iiinova — where your digital self keeps evolving

Reposted by Women in AI Research - WiAIR

Vered Shwartz @veredshwartz.bsky.social · 13/10/2026

Check out our new paper taking the guesswork out of value alignment! We show how to best represent values, and predict cross-value generalization, and build the first taxonomy of LLM values grounded in generalization dynamics.

021

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

The results support extending the noisy representation hypothesis to compositional language: distinctive meanings are less confusable in memory, while similar meanings are more susceptible to interference. 🎙️ www.youtube.com/watch?v=efnZ... 📄 doi.org/10.1016/j.jm... (6/6)

youtube.com

What Makes a Sentence Memorable? Inside Language, Memory and LLMs, with Dr. Greta Tuckute

YouTube video by Women in AI Research WiAIR

000

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

🔁 Context matters too. False alarms were more likely when a sentence shared a content word with, or was semantically similar to, a previously seen sentence. The closest semantic distractor shared a content word only 6.7% of the time. (5/6)

110

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

✨ Among six semantic-distinctiveness measures, SBERT-based distinctiveness had the highest predictive power. It predicted recognition accuracy beyond word-level features and was also associated with faster responses when participants endorsed a sentence as familiar. (4/6)

100

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

🧩 Constituent words matter: average word memorability significantly predicts sentence memorability. But sentence memory is not reducible to the words alone. Semantic distinctiveness explains additional variance beyond word memorability and word frequency. (3/6)

100

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

The study tested 2,500 six-word target sentences with 443 participants. Memorability varied substantially across sentences and was moderately consistent across participants, suggesting that intrinsic sentence properties affect how likely a sentence is to be remembered. (2/6)

100

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

🧠Why are some sentences easier to remember than others? In our latest WiAIR podcast, we spoke with Dr. Greta Tuckute about her paper, “A distinctive meaning makes a sentence memorable”, on what predicts sentence recognition memory. (1/6)

100

Women in AI Research - WiAIR @wiair.bsky.social · 06/10/2026

Do all language models learn things in the same order - even when they are trained on different data? Emmy Liu gives the answer in the video. #WiAIR is covering COLM 2026 this week, with a new paper every day youtu.be/LZsEasfVc7Q #COLM2026 #NLProc

youtu.be

COLM 2026 in San Francisco. Do All LLMs Learn in the Same Order?

YouTube video by Women in AI Research WiAIR

010

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

🎙️ Full episode with Dr. Tuckute 👇 📺 YouTube: www.youtube.com/watch?v=efnZ... 🎧 Spotify: open.spotify.com/episode/1Bkp... 🍏 Apple Podcasts: podcasts.apple.com/ca/podcast/w... 📄 Paper: arxiv.org/abs/2606.06857

youtube.com

What Makes a Sentence Memorable? Inside Language, Memory and LLMs, with Dr. Greta Tuckute

YouTube video by Women in AI Research WiAIR

010

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

📊 Frontal and temporal language regions differ: MFG shows no significant gain beyond surprisal, while temporal fROIs show a significant gain from feature-based models. Within the Matryoshka hierarchy, the most general 0.4% of features are most predictive. (5/6🧵)

101

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

🧠 The method identifies a coherent 32-voxel “Ghost” subpopulation associated with people-centered content, including relationships/emotions, actions involving people, and pronouns. The cluster is concentrated in a subset of participants. (4/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

🔍 Sparse SAE features predict voxel responses nearly as well as dense LM residual-stream features. Hard-to-Process voxels show no significant gain beyond surprisal, while Abstract and Concrete voxels benefit significantly from LM-derived features. (3/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

The study introduces Augmented Sparse Encoding Models, combining sparse autoencoder (SAE) features with surprisal. It tests them on 7T fMRI data from 8 participants listening to 200 linguistically diverse sentences. (2/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 30/09/2026

Can sparse, interpretable LM features help us understand how the brain responds to language? 🧠 In our latest WiAIR podcast, we spoke with Dr. Greta Tuckute about her co-authored paper, “Interpreting Brain Responses to Language with Sparse Features from Language Models.” (1/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

🎙️ Full episode with Dr. Greta Tuckute: 📺 YouTube: lnkd.in/gscRVkwp 📄 Paper: arxiv.org/abs/2503.01830 #WiAIR #AIResearch #Neuroscience #LLMs (6/6🧵)

lnkd.in

This link will take you to a page that’s not on LinkedIn

000

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

🔄 Early in training, next-word prediction and behavioral alignment correlate with brain alignment. These links weaken later, as models surpass human next-word prediction proficiency. With feature size controlled, model size is not a reliable predictor of brain alignment. (5/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

📚 Brain alignment tracks formal linguistic competence, including grammar, syntax, and compositional rules, more closely than functional competence involving world knowledge and reasoning. Functional competence keeps improving after brain alignment saturates. (4/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

🧠 Brain alignment rises sharply early in training, peaks around 2 to 8B tokens, then largely saturates or declines. The trajectory is similar across very different model sizes. (3/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

The study evaluates 8 Pythia model sizes across 34 training checkpoints spanning ~300B tokens, asking how brain alignment relates to linguistic competence throughout training. (2/6🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 23/09/2026

🧠 Does training an LLM longer make it more aligned with the human language network? In our latest #WiAIR episode, we spoke with Dr. Greta Tuckute about her co-authored paper, “From Language to Cognition: How LLMs Outgrow the Human Language Network.” (1/6🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 17/09/2026

The conversation covers how brain alignment develops over training, how sparse autoencoders can turn the "you are just comparing one black box to another" critique into something testable, and why "pineapple" sticks in memory and "light" does not

100

Women in AI Research - WiAIR @wiair.bsky.social · 17/09/2026

New episode of #WiAIRpodcast: Greta Tuckute @gretatuckute.bsky.social talks to @j-novikova-nlp.bsky.social about what brain-LLM alignment does and does not tell us, and why, asked to keep one label out of neuroscience, cognitive science and AI, she drops AI first

152

Women in AI Research - WiAIR @wiair.bsky.social · 14/09/2026

🎙️ Our new guest is live on #WiAIR soon! Dr.Greta Tuckute @gretatuckute.bsky.social, Research Fellow at the Kempner Institute at Harvard University, previously PhD at MIT. Greta will speak with us about language in biological brains and in artificial language models. Don't miss this episode!

051

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

We have a long way to go to build AI that truly understands data! Read the full breakdown and listen to the episode here 👇 📺 YT: www.youtube.com/watch?v=jtQq... 🎧 Spotify: open.spotify.com/episode/40RF... 🍏 Apple: podcasts.apple.com/ca/podcast/i... 📄 Paper: arxiv.org/abs/2505.17202 #WiAIR (7/7🧵)

youtube.com

Is an Image Worth a Thousand Words? Hidden Failures of Multimodal Metrics, with Dr. Elisa Kreiss

YouTube video by Women in AI Research WiAIR

000

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

Most interestingly, model error patterns are reliably distinct from human error patterns. This means AI doesn't just fail; it fails in ways that are entirely non-human. (6/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

Even when researchers gave the models the benefit of the doubt—only looking at valid, correctly formatted responses—the AI still fell short of human performance levels. (5/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

The verdict? 📉 Current vision-language models consistently underperform humans on graphical reasoning tasks. (4/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

🖼 The researchers created CHART-6, a suite of six assessments designed to measure data visualization literacy, and tested eight top VLMs (including GPT-4V) against human participants. (3/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

The paper, "CHART-6: Human-Centered Evaluation of Data Visualization Understanding in VLMs," tests if AI truly emulates human reasoning when looking at graphs. (2/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 08/09/2026

Do AI models really understand charts and graphs the same way we do? 📊 In the latest WiAIR Women in AI Research podcast episode, we are diving into another brilliant paper co-authored by our guest, Dr. Elisa Kreiss! 🎙 (1/7🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

We need evaluation metrics that prioritize specificity over verbosity! Don't miss this deep dive into VLM evaluation! 👇 📺 YT: www.youtube.com/watch?v=jtQq... 🎧 Spotify: open.spotify.com/episode/40RF... 🍏 Apple: podcasts.apple.com/ca/podcast/i... 📄 Paper: arxiv.org/pdf/2601.04609 #WiAIR (7/7🧵)

youtube.com

Is an Image Worth a Thousand Words? Hidden Failures of Multimodal Metrics, with Dr. Elisa Kreiss

YouTube video by Women in AI Research WiAIR

000

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

Prompting also changes everything 💡! The researchers found that giving a VLM a strict character limit decreased specificity, but simply prompting the model to be "concise" actually increased it. (6/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

The findings? Humans reliably prefer more specific descriptions regardless of their length. Controlling for length alone simply cannot account for differences in a description's specificity. (5/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

🖼 To prove this, the team constructed a dataset featuring "verbose" descriptions (long but empty) and "composite" descriptions (concise but highly detailed). (4/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

But length does not equal quality! The paper argues these concepts must be disentangled. Descriptions can be lengthy yet vacuous, or concise yet dense with information. (3/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

We are discussing her co-authored paper, "When More Words Say Less," which highlights a massive flaw in how we evaluate Vision-Language Models (VLMs) 📉. In current systems, the specificity of an image description is frequently conflated with its length. (2/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 01/09/2026

Are longer AI image descriptions actually better, or just more verbose? 🤔 We dive into this in the latest WiAIR Women in AI Research podcast with Dr. Elisa Kreiss! 🎙 (1/7🧵)

100

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

Don't miss this eye-opening conversation hosted by Jekaterina Novikova, PhD! 👇 📺 YT: www.youtube.com/watch?v=jtQq... 🎧 Spotify: open.spotify.com/episode/40RF... 🍏 Apple: podcasts.apple.com/ca/podcast/i... 📄 Paper: arxiv.org/pdf/2205.10646 #WiAIR #MultimodalAI (7/7🧵)

youtube.com

Is an Image Worth a Thousand Words? Hidden Failures of Multimodal Metrics, with Dr. Elisa Kreiss

YouTube video by Women in AI Research WiAIR

010

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

The path forward? 💡 As a proof-of-concept, the paper shows that upgrading existing metrics to include context-awareness significantly aligns AI evaluations with actual BLV user needs. We must evaluate AI based on what information matters to the end-user! (6/7🧵)

130

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

🖼 Look at the exact same image of a gazebo and sculpture. A description that is "good" for a Wikipedia page on Gazebos becomes a "bad" description when placed on a Sculpture page. One-size-fits-all AI descriptions simply do not work! (5/7🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

To prove this, a study asked BLV and sighted participants to rate descriptions embedded within distinct Wikipedia articles 📊. The result? Contextual relevance is a major driver of overall quality assessments. Raw accuracy isn’t enough; relevance matters! (4/7🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

But for blind and low vision (BLV) users, a description's usefulness isn't just about identifying objects—contextual information is highly valued 🦯! The immediate context of an image dictates what details are actually necessary to describe. (3/7🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

Dr. Kreiss's paper, "Context Matters for Image Descriptions for Accessibility," reveals a huge blindspot in multimodal AI 📉. Current referenceless evaluation metrics like CLIPScore and SPURTS assess image descriptions in a vacuum, ignoring context entirely! (2/7🧵)

140

Women in AI Research - WiAIR @wiair.bsky.social · 25/08/2026

Have you ever wondered why an AI-generated image description that seems perfectly accurate might be completely useless to someone using a screen reader? 🤔 We tackle this in the latest WiAIR Women in AI Research podcast with Dr. Elisa Kreiss! 🎙 (1/7🧵)

110

Women in AI Research - WiAIR @wiair.bsky.social · 20/08/2026

🎙️ New #WiAIR episode's out! We speak with @elisakreiss.bsky.social about how context matters in image descriptions, why the metrics still assume there's one correct answer, how blind and low-vision users were left out of accessibility research & much more Youtube: youtu.be/jtQq8NspkY4

032

Women in AI Research - WiAIR @wiair.bsky.social · 17/08/2026

Her argument: an image isn't worth a thousand words. It's worth an indefinite number of them. There's no single "correct" description of an image, it depends on who's asking, why, and what they already know.

100

Women in AI Research - WiAIR @wiair.bsky.social · 17/08/2026

🎙️ New guest alert! @elisakreiss.bsky.social joins the next episode of the #WiAIRpodcast. She's an Assistant Professor of Communication at UCLA, PhD in Linguistics from Stanford, now directing the Coalas Lab and studying how language and images work together to communicate meaning.

100