iiinova — where your digital self keeps evolving

Reposted by Divya Shanmugam

Ira Globus-Harris @iraglobusharris.bsky.social · 03/07/2026

Are you at ICML next week? Feel like your decision-making for which sessions to attend might not be risk minimizing? Don't incur (swap) regret and come to my, @aaroth.bsky.social, and @ncollina.bsky.social's tutorial Monday on multicalibration, decision-making, and collaborative learning!

22011

Divya Shanmugam @dmshanmugam.bsky.social · 24/03/2026

thanks for these comments, Ira :) :)

020

Divya Shanmugam @dmshanmugam.bsky.social · 23/03/2026

The goal is not just to remove race from models. It's to build systems where race no longer adds predictive power — because the factors it proxies for have been directly measured and the inequities it captures have been addressed.

141

Divya Shanmugam @dmshanmugam.bsky.social · 23/03/2026

Computational approaches to removing race are often insufficient. Why? Race often correlates with (1) unmeasured but relevant factors, like genetic traits (which should be measured directly) and (2) systemic disparities, like racism (which should be addressed, not adjusted for).

121

Divya Shanmugam @dmshanmugam.bsky.social · 23/03/2026

We lay out principles for these evaluations: measure not just model fit but downstream effects on treatment/resource allocation; report results by racial subgroup; and pair simulated analyses with post-deployment studies tracking real-world consequences. Lots of room here for more work!

110

Divya Shanmugam @dmshanmugam.bsky.social · 23/03/2026

Before changing the inputs to a clinical algorithm, it is critical to evaluate the consequences. Past work in nephrology and pulmonology has shown that removing race can have unpredictable effects, and can both improve and worsen disparities, making it essential to conduct rigorous evaluations.

120

Divya Shanmugam @dmshanmugam.bsky.social · 23/03/2026

New in Nature Health: how might we move towards a world in which race is not used in clinical algorithms? We need (1) careful comparison of race-aware and race-neutral algorithms and (2) systemic efforts to address underlying disparities.

1219

Reposted by Divya Shanmugam

Kenny Peng @kennypeng.bsky.social · 17/02/2026

New paper! The Linear Representation Hypothesis is a powerful intuition for how language models work, but lacks formalization. We give a mathematical framework in which we can ask and answer a basic question: how many features can be stored under the hypothesis? 🧵 arxiv.org/abs/2602.11246

14514

Reposted by Divya Shanmugam

Gabriel Agostini @gsagostini.bsky.social · 05/02/2026

We found, for example, racial disparities in upward mobility —that is, the rate at which people move to higher-income areas varies according to the racial composition of their current area of residence, even after controlling for income levels. 6/9

152

Reposted by Divya Shanmugam

Gabriel Agostini @gsagostini.bsky.social · 05/02/2026

Our paper “Inferring fine-grained migration patterns across the United States” is now out in @natcomms.nature.com! We released a new, highly granular migration dataset. 1/9

27327

Reposted by Divya Shanmugam

Maria Antoniak @mariaa.bsky.social · 29/01/2026

CS ArXiv recently banned “review and position” papers, but what are those? Do they include more generated content? Who is most affected by this change? @yanai.bsky.social and I dug into the data to find out! Nearly 50% of Computers & Society papers might be censored, vs 3% of Computer Vision ‼️

24319

Reposted by Divya Shanmugam

David Liu @david-m-liu.bsky.social · 28/01/2026

🎙️ I had a great time joining the Data Skeptic podcast to talk about my work on recommender systems If you're interested in embeddings, aligning group preferences, or music recommendations, check out the episode below 👇 open.spotify.com/episode/6IsP...

open.spotify.com

Fairness in PCA-Based Recommenders

1145

Reposted by Divya Shanmugam

Shuvom Sadhuka @shuvoms.bsky.social · 18/11/2025

I’m excited to share our new paper A Bayesian Model for Multi-stage Censoring, which I will present at #ML4H2025 in San Diego! 🧵 below:

171

Reposted by Divya Shanmugam

Jessica Hullman @jessicahullman.bsky.social · 05/11/2025

🧠⚙️ Interested in decision theory+cogsci meets AI? Want to create methods for rigorously designing & evaluating human-AI workflows? I'm recruiting PhDs to work on: 🎯 Stat foundations of multi-agent collaboration 🌫️ Model uncertainty & meta-cognition 🔎 Interpretability 💬 LLMs in behavioral science

13915

Reposted by Divya Shanmugam

Kate Donahue @kpaxdonahue.bsky.social · 06/11/2025

I’m recruiting students this upcoming cycle at UIUC! I’m excited about Qs on societal impact of AI, especially human-AI collaboration, multi-agent interactions, incentives in data sharing, and AI policy/regulation (all from both a theoretical and applied lens). Apply through CS & select my name!

14018

Divya Shanmugam @dmshanmugam.bsky.social · 06/11/2025

if you think about AI, healthcare, women's health, or all of the above, i highly recommend this article on the role of fetal heart rate monitors in the rise of C-sections: www.nytimes.com/2025/11/06/h...

nytimes.com

The ‘Worst Test in Medicine’ is Driving America’s High C-Section Rate

051

Divya Shanmugam @dmshanmugam.bsky.social · 03/11/2025

Super cool, and something I wish existed within machine learning for healthcare too! I'm often wondering what people are actually doing in practice and assembling evidence for my guesses.

040

Reposted by Divya Shanmugam

Angelina Wang @ COLM @angelinawang.bsky.social · 28/10/2025

Cornell (NYC and Ithaca) is recruiting AI postdocs, apply by Nov 20, 2025! If you're interested in working with me on technical approaches to responsible AI (e.g., personalization, fairness), please email me. academicjobsonline.org/ajo/jobs/30971

academicjobsonline.org

Cornell University, Empire AI Fellows Program

Job #AJO30971, Postdoctoral Fellow, Empire AI Fellows Program, Cornell University, New York, New York, US

13120

Reposted by Divya Shanmugam

Harini Suresh @harinisuresh.bsky.social · 25/04/2025

@michelleding.bsky.social has been doing amazing work laying out the complex landscape of "deepfake porn" and distilling the unique challenges in governing it. We hope this work informs future AI governance efforts to address the severe harms of this content - reach out to us to chat more!

042

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

p.s. we pronounce SSME as "Sesame" but you're welcome to your favorite pronunciation :)

000

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

Thanks also to our wonderful set of co-authors - Manish Raghavan (@manishraghav.bsky.social) , John Guttag, Bonnie Berger, and Emma Pierson (@emmapierson.bsky.social)-- without whom this work would not be possible!

110

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

Last but not least, thanks to @shuvoms.bsky.social, who co-led this work with me, and is an excellent thinking partner. Collaborate with him if you can!!

100

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

The paper includes much more, including theoretical connections to the literature on semi-supervised mixture models. Lots of exciting directions ahead – come chat with me and Shuvom at NeurIPS this December in San Diego! 📄 Paper: arxiv.org/abs/2501.11866 💻 Code: github.com/divyashan/SSME

100

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

Across 8 tasks, 4 metrics, and dozens of classifiers, SSME consistently outperforms prior work, reducing estimation error by 5.1× vs. using labeled data alone and 2.4× vs. the next-best method!

100

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

SSME starts with a set of classifiers, unlabeled data, and bit of labeled data, and estimates the joint distribution of classifier scores and ground truth labels using a mixture model. SSME benefits from three sources of info: multiple classifiers, unlabeled data, and classifier scores.

100

Divya Shanmugam @dmshanmugam.bsky.social · 17/10/2025

New #NeurIPS2025 paper: how should we evaluate machine learning models without a large, labeled dataset? We introduce Semi-Supervised Model Evaluation (SSME), which uses labeled and unlabeled data to estimate performance! We find SSME is far more accurate than standard methods.

1217

Divya Shanmugam @dmshanmugam.bsky.social · 15/10/2025

thank you, gabriel!! glad i've gotten to learn so much about maps from you :')

020

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

thank you, Erica 🥹 so glad we got to work together this year!

010

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

thank you, Emma!!! likewise, i'm so grateful for our collaborations over the years!

010

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

thank you, Kenny!!! that's so nice of you to say.

000

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

for an up-to-date review of recent and ongoing work, you can learn more at dmshanmugam.github.io :)

dmshanmugam.github.io

Divya Shanmugam

personal website

000

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

I'll be at INFORMS, ML4H, and NeurIPS later this year, presenting on recent work related to challenges of imperfect data & models in healthcare (and what we can do about them) -- more on these pieces of work soon!!

100

Divya Shanmugam @dmshanmugam.bsky.social · 14/10/2025

I am on the job market this year! My research advances methods for reliable machine learning from real-world data, with a focus on healthcare. Happy to chat if this is of interest to you or your department/team.

22812

Reposted by Divya Shanmugam

Gabriel Agostini @gsagostini.bsky.social · 03/09/2025

Are you a researcher using computational methods to understand cities? @mfranchi.bsky.social @jennahgosciak.bsky.social and I organize an EAAMO Bridges working group on Urban Data Science and we are looking for new members! Fill the interest form on our page: urban-data-science-eaamo.github.io

urban-data-science-eaamo.github.io

Urban Data Science & Equitable Cities | EAAMO Bridges

EAAMO Bridges Urban Data Science & Equitable Cities working group: biweekly talks, paper studies, and workshops on computational urban data analysis to explore and address inequities.

188

Reposted by Divya Shanmugam

Monica Agrawal @monicaagrawal.bsky.social · 15/07/2025

Excited to be at #ICML2025 to present our paper on 'pragmatic misalignment' in (deployed!) RAG systems: narrowly "accurate" responses that can be profoundly misinterpreted by readers. It's especially dangerous for consequential domains like medicine! arxiv.org/pdf/2502.14898

A person searching for risks of surgery. A traditional search engine would surface websites that would likely include both pros and cons of the surgery. However, RAG results only excerpt the cons.

0132

Reposted by Divya Shanmugam

Serena Booth @reniebird.bsky.social · 14/07/2025

I'll be presenting a position paper about consumer protection and AI in the US at ICML. I have a surprisingly optimistic take: our legal structures are stronger than I anticipated when I went to work on this issue in Congress. Is everything broken rn? Yes. Will it stay broken? That's on us.

A poster for the paper "Position: Strong Consumer Protection is an Inalienable Defense for AI Safety in the United States"

1195

Reposted by Divya Shanmugam

Allison Koenecke @allisonkoe.bsky.social · 22/06/2025

🎉Excited to present our paper tomorrow at @facct.bsky.social, “Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinese”, with @brucelyu17.bsky.social, Jiebo Luo and Jian Kang, revealing 🤖 LLM performance disparities. 📄 Link: arxiv.org/abs/2505.22645

"Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinese" Abstract:

While the capabilities of Large Language Models (LLMs) have been studied in both Simplified and Traditional Chinese, it is yet unclear whether LLMs exhibit differential performance when prompted in these two variants of written Chinese. This understanding is critical, as disparities in the quality of LLM responses can perpetuate representational harms by ignoring the different cultural contexts underlying Simplified versus Traditional Chinese, and can exacerbate downstream harms in LLM-facilitated decision-making in domains such as education or hiring. To investigate potential LLM performance disparities, we design two benchmark tasks that reflect real-world scenarios: regional term choice (prompting the LLM to name a described item which is referred to differently in Mainland China and Taiwan), and regional name choice (prompting the LLM to choose who to hire from a list of names in both Simplified and Traditional Chinese). For both tasks, we audit the performance of 11 leading commercial LLM services and open-sourced models -- spanning those primarily trained on English, Simplified Chinese, or Traditional Chinese. Our analyses indicate that biases in LLM responses are dependent on both the task and prompting language: while most LLMs disproportionately favored Simplified Chinese responses in the regional term choice task, they surprisingly favored Traditional Chinese names in the regional name choice task. We find that these disparities may arise from differences in training data representation, written character preferences, and tokenization of Simplified and Traditional Chinese. These findings highlight the need for further analysis of LLM biases; as such, we provide an open-sourced benchmark dataset to foster reproducible evaluations of future LLM behavior across Chinese language variants (this https URL).

Figure showing that three different LLMs (GPT-4o, Qwen-1.5, and Taiwan-LLM) may answer a prompt about pineapples differently when asked in Simplified Chinese vs. Traditional Chinese.

Figure showing that LLMs disproportionately answer questions about regional-specific terms (like the word for "pineapple," which differs in Simplified and Traditional Chinese) correctly when prompted in Simplified Chinese as opposed to Traditional Chinese.

Figure showing that LLMs have high variance of adhering to prompt instructions, favoring Traditional Chinese names over Simplified Chinese names in a benchmark task regarding hiring.

1174

Reposted by Divya Shanmugam

Shaily @shaily99.bsky.social · 10/06/2025

🖋️ Curious how writing differs across (research) cultures? 🚩 Tired of “cultural” evals that don't consult people? We engaged with interdisciplinary researchers to identify & measure ✨cultural norms✨in scientific writing, and show that❗LLMs flatten them❗ 📜 arxiv.org/abs/2506.00784 [1/11]

An overview of the work “Research Borderlands: Analysing Writing Across Research Cultures” by Shaily Bhatt, Tal August, and Maria Antoniak. The overview describes that We survey and interview interdisciplinary researchers (§3) to develop a framework of writing norms that vary across research cultures (§4) and operationalise them using computational metrics (§5). We then use this evaluation suite for two large-scale quantitative analyses: (a) surfacing variations in writing across 11 communities (§6); (b) evaluating the cultural competence of LLMs when adapting writing from one community to another (§7).

17130

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

it brings me tremendous joy you noticed!!!

000

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

Last but not least, thanks to Helen Lu, @swamiviv1, and John Guttag, my wonderful collaborators on this work! One of my last from the PhD 🥹

010

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

Empirically, TTA reduces prediction set sizes by 10-14% on average, with larger improvements for (1) classes with the largest prediction set sizes and (2) stronger coverage guarantees.

120

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

We also present a new finding on TTA that explains its value to conformal scores: it promotes the true class to be more likely even when it is predicted to be unlikely, which is valuable for conformal scores that rely on orderings over predicted probabilities (e.g. APS, RAPS)!

120

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

We show that test-time augmentation (TTA)—a classic vision technique—is a simple and surprisingly effective way to shrink sets while maintaining coverage. TTA aggregates predictions over transformations of an input (a neat way to create an ensemble out of a single classifier!)

120

Divya Shanmugam @dmshanmugam.bsky.social · 14/06/2025

New work 🎉: conformal classifiers return sets of classes for each example, with a probabilistic guarantee the true class is included. But these sets can be too large to be useful. In our #CVPR2025 paper, we propose a method to make them more compact without sacrificing coverage.

A gif explaining the value of test-time augmentation to conformal classification. The video begins with an illustration of TTA reducing the size of the predicted set of classes for a dog image, and goes on to explain that this is because TTA promotes the true class's predicted probability to be higher, even when it's predicted to be unlikely.

3226

Divya Shanmugam @dmshanmugam.bsky.social · 12/06/2025

One place you can find me is Poster Session 4 on Saturday, at 5PM, presenting recent work on how you can use test-time augmentation to reduce the size of sets produced by conformal prediction. Full paper thread coming shortly :) here is the paper in the meantime: arxiv.org/abs/2505.22764

arxiv.org

Test-time augmentation improves efficiency in conformal prediction

A conformal classifier produces a set of predicted classes and provides a probabilistic guarantee that the set includes the true class. Unfortunately, it is often the case that conformal classifiers p...

020

Divya Shanmugam @dmshanmugam.bsky.social · 12/06/2025

I’m in Nashville this week for #CVPR2025! DM me to chat about conformal prediction, test-time adaptation, or model reliability. Excited to see new work and to catch up with friends old and new!!

140