Twitxy Saved

Rushil Bhat

Verified account Doubleword

@rushilbhat

Joined April 2022
403 Following 23 Followers 27 Posts

Latest posts 5

Rushil Bhat Verified account @rushilbhat
Came away from this amazed at how much latent ability models possess. DeepSeek-V4-Pro solves 51% of SWE-bench Pro with one attempt and 71% with 64 attempts. Efficient inference turns parallel sampling into a lever worth pulling, especially on tasks you can verify.
Meryem Arik Verified account @MeryemArik9 Sep 22
1/ We think parallel test-time scaling is one of the most underexplored ways to improve model quality

We spend a lot of time on efficiency for long-horizon agentic inference

We showed that a swarm of DeepSeek-V4-Pro agents lifts the SWE-Bench-Pro score by ~20 points, beating the cost-performance trade-off of the leading frontier models🧵
View quoted post on X
Rushil Bhat Verified account @rushilbhat
While writing a minimal TP setup, I noticed the loss was dropping sluggishly and the gradient norms were unstable vs the single-GPU baseline. Turns out the way I parallelised CE loss ended up erasing gradients.

New post on what went wrong + how to fix it.
Rushil Bhat Verified account @rushilbhat
Inspired by @jacobrintamaki's posts, I've sketched out some ideas about what could help maximise the project's impact.

@eladgil @shreyanj98
Jacob Rintamaki Verified account @jacobrintamaki Feb 16, 2025
Made another post after having a few conversations with the great @shreyanj98 on evals for this project. It made me believe that to have great evals (if you're paranoid) you need to go all the way back to how you source your PDFs. link below
View quoted post on X
Rushil Bhat Verified account @rushilbhat
The recent Kimi paper gives really nice intuition on why MCTS failed for R1.

In AlphaGo the value function was necessary to steer the system towards a strategy that was difficult to exploit—optimal for adversarial play.

For language models straight-line drives of optimal actions seem actually counterproductive. ‘Mistakes’ allow for reflection and backtracking, increasing inference compute by extending the CoT and triggering more forward passes.