Johannes Hoffart

SAP-RPT-1 is here: The foundation model for tabular data.
Predict payment delays, ticket severity, churn and more with high accuracy and minimal setup.

Try it now:
* Playground: rpt.cloud.sap
* Get open source: huggingface.co/SAP/sap-rpt-...
* Read the paper: arxiv.org/abs/2506.10707

A new player enters the arena of Foundation Models on Tabular Data: www.limix.ai - novel methods for pre-training and data generation that look highly relevant. Their evaluation on selected datasets is showing strong performance. Exciting times, looking forward to further in depth comparisons!

At #VLDB2025 London I joined a panel on Neural Relational Data. My Take: LLMs solve some data management tasks, but the next wave is Foundation Models on Relational Data and Semantically Linked Tables. More on this and further trends in #AI and #DataManagement - www.hoffart.ai/vldb-2025-ai...

Senior/Principal Applied Research Scientist (f/m/d): Foundation Models on Linked Business Data Senior/Principal Applied Research Scientist (f/m/d): Foundation Models on Linked Business Data

Our team developing Foundation Models on Tables & Linked Business Data is looking for a new Senior Applied Research Scientist! Excited about pushing the frontier in foundation models on tabular data? Want to have business impact and academic visibility?

Look no further: jobs.sap.com/job/Walldorf...

5 Minute Papers on AI for the Planet AI is more than just chatbots! Learn about how AI can be used to protect biodiversity, fight climate change, and just better understand our planet through 5-minute explainers covering academic papers ...

For the past 3 years, I've taught a course on Machine Learning for Climate Change to undergrads. At times, people have asked if the course lectures could be made available online. While I can't offer that, I have decided to start making "5 Minute Papers on AI for the Planet" videos. Hope its useful!

Can you train a performant language model using only openly licensed text?

We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2

Building software on top of Large Language Models I presented a three hour workshop at PyCon US yesterday titled Building software on top of Large Language Models. The goal of the workshop was to give participants everything they …

Here's the full workshop handout plus annotated slides from "Building software on top of Large Language Models", a three hour tutorial I presented yesterday at PyCon US #PyConUS simonwillison.net/2025/May/15/...

I asked "on the other platform" what were the most important improvements to the original 2017 transformer.

That was quite popular and here is a synthesis of the responses:

This was helpful.

Also worth noting that Bluesky remains a very fraught place for AI discussions for a variety of reasons, good & bad, but with the impact of keeping a lot of the most relevant AI news, paper discussions & biggest names on X

That might change, but it hasn’t yet. Still posting, tho.

AI assisted search-based research actually works now For the past two and a half years the feature I’ve most wanted from LLMs is the ability to take on search-based research tasks on my behalf. We saw the …

It's been a couple of years since GPT-4 powered Bing, but with the various Deep Research products and now o3/o4-mini I'm ready to say that AI assisted search-based research actually works now simonwillison.net/2025/Apr/21/...

Very long time ago, ACM SIGMOD 2012. Submitted my thesis outline and got very good feedback both during the session and afterwards during an individual lunch with a very senior colleague. Would recommend!

The State of LLM Reasoning Models Part 1: Inference-Time Compute Scaling Methods

I just shared a new article, "The State of Reasoning Models", where I am exploring 12 new research articles on improving the reasoning capabilities of LLMs (all published after the release of DeepSeek R1): magazine.sebastianraschka.com/p/state-of-l...

Happy reading!

I shared a controversial take the other day at an event and I decided to write it down in a longer format: I’m afraid AI won't give us a "compressed 21st century"

Here: thomwolf.io/blog/scienti...

It's an extension of this interview discussion from the AI summit: youtu.be/AxBd3G0lFLs?...

When using LLM-as-a-judge, practitioners often use greedy decoding to get the most likely judgment. But we found that deriving a score from the judgment distribution (like taking the mean) works better!
❌LLM-as-a-judge with greedy decoding
😎Using the distribution of the judge’s labels

Discover European cities ✈️ while building your career! Check out the ELLIS PhD/Postdoc Program's 2025 Winter & Summer School Schedule! Dive deep into cutting-edge #AI research, learn from top researchers & connect with peers across Europe. Learn more: bit.ly/42iow66 #PhD #machinelearning

Our first release of 2025: 𝙨𝙢𝙤𝙡𝙖𝙜𝙚𝙣𝙩𝙨, 𝘁𝗵𝗲 𝘀𝗶𝗺𝗽𝗹𝗲𝘀𝘁 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘀𝘆𝘀𝘁𝗲𝗺𝘀!

💥 Main logic in ~1000 LoC
🧑‍💻 Agent writes its actions in code! LLMs are much better at writing code than current standard of writing JSON => higher perf
🌍 Any LLM support (h/t LiteLLM)
🛡️ Secure code exec (h/t E2B)

Could not be more proud of our crew! Kudos to @marcospinaci.bsky.social Marek Polewczyk Markus Kohler Sam Thelin @tj-klein.bsky.social Clemens Biehl @margaridacosta15.bsky.social Margarida Costa André Sreš Jonas Kolk - tremendous achievement!

SALT: Sales Autocompletion Linked Business Tables Dataset Foundation models, particularly those that incorporate Transformer architectures, have demonstrated exceptional performance in domains such as natural language processing and image processing....

▶️ Open source, multi-table data set containing millions of sales orders sourced from real, production linked business data:

Paper 📜 "SALT: Sales Autocompletion Linked Business Tables Dataset" openreview.net/forum?id=UZb...
Data 💻 github.com/SAP-samples/...

PORTAL: Scalable Tabular Foundation Models via Content-Specific... Self-supervised learning on tabular data seeks to apply advances from natural language and image domains to the diverse domain of tables. However, current techniques often struggle with integrating...

▶️ Foundation model trained on tabular data, including open weights and code

Paper 📜 "PORTAL: Scalable Tabular Foundation Models via Content-Specific Tokenization" openreview.net/forum?id=TSZ...
Model & Weights 💻 github.com/SAP-samples/...

Have a look at our work on foundation models on tabular data, published today at #TRL @ #NeurIPS2024:

📜 PORTAL, an open weight and code foundation model trained on tabular data, and
📜 SALT, a real business data set containing millions of sales orders across multiple tables.

Further details 👇

Gemini 2.0 Flash: An outstanding multi-modal LLM with a sci-fi streaming mode Huge announcment from Google this morning: Introducing Gemini 2.0: our new AI model for the agentic era. There’s a ton of stuff in there (including updates on Project Astra and …

Wrote up my initial impressions of the new Google Gemini 2.0 Flash model - it's really good, and the streaming mode (where you can stream video and audio to it and get audio streamed right back) is pure science-fiction simonwillison.net/2024/Dec/11/...

The 3rd Table Representation Learning (TRL) workshop at NeurIPS 2024 is approaching soon ✨

Join us Saturday 14 Dec from 8:30AM for an amazing program and discussions about all things neural models + tabular data (table-representation-learning.github.io ).

Not in Vancouver? Join online neurips.cc 😎

SAP Knowledge Graph - SAP Jobs Find SAP Knowledge Graph at SAP

We are growing the team building the SAP Knowledge Graph and are #hiring AI & Data Scientists, Data Engineers, Knowledge Engineers and Applied Research Scientists in Germany (Berlin, Walldorf) and India (Bangalore): jobs.sap.com/search/?crea...

Let's take GenAI to the next level with #KG!

Tired of saturated benchmarks? Want scope for a significant leap in capabilities?

🔥 Introducing BALROG: a Benchmark for Agentic LLM and VLM Reasoning On Games!

BALROG is a challenging benchmark for LLM agentic capabilities, designed to stay relevant for years to come.

1/🧵

What are data spaces and what do they do? Learn more about data spaces: what they are, what they do and what’s next.

Great blog post from @odihq.bsky.social @esimperl.bsky.social on the current development state of #dataspaces in Europe.

theodi.org/news-and-eve...

How AutoML Creates New Opportunities for Europe - Frank Hutter // CyberValley Podcast #5 YouTube video by Cyber Valley

Tabular DL and AutoML podcast just dropped. For sure watching this

youtu.be/3qpQ-sMRafE

Let me surface this again now that this place is more lively: Come join us at SAP in the US or Germany for a PhD Summer Internship in 2025 in Foundation Models on Structured Data, Table Representation Learning, LLMs and Knowledge Graphs! #MLInternships

Added some more folks to the Open Source AI Starter Pack:

go.bsky.app/N8yVZdW

AI@HPI Conference logo

I am chairing the
AI@HPI Conference: Responsible AI

December 3-4 in Potsdam (Berlin metropolitan area)

Discussing AI with regard to bias, elections/society, trustworthiness, copyright, the EU AI Act, and best practices.

Registration:
hpi.de/en/ai-hpi-co...

Please spread the word!

Hi 👋 We're glad to be here on @bsky.app and looking forward to engaging in this community. But first, learn a little more about us...

#ELLISforEurope #AI #ML #CrossBorderCollab #PhD