Lightnews — Scholar-powered news

Maria Ryskina

@mryskina.bsky.social

100 followers 140 following 23 posts

Postdoc @vectorinstitute.ai | organizer @queerinai.com | previously MIT, CMU LTI | 🐀 rodent enthusiast | she/they 🌐 https://ryskina.github.io/

ryskina.github.io

Posts Media Videos Starter Packs

Pinned

Maria Ryskina @mryskina.bsky.social · 11d

Interested in language models, brains, and concepts? Check out our COLM 2025 🔦 Spotlight paper!

(And if you’re at COLM, come hear about it on Tuesday – sessions Spotlight 2 & Poster 2)!

Paper title: Language models align with brain regions that represent concepts across modalities.
Authors: Maria Ryskina, Greta Tuckute, Alexander Fung, Ashley Malkin, Evelina Fedorenko.
Affiliations: Maria is affiliated with the Vector Institute for AI, but the work was done at MIT. All other authors are affiliated with MIT.
Email address: maria.ryskina@vectorinstitute.ai.

1 5 26

Reposted by Maria Ryskina

Maria Antoniak @mariaa.bsky.social · 19h

Inspired to share some papers that I found at #COLM2025!

"Register Always Matters: Analysis of LLM Pretraining Data Through the Lens of Language Variation" by Amanda Myntti et al. arxiv.org/abs/2504.01542

Title, authors, and abstract of the paper.

Figure 3: Change of accuracy from first to final checkpoint on individual benchmarks shown as a range, with grey indicating the first checkpoint and colours indicating the last checkpoint. The random-guess threshold is shown as a grey vertical line in cases where at least one model falls below it. Bars and legend shown in order of average accuracy.

1 8 23

Maria Ryskina @mryskina.bsky.social · 1d

LLMs Assume People Are More Rational Than We Really Are by Ryan Liu* & Jiayi Geng* et al.:

LMs are bad (too rational) at predicting human behaviour, but aligned with humans in assuming rationality in others’ choices.

arxiv.org/abs/2406.17055

Title: Large Language Models Assume People are More Rational than We Really are
Authors: Ryan Liu*, Jiayi Geng*, Joshua C. Peterson, Ilia Sucholutsky, Thomas L. Griffiths
Affiliations: Department of Computer Science & Department of Psychology, Princeton University; Computing & Data Sciences, Boston University; Center for Data Science, New York University
Email: ryanliu at princeton.edu and jiayig at princeton.edu

Maria Ryskina @mryskina.bsky.social · 1d

Neologism Learning by John Hewitt et al.:

Training new token embeddings on examples with a specific property (e.g., short answers) leads to finding “machine-only synonyms” for these tokens that elicit the same behaviour (short answers=’lack’).

arxiv.org/abs/2510.08506

Title: Neologism Learning for Controllability and Self-Verbalization
Authors: John Hewitt, Oyvind Tafjord, Robert Geirhos, Been Kim
Affiliation: Google DeepMind
Email: {johnhew, oyvindt, geirhos, beenkim} at google.com

Maria Ryskina @mryskina.bsky.social · 1d

Hidden in Plain Sight by Stephanie Fu et al. [Outstanding paper award]:

VLMs are worse than vision-only models on vision-only tasks – LMs are biased and underutilize their (easily accessible) visual representations!

hidden-plain-sight.github.io

Title: Hidden in plain sight: VLMs overlook their visual representations
Authors: Stephanie Fu, Tyler Bonnen, Devin Guillory, Trevor Darrell
Affiliation: UC Berkeley

1 1

Maria Ryskina @mryskina.bsky.social · 1d

UnveiLing by Mukund Choudhary* & KV Aditya Srivatsa* et al.:

Linguistic olympiad problems about certain linguistic features (e.g., morphological ones) are tougher for LMs, but morphological pre-tokenization helps!

arxiv.org/abs/2508.11260

Title: UnveiLing: What Makes Linguistics Olympiad Puzzles Tricky for LLMs?
Authors: Mukund Choudhary*, KV Aditya Srivatsa*, Gaurja Aeron , Antara Raaghavi Bhattacharya, Dang Khoa Dang Dinh, Ikhlasul Akmal Hanif, Daria Kotova, Ekaterina Kochmar, Monojit Choudhury
Affiliations: Mohamed Bin Zayed University of Artificial Intelligence, IIT Gandhinagar, Harvard University, VinUniversity, Universitas Indonesia

Maria Ryskina @mryskina.bsky.social · 1d

A Taxonomy of Transcendence by Natalie Abreu et al.:

LMs outperform the experts they are trained on through skill denoising (averaging out experts’ errors), skill selection (relying on the most appropriate expert), and skill generalization (combining experts’ knowledge).

arxiv.org/abs/2508.17669

Title: A Taxonomy of Transcendence
Authors: Natalie Abreu, Edwin Zhang, Eran Malach, & Naomi Saphra
Affiliations: Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University
Email: {natalieabreu, ezhang} at g.harvard.edu and {emalach, nsaphra} at fas.harvard.edu

1 3

Maria Ryskina @mryskina.bsky.social · 1d

The Zero Body Problem by @rmmhicke.bsky.social et al.:

LMs use sensory language (olfactory, auditory, …) differently from people + evidence that RLHF may discourage sensory language.

arxiv.org/abs/2504.06393

Title: The Zero Body Problem: Probing LLM Use of Sensory Language
Authors: Rebecca M. M. Hicke, Sil Hamilton & David Mimno
Affiliations: Department of Computer Science & Department of Information Science, Cornell University, Ithaca, New York, USA
Email: {rmh327, srh255, mimno} at cornell.edu

1 4

Maria Ryskina @mryskina.bsky.social · 1d

Readability ≠ Learnability by Ivan Lee & Taylor Berg-Kirkpatrick:

Developmentally plausible LM training works not because of simpler language but because of lower n-gram diversity! Warning against anthropomorphizing / equating learning in LMs and in children.

openreview.net/pdf?id=AFMGb...

Title: Readability ≠ Learnability: Rethinking the Role of Simplicity in Training Small Language Models
Authors: Ivan Lee & Taylor Berg-Kirkpatrick
Affiliation: UC San Diego
Email: {iylee, tberg} at ucsd.edu

1 5

Maria Ryskina @mryskina.bsky.social · 1d

⭐ A thread for some cool recent work I learned about at #COLM2025, either from the paper presentations or from the keynotes!

1 6 13

Reposted by Maria Ryskina

Juan Diego Rodriguez @juand-r.bsky.social · 1d

👀

Phillip Isola @phillipisola.bsky.social · 4d

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis.

Today I want to share two new works on this topic:

Eliciting higher alignment: arxiv.org/abs/2510.02425
Unpaired learning of unified reps: arxiv.org/abs/2510.08492

1/9

1 3

Reposted by Maria Ryskina

Queer in AI @queerinai.com · 6d

We are launching our Graduate School Application Financial Aid Program (www.queerinai.com/grad-app-aid) for 2025-2026. We’ll give up to $750 per person to LGBTQIA+ STEM scholars applying to graduate programs. Apply at openreview.net/group?id=Que.... 1/5

Grad App Aid — Queer in AI

www.queerinai.com

1 9 7

Reposted by Maria Ryskina

Maria Antoniak @mariaa.bsky.social · 6d

Keynote at #COLM2025: Nicholas Carlini from Anthropic

"Are language models worth it?"

Explains that the prior decade of his work on adversarial images, while it taught us a lot, isn't very applied; it's unlikely anyone is actually altering images of cats in scary ways.

2 22 79

Reposted by Maria Ryskina

Conference on Language Modeling @colmweb.org · 8d

Outstanding paper 🏆 1: Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling
openreview.net/forum?id=3Bm...

1 1 3

Reposted by Maria Ryskina

Naomi Saphra @nsaphra.bsky.social · Aug 29

How can an imitative model like an LLM outperform the experts it is trained on? Our new COLM paper outlines three types of transcendence and shows that each one relies on a different aspect of data diversity. arxiv.org/abs/2508.17669

3 17 95

Reposted by Maria Ryskina

Greta Tuckute @gretatuckute.bsky.social · 11d

Check out @mryskina.bsky.social's talk and poster at COLM on Tuesday—we present a method to identify 'semantically consistent' brain regions (responding to concepts across modalities) and show that more semantically consistent brain regions are better predicted by LLMs.

Maria Ryskina @mryskina.bsky.social · 11d

Interested in language models, brains, and concepts? Check out our COLM 2025 🔦 Spotlight paper!

(And if you’re at COLM, come hear about it on Tuesday – sessions Spotlight 2 & Poster 2)!

4 14

Maria Ryskina @mryskina.bsky.social · 11d

Also, if you're at COLM, come to Gillian's keynote and find out what our lab is working on!

Conference on Language Modeling @colmweb.org · 22d

Keynote spotlight #4: the second day of COLM will close with @ghadfield.bsky.social from JHU talking about human society alignment, and lessons for AI alignment

2 1

Maria Ryskina @mryskina.bsky.social · 11d

🧠 This paper is a part of my first brain project, and I had a lot of fun working on it with the amazing @gretatuckute.bsky.social, @alexanderfung.bsky.social, Ashley Malkin, and @evfedorenko.bsky.social!

Paper link: arxiv.org/abs/2508.11536
Code+data: github.com/ryskina/conc...

Language models align with brain regions that represent concepts across modalities

Cognitive science and neuroscience have long faced the challenge of disentangling representations of language from representations of conceptual meaning. As the same problem arises in today's language...

arxiv.org

1 1 14

Maria Ryskina @mryskina.bsky.social · 11d

We also compare the representational geometries of the models and the brain using RSA, and find significant alignment in all models. For VLMs, it further increases when both text and image stimuli are used, especially in the ventral ROI:

Title: 5. Representational similarity analysis.

Left: RSA diagram. Stimuli (word clouds, pictures, and sentences for concepts Help, Bird, and Broken) are encoded first via LM hidden states and then via brain activation. Each set of representations is then converted into a distance matrix showing how dissimilar each pair of concepts is in this representational space. Finally, a correlation is computed between the distance matrices.

Right: 3 bar charts of RSA correlations, labeled ROI 1, 2, and 3. Each chart includes two groups of bars labeled LMs and VLMs. Both groups include bars for Permuted baseline and for Result (text only). The VLMs group includes an additional bar for Result (text + images). All Result bars are significantly higher than the Permuted baseline bars. Result (text only) bars are comparable between LMs and VLMs. For VLMs, Result (text + image) is higher than Result (text only) in ROI 1 and especially ROI 3, but comparable in ROI 2.

1 2

Maria Ryskina @mryskina.bsky.social · 11d

2. Within our ROIs, LM predictivity is correlated with semantic consistency, even where response to meaningful language (preference for sentences over non-words) is low:

3 rows and 6 columns of line charts. Rows are labeled ROI 1, 2, and 3. Pairs of columns are labeled Sentence paradigm, Picture paradigm, or Word cloud paradigm. The X axis is Predictivity. Odd columns are labeled Controlled for language, Y axis is Quartile (cons.), and the chart lines are red. The red lines all trend upwards. Even columns are labeled Controlled for consistency, Y axis is Quartile (lang.), and the chart lines are blue. The blue lines show a moderate to strong upward trend for ROI 1 and 3 in Sentence and Word cloud paradigms but no clear trend elsewhere.

Maria Ryskina @mryskina.bsky.social · 11d

In our brain encoding experiments (using LM representations to predict brain responses), we find that:

1. Across the whole brain, areas with higher semantic consistency – plausibly representing concepts across modalities – are better predicted by LMs:

Title: 4. LMs for brain encoding.

Left: Brain encoding diagram. Stimulus (He hit the ball out of the field) is encoded first via LM hidden states and then via brain activation. Then a regression model is used to predict the brain activation from LM hidden states.

Right: Scatter plots of brain encoding performance for the Sentence, Picture, and Word cloud paradigms. Y axis is Mean LM predictivity. X axis is Semantic consistency. Each point corresponds to a Glasser anatomical area, with areas within ROI 1, 2, or 3 specially marked. A line shows a positive trend in each plot. A cluster of points is highlighted above the line in the Picture paradigm plot.

Maria Ryskina @mryskina.bsky.social · 11d

…we define semantic consistency – a measure of how consistently a brain area responds to the same concepts across paradigms – and use it to identify three regions of interest (ROI) where semantically consistent voxels are found:

Left: figure titled 2. Semantic consistency metric. Subtitle: Correlation between brain responses to concepts across stimuli formats. Three columns of squares represent three vectors, one square per element, with top and bottom elements labelled Concept 1 and Concept 180 respectively. Vectors are labelled Beta with subscript S, P, or WC for paradigm. Each pair of vectors is connected by an arc labelled R for Pearson correlation. From each R, a line leads to a circle with C in it.

Right: figure titled: 3. Consistent brain areas.
Top: heatmap of the left hemisphere titled Probabilistic map. The jet colormap ranges from blue for 0% to red for 30%. The heatmap shows higher values in the posterior temporal, inferior frontal, and ventrotemporal areas.
Bottom: a segmentation of the left hemisphere titled Regions of interest. Posterior temporal, inferior frontal, and ventrotemporal areas are colored in pink (labeled ROI 1), blue (ROI 2), and green (ROI 3) respectively.

Maria Ryskina @mryskina.bsky.social · 11d

Using an fMRI dataset of brain responses to stimuli that encode concepts in different paradigms (sentences, pictures, word clouds) and modalities…

Figure title: 1. Brain responses to concepts (Pereira et al., 2018, Exp. 1).

Left: an icon of a human brain, captioned n=17, for the number of participants. and an icon of an MRI scanner, captioned fMRI.

Right: three boxes connected with arrows, representing three experimental paradigms. Each box includes text: 180 concepts × 4–6.
The first box is yellow, titled Sentences and contains a sentence: He hit the ball out of the field. The word ball is highlighted in bold.
The second box is purple, titled Pictures and contains a picture of a brown snake coiled on the ground, with the caption Dangerous highlighted in bold.
The third box is green, titled Word clouds and contains a word cloud with the word Bear in the middle highlighted in bold and surrounded by related words: Squirrel, Duck, Deer, Wolf, Raccoon.

Maria Ryskina @mryskina.bsky.social · 11d

Interested in language models, brains, and concepts? Check out our COLM 2025 🔦 Spotlight paper!

(And if you’re at COLM, come hear about it on Tuesday – sessions Spotlight 2 & Poster 2)!

1 5 26

Reposted by Maria Ryskina

Queer in AI @queerinai.com · 13d

Attending COLM next week in Montreal? 🇨🇦 Join us on Thursday for a 2-part social! ✨ 5:30-6:30 at the conference venue and 7:00-10:00 offsite! 🌈 Sign up here: forms.gle/oiMK3TLP8ZZc...

4 4

Maria Ryskina @mryskina.bsky.social · 29d

Yay! Go Leonie!