Pattern Recognition Lab

Laya, Jev and the Return of the Discriminative Model

For three years the word "model" in public conversation has meant a machine that writes. The interesting release of the past week writes nothing at all. On 18 September a single developer in India published Laya, a 421-million-parameter model that takes a piece of text and a set of typed questions and returns typed answers with probabilities, in one forward pass, under Apache 2.0. Within a day the announcement collected more than twelve hundred points on Hacker News and roughly three hundred comments, most of them arguing about who had the idea first. The quarrel is the least interesting part of the story, and I will come to it, but it is worth stating the useful part up front: this is a classifier, it is very fast, and for a large class of production problems that is exactly what was needed.

Laya, Jev and the Return of the Discriminative Model For three years the word “model” in public conversation has meant a machine that writes.

Congratulations to all the award winners at RIME 2026 at MICCAI in Strasbourg — the 2nd Workshop on Reconstruction and Imaging Motion Estimation!

A wonderful workshop, and really good times with Thomas Küstner and Lina Felsner. Thank you to NVIDIA for sponsoring the Best Paper and Best Poster Awards, and to everyone who submitted, presented and stayed for the discussion.

Great to have FAU Erlangen-Nürnberg among the organising institutions, alongside the University of Tübingen, Imperial College London and TUM.

Congratulations to Haobo Song — first place in Task One and fourth place in Task Two of the PENGWIN Challenge 2026 at MICCAI in Strasbourg!

PENGWIN is the Peripelvic Fracture Segmentation and Reduction Planning Challenge. The winning entry in Task One, on automatic peripelvic fracture segmentation, was "Anatomy-guided Core/Contact Instance Reconstruction with Selective Affinity Refinement".

Full team: Haobo Song, Daiqi Liu, Chang Liu, He Lyu, Chengze Ye and myself — a great collaboration between FAU Erlangen-Nürnberg and Siemens Healthineers.

Very proud of this one. Well done, everyone!

Glad to meet Julia Schnabel at MICCAI in Strasbourg, where we talked about the 2nd Bavarian Conference on AI in Medicine.

It takes place 30 November to 1 December at the Bayerische Akademie der Wissenschaften in Munich, organised jointly by FAU Erlangen-Nürnberg, LMU München, TU München, Helmholtz München, Würzburg, Bayreuth and Regensburg.

The abstract deadline is 5 October — just over a week away. If you work on AI in medicine in or near Bavaria, this is the room to be in. Don't forget to submit!

https://events.hifis.net/event/3582/

A real pleasure to welcome Prof. Huan Song and the team from West China Hospital to Erlangen.

We spent time at Siemens Healthineers headquarters and at Friedrich-Alexander-Universität Erlangen-Nürnberg, including a look behind the scenes — always the part that makes the engineering feel real.

Thank you, Prof. Song, for making the long journey, and thanks to everyone on both sides who made the visit happen. This is how collaborations actually start: in person, in front of the machines, with enough time to talk properly.

Before Mythos There Was SATAN: What a 1995 Panic Does and Does Not Teach Us

When Anthropic's Mythos, a model that finds software vulnerabilities on its own, became the story of the summer, the reaction had a familiar sound to anyone who remembers the internet of the mid-1990s. In a June essay for the *Pessimists Archive* newsletter, Louis Anslow supplies the memory: in April 1995 a piece of software called SATAN, the Security Administrator Tool for Analyzing Networks, was released to the public, and the press treated it as the end of the online world. His essay is a short, well-sourced piece of history with a clear opinion attached, and it deserves both a retelling and an argument. The retelling follows; the argument is mine.

Before Mythos There Was SATAN: What a 1995 Panic Does and Does Not Teach Us When Anthropic’s Mythos, a model that finds software vulnerabilities on its own, became the story of the summer, the reaction had a familiar sound to anyone who remembers the internet of the mid-1990s.

From Sparse Triples to Dense Wiki – Why Agents Need a Better Memory

Large language models have become remarkably good at answering questions, but when a virtual assistant has to plan over days, keep track of a user’s preferences, or stitch together evidence from many documents, the thin “knowledge-graph” structures that many systems use start to look like a skeletal outline. Those graphs store facts as (head, relation, tail) triples – a format that is easy for machines to index but strips away the rich flow of natural-language text. The 2026 arXiv pre-print *WFM: Wiki Foundation Model for Complex Agentic Reasoning* argues that the next step for intelligent agents is to replace those over-sparse graphs with a denser “LLM-Wiki” representation: a web of Markdown pages and passage-level text linked together by multi-layered edges. The paper’s authors set out to test whether a foundation model trained directly on such a hybrid graph can boost both multi-hop question answering and long-term memory retrieval for autonomous agents.

From Sparse Triples to Dense Wiki – Why Agents Need a Better Memory Large language models have become remarkably good at answering questions, but when a virtual assistant has to plan over days, keep track of a user’s preferences, or stitch together evidence from many documents, the thin “knowledge-graph” structures that many systems use start to look like a sk...

Crossing the Digital Divide: A New Benchmark for Multi-Device GUI Agents

In everyday life we rarely stay glued to a single screen. A typical morning might begin with a photo taken on a smartphone, an immediate upload to the cloud, a quick edit on a laptop, and a final save back to the phone’s gallery. Such cross-device workflows have long been a cliché in user-experience studies, yet most autonomous “GUI agents” – software that watches a screen and clicks, types, or scrolls on our behalf – are tested only on isolated, single-device tasks. The result is an overly rosy picture of readiness: an agent can close a dialog on a Windows PC, but it may stumble the moment it has to remember that a file it just produced on Android lives in a different folder on a Linux laptop.

Crossing the Digital Divide: A New Benchmark for Multi-Device GUI Agents In everyday life we rarely stay glued to a single screen.

Self-Improving Agents, Surveyed: A Clean Formalism and a Modest Reality

When I. J. Good wrote in 1966 that “the first ultraintelligent machine is the last invention that man need ever make,” he imagined a system that redesigns itself faster than any human could follow. Six decades later, that idea has found a practical foothold in foundation models, large language and vision-language networks that read and write text, images and code, and that speak the same natural language we use to instruct them. They have become the cognitive core of a new generation of autonomous agents, and the question in many research meetings is no longer whether a model can pass a benchmark but whether an agent can persistently improve itself as it meets new tasks, errors and opportunities.

Self-Improving Agents, Surveyed: A Clean Formalism and a Modest Reality When I.

Theory Is All You Need: A Strong Case Against Prediction That Underrates the Machines

Artificial intelligence has mastered chess and Go and passes professional exams, and the headlines suggest that the mind of a computer is converging on the mind of a human, especially now that large language models write essays indistinguishable from a graduate student’s. The 2024 article **“Theory Is All You Need: AI, Human Cognition, and Causal Reasoning”** by Teppo Felin of Utah State University and Matthias Holweg of the University of Oxford, published open access in *Strategy Science*, asks a different question: is the data-driven, backward-looking prediction that powers today’s AI the same kind of intelligence that lets humans generate genuinely new ideas?

Theory Is All You Need: A Strong Case Against Prediction That Underrates the Machines Artificial intelligence has mastered chess and Go and passes professional exams, and the headlines suggest that the mind of a computer is converging on the mind of a human, especially now that large language models write essays indistinguishable from a graduate student’s.