Prasann Singhal

This account is verified. Learn more

@prasann_singhal

2nd-year #NLProc PhD at UC Berkeley working with @sewon__min @JacobSteinhardt , formerly advised by @gregd_nlp

Berkeley, California prasanns.github.io Joined January 2021

837 Following 512 Followers 169 Posts

Latest posts 5

Pinned post

Prasann Singhal @prasann_singhal Sep 28

Could long-context architectures “find all contradictions” in a science literature? Not yet! 🧵

We study a new class of "high-complexity” tasks whose difficulty scales quadratically with corpus size (as opposed to linearly), reversing common LCLM decisions! (block-sparse attention, hybrid models…)

Prasann Singhal @prasann_singhal May 8

I love this paper so much! Results in Figure 3 / 5, and Appendix B.2 are in particular my favorites. "Emergent modularity" is such a cool concept. I think this paper represents a really fresh / cool style of pre-training work which I'm hoping to see a lot more of in coming years!

Ryan Yixiang Wang Verified account @RyanYixiang May 8

MoEs are everywhere in frontier models, and they are deployed as a monolith system.

But many applications only need a narrow slice of capabilities, e.g., math, code, biomedical, etc.

So what if "modularity" is actually the missing opportunity for MoEs?

Today, we're releasing EMO: an end-to-end pretrained MoE where modularity emerges naturally, enabling selective use of experts!

View quoted post on X

Prasann Singhal @prasann_singhal Sep 19, 2024

This is a gigantic / awesome project, it was an honor to help out on. I thought the Section 3 meta-analysis was especially cool (500+ papers *manually* filtered!): extracting actual results from papers, rigorous categorization, detailed outlier analysis... Also section 4.3

Zayne Sprague Verified account @ZayneSprague Sep 19, 2024

To CoT or not to CoT?🤔

300+ experiments with 14 LLMs & systematic meta-analysis of 100+ recent papers

🤯Direct answering is as good as CoT except for math and symbolic reasoning
🤯You don’t need CoT for 95% of MMLU!

CoT mainly helps LLMs track and execute symbolic computation

View quoted post on X

Prasann Singhal @prasann_singhal May 3, 2024

Labeling preferences online for LLM alignment improves DPO vs using static prefs. We show we can use online prefs to train a reward model and label *even more* preferences to train the LLM.

D2PO: discriminator-guided DPO

Work w/ @natolambert @scottniekum @tanyaagoyal @gregd_nlp

Prasann Singhal @prasann_singhal Nov 10, 2023

Super excited to talk about our RLHF work at the awesome
@istecnico & @Unbabel seminar series this Monday!

ELLIS Unit Lisbon @Lisbon_ELLIS Nov 10, 2023

This monday we welcome @prasann_singhal for our @istecnico & @Unbabel seminar series to talk about "A Long Way to Go: Investigating Length Correlations in RLHF". Join us!

📅 13th November, 5pm (WET)

🔗 to join online: tinyurl.com/3uns3pe9

View quoted post on X