Ashutosh Baheti

I’m super excited about the launch of Genie Code! It extends the power of AI coding to agentic data work, answering questions 2-3x more accurately than coding agents and automatically engineering and monitoring high quality pipelines. It’s transformed my own work at Databricks!

We just published OfficeQA Pro - a set of 133 challenging questions from the original OfficeQA benchmark. Even the best frontier agents still struggle on OfficeQA Pro with common issues stemming from errors in parsing, retrieval, and visual reasoning.

Most AI benchmarks test reasoning in isolation. Real enterprise tasks require grounded reasoning: 1️⃣ Find the right documents 2️⃣ Extract the right values 3️⃣ Perform analyses OfficeQA Pro evaluates this end-to-end. Frontier agents still score <50%. 🧵Paper & details below!

This is a great pattern to build your own specialized AI, described in our report. - Generate synthetic data with current version of model - Apply efficient large-batch off-policy RL (OAPL) - Generate harder data with updated model - Produce efficient, generalizable small model

We just released KARL — a knowledge agent trained with reinforcement learning that beats Claude Opus 4.6 and GPT-5.2 on enterprise search, at a fraction of the cost and latency. 🧵

Super excited to talk about KARL! A few of my favorite things about the report: - we show RL is doing more than just sharpening - TTS that works on non-verifiable tasks - multitask RL > multi expert distillation - OAPL working at scale and more!

Going to do a more technical deep dive on our enterprise knowledge agents and how we train them with RL. Overall we found that simple, yet principled off-policy RL works at scale for complex agentic tasks with hundreds of steps of tool use and context management. Here are the key takeaways from our 80 page technical report. (1) RL does not just sharpen base model's distribution. We see test-time scaling improves consistently over the iterations of the RL training. Skills learned during RL transfers to unseen prompts and agent learns to solve prompts where base model has zero accuracy under pass@16. (2) Multi-task RL generalizes really well Simple mixing of training data from multiple tasks works well and allows multi-task RL scale beyond your in-distribution training tasks. We found that multi-task RL just works better than multi-expert distillation. (3) End-to-end RL for tools and context management works best. We skipped mid-training, and directly trained everything end-to-end using RL at scale (2m tokens per gradient computation). Models learned to use vector database tools and context compression at the same time.

Today, we're sharing 🌁 Knowledge Agents from Reinforcement Learning (KARL) 🌁 We trained an agent that excels on challenging grounded reasoning tasks. KARL matches Sonnet 4.5 quality at a fraction of the cost, and with test-time scaling reaches Opus 4.6 levels. This was a fun project that I learned a lot from. Here are a few pieces that resonated with me.

Incredibly proud of the team behind this!! KARL solves a genuinely hard problem and it's only the first of many agents we're building @DbrxMosaicAI 🚀🤖

🚀 Today we’re releasing FlashOptim: better implementations of Adam, SGD, etc, that compute the same updates but save tons of memory. You can use it right now via `pip install flashoptim`. 🚀 arxiv.org/abs/2602.23349 A bunch of cool ideas make this possible: [1/n]

@jefrankle @j_nadan_chang @databricks Why this bias? Why not both off-policy and on-policy!?

Online RL? 🥱 Off-policy RL 🥳 OAPL is what we're using internally at @databricks for some really cool RL results that you'll hear about soon...

Many recent works try to force GRPO to be on-policy by adding things like extra importance weighting, clipping, masking, data deletion, inference engine edits, router replay… But are these actually needed? We push the other direction: make it maximally off-policy and keep it simple! Turns out that nothing is wrong with off-policy. It works well, avoids entropy collapse, and improves test-time scaling (RL doesn't just sharpen base model's distribution). While we only presented dense models in this report for math and coding tasks, similar recipe works for large-scale MoEs in agentic settings. Will share more results very soon!

I now constantly get questions about the SAAS meltdown, role of AI, system of records etc. I don't have an answer to all these. But I do know that we saw an acceleration in our business in Q2, Q3, and now finished the year with accelerating Q4. The question is, why? Short answer: AI. But the underlying reason is subtle. We are growing fast because we are finally removing the biggest bottleneck in data: the technical barrier to entry. For years, if you didn’t know SQL, Python, you were locked out of the value chain. That has changed fundamentally with the 𝐆𝐞𝐧𝐢𝐞 𝐟𝐚𝐦𝐢𝐥𝐲, and it is the "secret sauce" behind our recent momentum: • 𝐆𝐞𝐧𝐢𝐞: Analysts can query data without any SQL. I use this every day myself. • 𝐃𝐚𝐭𝐚 𝐒𝐜𝐢𝐞𝐧𝐜𝐞 𝐆𝐞𝐧𝐢𝐞: Builds end-to-end AI models for you, similar to Cursor for ML on your data. • 𝐃𝐚𝐭𝐚 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫 𝐆𝐞𝐧𝐢𝐞: Write Spark pipelines, does plumbing, troubleshooting. We've been talking about DATA + AI democratization, but generative AI finally enabled it in a way that wasn't possible before. That's why we're seeing a market response. Take 𝐋𝐚𝐤𝐞𝐛𝐚𝐬𝐞 𝐏𝐨𝐬𝐭𝐠𝐫𝐞𝐬. We launched this serverless engine for agents and apps recently. At 8 months into its journey, its revenue is already 2x what our Data Warehouse product was at the same stage. All this taken together, we ended up with the following stats for Q4: 🚀 $5.4B Revenue Run-Rate, growing >65% YoY 🚀 $1.4B AI Revenue Run-Rate 🚀 FCF Positive for the year 🚀 NRR >>140% databricks.com/company/newsro…

Agent memory is a simple and powerful way to do continual learning! With the new MemAlign method from Databricks Research, we can build better LLM judges from examples of human ratings, and they scale with more data. Now in Databricks and @MLflow. databricks.com/blog/memalign-…

Lakebase is GA! We think this is going to make it radically simpler and more reliable to work with online databases. You can instantly branch your DB, take snapshots, roll back to a point in time, or create a copy for offline analysis, whether it's humans or AI doing the work.

Instructed Retriever is a multi-tiered declarative approach for building high quality search agents. It's an example of an "instructed system", which goes beyond prompt tuning and tool calling by passing data among modules which work together to fulfill an information need.

Interested in training RL agents for long horizon enterprise tasks? 🤖👾 Come work with our phenomenal and cracked team! 🚀

Today we’re introducing OfficeQA, a new benchmark grounded in ~89,000 pages of U.S. Treasury Bulletins that reflects the complex, document-heavy tasks enterprises actually face. Unlike existing benchmarks, OfficeQA measures economically valuable, real-world reasoning: parsing dense tables, navigating scanned PDFs, and retrieving facts across decades of documents. Even strong agents reach only ~45% accuracy, showing how far the field has to go. The benchmark is now open to the community, and the Databricks Grounded Reasoning Cup in Spring 2026 will challenge teams to push these capabilities forward. databricks.com/blog/introduci…

Today we’re releasing OfficeQA — a new benchmark for end-to-end grounded reasoning that reflects the real work enterprises need AI agents to do. More details below 👇

@jefrankle Will I be getting this swag? 🤣

Special Databricks swag for the first five people to send me a selfie with Ashu in the Databricks booth at NeurIPS!