Sumit Mishra

Sumit Mishra

@ossm.me

Reinforcement Learning Researcher Website: ossm.me

8 followers5 following21 posts

Sumit Mishra
@ossm.me

I implemented Dynamic Programming in the FrozenLake-v1 GridWorld, directly following the concepts from the book by Richard S. Sutton and Andrew G. Barto. gif where: is_slippery=False;

Sumit Mishra
@ossm.me

This was a good exercise in translating the mathematical formulation of Dynamic Programming into working code and validating it experimentally. Next, I’ll work on Monte Carlo methods, followed by implementing and reproducing research papers.

Sumit Mishra
@ossm.me

Building a Tabular RL project from scratch. Starting with Dynamic Programming --> Policy Evaluation --> Policy Iteration --> Value Iteration, then moving into Monte Carlo and TD. I’m using Gymnasium environments to run and test my algorithms. After that, I’ll create my own GridWorld environment

Sumit Mishra
@ossm.me

I’m using Sutton & Barto’s book (chapter 3, 4, 5, 6) and trying to implement the algorithms myself without using any built-in RL libraries. This is my first project where I’m directly implementing algorithms from a book and I want to get better at this so I can reproduce research papers someday.

Sumit Mishra
@ossm.me

ICLR 2027 has 50k+ abstracts (vs. 19.5k valid submissions in 2026), with a 20 paper author cap and a new 1 paper limit for submissions where no author is an eligible reciprocal reviewer.

Sumit Mishra
@ossm.me

Authors on 3+ papers must review at least 6 papers, and ICLR says the rapid growth in submissions has begun to strain the review system.

Sumit Mishra
@ossm.me

With decisions due Dec 16, the scale of the process raises a serious question: how well can peer review distinguish strong work at this volume?

Sumit Mishra
@ossm.me

Over the next year, I’m going to put my full focus into building projects and making open-source contributions. My main areas of focus will be reinforcement learning and applied mathematics, especially at the intersection of the two.

View full thread
Sumit Mishra
@ossm.me

For everything I work on, I’ll write a detailed article documenting what I learned and how I implemented it. You can read these articles and try the implementations yourself at ossm.me

Sumit Mishra
@ossm.me

I’ll also share daily updates here about what I’m working on. If I come across a project or research paper that I find particularly interesting, I’ll also try to make a YouTube video about it.

Sumit Mishra
@ossm.me

Over the next year, I’m going to put my full focus into building projects and making open-source contributions. My main areas of focus will be reinforcement learning and applied mathematics, especially at the intersection of the two.

Sumit Mishra
@ossm.me

I really enjoy implementing research papers, and I want to become better in it, so this will be one of my main priorities. I’ll start with simpler papers or selected chapters from books, implement the concepts practically through code and gradually work my way toward deeper and more advanced topics.

Sumit Mishra reposted

Shubhendu Trivedi
@shubhendu.bsky.social

Why some journalist? Just imagine our scientists, who are puzzling over why the Fields Medalist trade union represents such anti-progress Luddites, explaining it.

Quantian@quantian.bsky.social

It would have been funny if Hodge was the first Milennium problem to be solved because imagine some journalist trying to explain that in the New York Times. Probably a majority of working mathematicians couldn’t even understand the problem statement!

Sumit Mishra
@ossm.me

AI for science should not be judged primarily by whether it can summarize papers or generate plausible hypotheses. A stronger test is whether it can identify a useful question, propose an experiment or derivation, interpret the evidence, and revise its hypothesis when the evidence disagrees.

Sumit Mishra
@ossm.me

Random thing I learned today: AI is being used for weather forecasting, mathematical research, robotics, protein design and a bunch of other things I wouldn't have expected a few years ago. Sometimes I think we focus way too much on chatbots when the more interesting stuff is happening elsewhere.

Sumit Mishra
@ossm.me

Interesting research direction: LLMs and Reinforcement Learning may complement each other rather than compete. LLMs can handle: → planning → decomposition → high-level reasoning RL can handle: → interaction → action optimization → sequential control A hybrid architecture could combine both.

Sumit Mishra reposted

PyTorch
@pytorch.org

IBM Research uses vLLM at the core of its RITS platform, providing shared access to model inference and tuning across its research community. Our case study outlines how it supports 1,300+ users and 100+ models. What this could mean for enterprise AI infrastructure: https://bit.ly/4vYx66J

bit.lyCan IBM's RITS Platform and vLLM Reset the Bar for Enterprise AI Access?IBM's RITS Platform integrates vLLM to democratize enterprise AI access. Explore how centralized model inferencing reshapes organizational AI strategy

You are up to date.