly, only for some multilingual models is this true. Aya knows China best in Chinese, but LLaMA's best in English always.
November 5, 2025 at 7:47 PM
having moved to Boston I do miss NYC politics a little. NYC is like, edge of your seat, always something dumb / triumphant happening. you can see boston's beloved mayor playing piano concerts with a big 5 orchestra on the side. a football heir decided to run against her then changed his mind.
November 5, 2025 at 2:24 PM
Reposted by Naomi Saphra
Flying out to @emnlpmeeting soon🇨🇳
I'll present our parametric CoT faithfulness work (arxiv.org/abs/2502.14829) on Wednesday at the second Interpretability session, 16:30-18:00 local time A104-105

If you're in Suzhou, reach out to talk all things reasoning :)
Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps
When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work o...
arxiv.org
October 31, 2025 at 1:30 PM
Reposted by Naomi Saphra
I wrote a short blog post about masked softmax layers in PyTorch (i.e., when you have structural constraints that tell you some classes _must_ have probability zero).

This was based on a real bug I found in a neural chess model implementation.
Masked Softmax Layers in PyTorch
Correctly computing masked softmax layers.
mcognetta.github.io
November 3, 2025 at 7:39 PM
The first time I didn't vote for Cuomo was because he betrayed unions.
The second time I didn't vote for Cuomo was because he made a fake party to steal votes from the WFP line.
The third time I didn't vote for Cuomo was because he's a creep.
Jealous of NYers who get to not vote for Cuomo tomorrow!
November 3, 2025 at 9:37 PM
Reposted by Naomi Saphra
November 2, 2025 at 4:18 AM
Reposted by Naomi Saphra
This game is like I got caught smoking a baseball and now I have to smoke a whole pack of baseball
October 28, 2025 at 5:01 AM
Seeing a weird number of people misunderstand the policy. It applies ONLY to position papers and surveys, which have no new results and only present opinions. go bluejays
November 2, 2025 at 3:54 AM
Reposted by Naomi Saphra
A dishwasher is already a robot that does your dishes
The loading-the-dishwasher struggle is real. Even—or especially—for the $20,000 1X Neo humanoid home robot.

🔗: on.wsj.com/47whlZ1
November 1, 2025 at 3:33 AM
Reposted by Naomi Saphra
Need to establish a norm against making the manifold chip-coloured so im not hungry reading papers
November 1, 2025 at 1:00 PM
I thought I'd never use an LLM to actually write until I discovered that grants all require a gruelingly self-aggrandizing personal statement
October 31, 2025 at 10:51 PM