Twitxy Saved

Eshaan Nichani

@EshaanNichani

postdoc @ microsoft research nyc · phd from princeton ece · developing the theoretical & scientific foundations of ml & ai
eshaannichani.com Joined April 2021
320 Following 893 Followers 101 Posts

Latest posts 5

Eshaan Nichani @EshaanNichani
We're presenting this paper today at NeurIPS! Drop by the 11am poster session (poster #4016) to learn more!
Eshaan Nichani @EshaanNichani May 5, 2025
Excited to announce a new paper with Yunwei Ren, Denny Wu, @jasondeanlee!

We prove a neural scaling law in the SGD learning of extensive width two-layer neural networks.

arxiv.org/abs/2504.19983

🧵below (1/10)
View quoted post on X
0 replies 4 reposts 23 likes 3.3K views View on X
Eshaan Nichani @EshaanNichani
Transformers can efficiently *express* compositional/multi-step reasoning tasks. But what about *learnability*?

Our new paper shows that learnability via GD is computationally hard, unless models are trained on a curriculum/mixture of easy-to-hard data!

See🧵below for more!
Zixuan Wang Verified account @zzZixuanWang May 30, 2025
LLMs can solve complex tasks that require combining multiple reasoning steps. But when are such capabilities learnable via gradient-based training?

In our new COLT 2025 paper, we show that easy-to-hard data is necessary and sufficient!

arxiv.org/abs/2505.23683

🧵 below (1/10)
View quoted post on X
1 replies 2 reposts 34 likes 3.3K views View on X