Atharv Sonwane

CS PhD @ Cornell
Prev. RF @ Microsoft Research India, CS @ BITS Goa

AI, PL, Robotics

[Daman Arora @amuseddaman]

1/ RL for improving task success has made LLMs remarkably capable. But what happens when the model fails? Turns out, training for graceful failure can be surprisingly tricky!

Very excited to share our new paper: Training Language Models to Fail Gracefully

🧵

[Atharv Sonwane @twm_as]

@VarshitaKolipa1 makes sense. thanks for sharing

[Jatin Prakash @bicycleman15]

What to do when you have zero rewards during RL?

We benchmarked RL baselines on a simple star-graph task where they underperform in zero reward scenarios.

Turns out, a dead simple data-centric intervention of just adding easy samples of the task helps unlock RL training! 👇

[Atharv Sonwane @twm_as]

@JulesJacobs5 great photo! is this around Ithaca?

[Aditya Kanade @adityakanade0]

Releasing MASAI: Modular Architecture for Software-engineering AI agents

Modularity helps achieve highest resolution rate (28.33%) at <$2 avg. cost/issue on SWE-bench Lite

@amuseddaman @twm_as @nalin_wadhwa @AbhavM @SaitejaUtpala @rkbairi @naga86

https://t.co/gidV4SIOsp

[Atharv Sonwane @twm_as]

@_carlosejimenez Really interesting. Will the logs be shared for SWE-agent with GPT4o?

[Atharv Sonwane @twm_as]

@paulgauthier Super interesting work. Quick question: while evaluating on SWE-bench, does aider make use of the "hints text" provided in the dataset?

[Aditya Kanade @adityakanade0]

Nice to see our work (CORE) on using LLMs to resolve code quality issues flagged by static analysis tools like CodeQL (Python) and Sorald (Java) accepted in FSE 2024! Thanks to all the collaborators for the great effort 👍@FSEconf #FSE2024

Pre-print: https://t.co/nzEJRsrNkp

[Atharv Sonwane @twm_as]

Work done at @MSFTResearch India. Learned a lot working from my amazing collaborators: @adityakanade0 @rkbairi @tengantsuu @SureshIyengar10 @SriramRajamani Vageesh DC, B Ashok and Shashank Shet

[Atharv Sonwane @twm_as]

CodePlan is accepted at FSE 2024!

We frame repository level coding as a planning problem over local edits and demonstrate how it can be solved using LLMs + static analysis.

[Aditya Kanade @adityakanade0]

LLMs are good at localized coding tasks. What if a task spans multiple inter-dependent files? These “repository-level coding tasks” cannot be solved directly using LLMs. We formulate these as a planning problem and design a task-agnostic, neuro-symbolic framework called CodePlan.

[Atharv Sonwane @twm_as]

At #NeurIPS2023 and interested in automating repository level coding with LLMs?

I'll be at our poster today on CodePlan at the Foundation Models for Decision Making Workshop!

Venue: Hall E2 till 5:30 PM

[Aditya Kanade @adityakanade0]

LLMs are good at localized coding tasks. What if a task spans multiple inter-dependent files? These “repository-level coding tasks” cannot be solved directly using LLMs. We formulate these as a planning problem and design a task-agnostic, neuro-symbolic framework called CodePlan.

[Arun Iyer @AIonGradFlow]

Microsoft Research India - who we are. https://t.co/HbunT0Kinj via @YouTube

[Aditya Kanade @adityakanade0]

LLMs are good at localized coding tasks. What if a task spans multiple inter-dependent files? These “repository-level coding tasks” cannot be solved directly using LLMs. We formulate these as a planning problem and design a task-agnostic, neuro-symbolic framework called CodePlan.

[SAiDL @SforAiDL]

We are excited to present to you the third edition of "AI Symposium," in association with @appcair - the AI Research Lab of @BITSPilaniGoa !

[1/6]

[Stats of India @Stats_of_India]

Who exactly is Indian middle class?

• 90% of Indians make less than 25,000 monthly.
• If you're making > 1L a month, you're among the top 3%.

[Nithin Kamath @Nithin0dha]

How large is the Indian market for B2C tech businesses in terms of users who can generate revenue?

Maybe 15 crores max!

Here's why, with Fintech as a reference, since some data is available. I guess it is important to know this, so we can all be rationally optimistic. 1/11

[Fifty Two @FiftyTwoDotIn]

It's time for another 🧵

IIT GIRLS: How women students at IIT Bombay in the 70s were radicalised by the sexism they faced on campus. And went on to change science in India forever.

[Rajaswa Patil @RajaswaPatil]

1/n] @SforAiDL LRG (https://t.co/CHgjbNrmWR) was established a couple years ago by a bunch of undergrad #NLProc enthusiasts from @BITSPilaniGoa. I am glad to share that the members from the group will be presenting **4** published works at @emnlpmeeting this year:

[SAiDL @SforAiDL]

Reminder! The Social session on Gathertown is starting in 15 mins.

Please note that only accepted people would be able to join this event.