Dipendra Misra

RL and LLM researcher interested in building an open-source AI ecosystem. Previously, @DbrxMosaicAI, @MSFTResearch, @Cornell.

All views are my own.

[Dipendra Misra @DipendraMisra]

1/n Data is crucial in post-training of LLMs. But what if you could repeatedly train on in-distribution data at no extra cost? This is what deployment data allows you to do. In our NeurIPS paper, we show how to post-train LLMs using user-edits logs – a common deployment data.

[Dipendra Misra @DipendraMisra]

This month I am saying goodbye to an amazing journey at Databricks! Huge thanks to @jefrankle & @matei_zaharia for bringing me onto the rocket ship 🚀 (and for the delicious farewell cake)

Had so much fun through various releases, & coffee discussions.

Next steps coming soon :)

[Dipendra Misra @DipendraMisra]

4/n Follows our prior work on learning from user-edits:

- Our NeurIPS '24 paper on an agentic approach. https://t.co/GliJAhEYIS

- Databricks blog on an enterprise application using user-edits to fine-tune LLM for coding. https://t.co/hpZ2r5M9qe

[Dipendra Misra @DipendraMisra]

1/n Data is crucial in post-training of LLMs. But what if you could repeatedly train on in-distribution data at no extra cost? This is what deployment data allows you to do. In our NeurIPS paper, we show how to post-train LLMs using user-edits logs – a common deployment data.

[Dipendra Misra @DipendraMisra]

3/n We derive sub-optimality bounds for algorithms that use these different feedback. Empirically, we show that combining these feedback provides better performance. There is still a lot to try here, both in algorithm development and benchmark creation.

[Dipendra Misra @DipendraMisra]

We are hiring interns for Summer 2026!

[Matei Zaharia @matei_zaharia]

My team is hiring AI research interns for summer 2026 at Databricks! Join us to learn about AI use cases at thousands of companies, and contribute to making it easier for anyone to build specialized AI agents and models for difficult tasks.

[Michael Bendersky @bemikelive]

Excellent work from @DbrxMosaicAI achieving single-model SoTA on text2sql BIRD benchmark using RL with verifiable rewards and Test-time Adaptive Optimization.

[Dipendra Misra @DipendraMisra]

Our result used neither additional training data nor closed-source LLMs. This simple learning recipe can be applied to many other real-world enterprise tasks.

And while not the main point, our result continues to be SOTA in the single-model BIRD leaderboard.

[Dipendra Misra @DipendraMisra]

Our paper showing that with a general-purpose RLVR recipe, we can get SOTA on the BIRD benchmark is out:

https://t.co/z0sbcysvCC

Our hybrid approach performs offline RL and online RL to fine-tune a 32B model. This, along with self-consistency, was sufficient to get us to SOTA.

[Matei Zaharia @matei_zaharia]

Prompt optimization is becoming a powerful technique for improving AI that can even beat SFT! Here are some of our research results with GEPA at Databricks, in difficult Agent Bricks info extraction tasks. We can match the best models at 90x lower cost, or improve them by ~6%.

[Dipendra Misra @DipendraMisra]

@aldopacchiano @ggaonlp You can check out other work on improving LLMs using user-edits:

1. Fine-tuning LLMs using edit feedback: https://t.co/0eQhtwvHLh…

2. A prompt-engineering approach: NeurIPS 2024 paper. https://t.co/GliJAhEYIS

[Dipendra Misra @DipendraMisra]

Our paper on fine-tuning LLMs using user-edits got accepted at NeurIPS 2025! Joint work with @aldopacchiano , @ggaonlp , and Ta-Chung. Paper forthcoming.

Training LLMs using deployment data (e.g., user-edits) provides a never-ending way to improve and personalize LLMs!

[Jonathan Frankle @jefrankle]

Not that I have a favorite recent project, but... 🧵 LLM judges are the popular way to evaluate generative models. But they have drawbacks. They're:
* Generative, so slow and expensive.
* Nondeterministic.
* Uncalibrated. They don't know how uncertain they are.
Meet PGRM!

[Dipendra Misra @DipendraMisra]

Agents that can learn from language feedback will become better with time. In particular, natural language feedback captures a lot more signal than scalar reward and is natural to us humans. Our research team's work on ALHF shows how to do this in practice!

[Matei Zaharia @matei_zaharia]

Really excited about ALHF, new work from our research team that lets users give natural language feedback to agents and optimizes them for it. It sort of upends the traditional supervision paradigm where you get a scalar reward, and it makes AI more customizable for non-experts.

[Dipendra Misra @DipendraMisra]

We now have the best single-model performance on the popular BIRD benchmark. However, more importantly, we accomplished this using our general-purpose RL recipe that is rolling out to customers in our Agent Bricks product.
https://t.co/S5s5h4gBmN

[Jonathan Frankle @jefrankle]

RLVR and test-time compute are a powerful combo for enterprises, so much so that @databricks now leads overall BIRD single-model leaderboard. This isn't about BIRD, though. It's an example of what our customers are accomplishing in their domains with our RL recipe in Agent Bricks

[Dipendra Misra @DipendraMisra]

If you are excited about the new OpenAI open-weight LLMs, then come and try them on Databricks!

[Databricks @databricks]

We’re excited to partner with @OpenAI to launch their new open source models natively on Databricks!

gpt-oss sets a new standard of quality for open language models, supporting advanced reasoning with the transparency, flexibility and control enterprises need. Running on Databricks, the gpt-oss models connect securely to your data and scale with built-in governance, and expand what you can build and do with GenAI.

Try both the 20B and 120B today in the Mosaic AI Playground. https://t.co/QMoGmxQqWj

[Jonathan Frankle @jefrankle]

More details in the blog: https://t.co/ZqmcWh6l2K

This work was led by @DipendraMisra with contributions from many others. If you're interested in taking this for a spin yourself, sign up here: https://t.co/hxl9R2I6Wn

[Dipendra Misra @DipendraMisra]

Further, unlike the previous best result, we were able to do this without relying on proprietary LLMs or additional datasets besides Bird. Perhaps most excitingly, we were able to do this in a short amount of time, meaning these benefits may disseminate quickly to other domains.

[Dipendra Misra @DipendraMisra]

At @databricks , we are using RLVR to solve real-world problems. This requires both great science & engineering! As an example of the power of our training stack, we were able to reach the top of the popular Bird single-model single-call leaderboard in our first attempt!

[Jonathan Frankle @jefrankle]

RLVR isn't just for math and coding! At @databricks, it's impacting products and users across domains. One example: SQL Q&A. We hit the top of the BIRD single-model single-generation leaderboard with our standard TAO+RLVR recipe - the one rolling out in our Agent Bricks product.

[Dipendra Misra @DipendraMisra]

AI for data science is going to be a big productivity boost across the industry. It is an important area for us, and we are hiring strong researchers and engineers in this space.

[Matei Zaharia @matei_zaharia]

Lakeflow Designer and our new Genie Deep Research really begin to show the power of AI for data science, which I think will be even more impactful than AI for coding. I'm proud that Databricks is leading the charge here.

We're hiring researchers in this area, so reach out!