Assistant Prof @ MIT. Interested in AI safety, evaluation, auditing, human-AI interaction. Outside of research: 🏊♀️+🚴♀️+🏃♀️
Cambridge, MA
- Tweets 163
- Following 389
- Followers 879
- Likes 864
🚨 New preprint out: arxiv.org/pdf/2610.26645! 🚨 This time, it's with with @ek8terina and @alesmorosini! 1/n 🧵
1
3
9
567
In our experiments, adding CoT-derived groups reduces Brier error by 4.5–15% compared to topic-only multicalibration, while preserving similar topic-level calibration. 5/n
1
1
52
Check our paper out here: arxiv.org/pdf/2610.26645 and do give a follow to my amazing students leading this work. 6/6
48
I've been thinking a lot about AI slop lately, and there are a few thoughts that I wanted to document, not because they are finalized and won’t change, but just in case others find them useful. 1/n
1
2
12
3,017
4. I don’t have a prescribed path ahead. But my path is: implement a “standards and expectations” guide for AI use for my lab, encourage frequent discussions with my students about the “proper” (in my view, at least) use of AI in research, ... 9/n
1
3
220
Chara Podimata retweeted
The discourse that informs our views is just as important as how we vote, but harder to architect well.
We headed to Boston on our summer Moonshots tour to meet the team building a platform for deliberation at national scale. It's designed to assemble better groups, nudge quiet voices, and surface topics that spark quality debate. More soon! laude.org/moonshots/one/deli…
@ArielProcaccia @bakkermichiel @baileyflan @Arfung @lessig @ChrisRytting
2
10
1,061
Also to appear in NeurIPS26! Congrats to my students Nico and Ellery!
Paper 2: Prediction-Powered Inference Across Many Tasks for AI Evaluation and Social Science Research
We develop methods for statistically valid inference from many related tasks, with many cheap proxy labels (via, e.g., an LM) and only a few high-quality labels per task. 🧵
8
1,704
This work will be appearing at NeurIPS26 — congrats to our whole team and especially Nico and @theo_olausson!
Paper 1: Conformal Language Modeling via Posterior Sampling
How can we control hallucination risk in LLM generations without sacrificing the usefulness and coherence of the response?
In our new paper, we use conformal methods to calibrate the generation process itself. 🧵
2
23
2,145
Chara Podimata retweeted
A @nytimes piece by @tiffkhsu spotlights MITHIC-funded research into voters and AI co-led by @MITPoliSci professors @AdamBerinsky and @cstewartiii and @MITSloan professor @charapod. nytimes.com/2026/09/10/busin…
1
3
3
549