Harshit Joshi (@harshitj_

Harshit Joshi

1,041 posts

Harshit Joshi

@harshitj__

CS phd @StanfordNLP, @StanfordOVAL | prev: @MSFTResearch | LLM systems for knowledge access, discovery and curation

Stanford, CA

Joined May 2012

Harshit Joshi
@harshitj__
Apr 18, 2023
I am beyond thrilled to share that I will attend @Stanford Computer Science department for my Ph.D. this fall!!!! I want to tame the powers of LLMs (Neural models) using symbolic methods to build robust neuro-symbolic systems.
Harshit Joshi
@harshitj__
Jan 12, 2025
Replying to @deedydas
Hi @deedydas, I dont think you can call STORM a Google Deep Research clone when the former was out in early 2024 and the latter just came out in Dec 2024. If anything, Google Deep Research is a STORM clone 😉
Harshit Joshi
@harshitj__
Jul 9, 2024
🚨 New preprint 🚨 Current task-oriented AI assistants often struggle with reliability, control, and integrating knowledge. Even OpenAI's customer service bot uses a dialog tree-based assistant. We set out to solve this!
Harshit Joshi
@harshitj__
Jun 25, 2023
Just realized that my works have received more than 150 citations now! definitely, not the best metric to measure impact/success, but at least it makes me feel I am doing something right.
Harshit Joshi
@harshitj__
Feb 3, 2023
I think that now I can surely say that I am going to grad school!!! 🤩
Harshit Joshi
@harshitj__
Dec 21, 2023
🔥 FLAME has been accepted at AAAI 2024! We show that you don't always need a LARGE language model if you have good data and domain insights which can be leveraged to carefully design the tokenizer and pretraining tasks. Especially if you want to serve in a low resource setting.
Harshit Joshi
@harshitj__
Apr 5, 2024
This guy (@aryaman2020) just keeps coming up with bangers!! Forget LoRA, LoReFT is in the town.
Aran Komatsuzaki
@arankomatsuzaki
Apr 5, 2024
ReFT: Representation Finetuning for Language Models 10x-50x more parameter-efficient than prior state-of-the-art parameter-efficient fine-tuning methods repo: github.com/stanfordnlp/py… abs: arxiv.org/abs/2404.03592
Harshit Joshi
@harshitj__
Sep 7, 2022
I am thrilled to share that our work on "Neurosymbolic Repair for Low-Code Formula Languages" has been accepted at OOPSLA 2022 (@splashcon), a premier programming languages conference! This is my first work at @Microsoft with the super amazing PROSE team (@ProseMsft)!!!
Harshit Joshi
@harshitj__
Dec 11, 2023
note to future self: good research code is different from good production/engineering code
Harshit Joshi
@harshitj__
Feb 6, 2023
I am excited to share that we are presenting our work: "Repair Is Nearly Generation: Multilingual Program Repair with LLMs" at #AAAI23 this week! My ever-charming advisor, @josepablocam, will present the poster on 12th Feb in person. If you are attending, please visit! (1/n)
Harshit Joshi
@harshitj__
Jul 27, 2025
flying to Vienna 🇦🇹 for ACL to present Genie Worksheets (Monday 11am)! come and say hi if you want to talk about how to create controllable and reliable application layers on top of LLMs, knowledge discovery and curation, or just wanna hang
Harshit Joshi
@harshitj__
Nov 2, 2023
A reviewer complained about not citing an arxiv work from Oct '23, which came out 3 months after we submitted our work lol 🥰🥰
Harshit Joshi
@harshitj__
Dec 10, 2021
Career Update: It's been a month since I joined @MSFTResearch as a Research Fellow. We are working on blending Symbolic and Neural approaches for program synthesis at @ProseMsft and integrating them with the diverse range of Microsoft products!
Harshit Joshi
@harshitj__
Jul 18, 2021
sometimes i think it's getting difficult to get into nlp research (few low hanging fruits) with great strides we have made through transformer models and amazing tools such as @huggingface, stanza by @stanfordnlp. but then I see the topic modeling done by @Twitter. lol