At Compute! Paris, @alonsosilva.bsky.social shows why combining both beats either alone, through real applications: extracting fields from unstructured text, classifying at scale for categorical features, and pulling out entities and relations to enrich your data.

Two techniques attack this from different angles: structured outputs, where the model fills a schema through tool calling, and structured generation, which constrains token sampling so the JSON is valid by construction.

Free text in, dataframe columns out. That's the promise of LLM-based extraction, until the model returns JSON that doesn't parse and your batch job dies halfway through.

How to evaluate models on your own use case with accessible open-source tools, and how to run those tests in parallel on standard cloud GPUs, no HPC cluster needed. One early finding: combining features from several models beats any single one.

AlphaEarth, TerraMind, TESSERA, Copernicus-FM: Earth observation now has more foundation models than anyone has time and capacity to test. At Compute! Paris, Gijs van den Dool offers a practical way out of this model fatigue.

Not random, they fall exactly where the meaning lies in the relationships themselves, which way a link runs, wheter it is a cause or association, or a step an expert infers across sentences. Illustrated with geophysics, this is built to generalize to other fields.

Every extracted relation gets traced back to the exact passage that justifies it, so the gap between "found it in the text" and "understood why it's true" becomes visible. The system's failures are particularly interesting.

It's tested against an expert-built benchmark in a hard domain: mass-transport deposits in subsurface geophysics with ambiguous terminology and lots of implicit geological reasoning.

AI can extract scientific facts from a paper. But can it understand the relationships between them? At Compute! Paris, Feryal Batoul Talbi presents OntoGeoRAG, an open-source pipeline that turns scientific PDFs into knowledge graphs.

Every finding, all software behind, even the research field's FAIR rules, all get published similarly, turning independent claims into a connected, reusable graph of evidence, with verified citations. Perfect for giving humans and AI agents solid scientific grounding. Get your ticket!

The talk draws on a year spent replicating 30+ published claims in ecology, biodiversity, and climate science. Recurring failure modes surfaced (biased representation choices, priors that strongly sway results, hidden train/test leakage), each catalogued with how to catch it.

The smallest unit of science to trust isn't a paper, it's a single claim. At Compute! Paris, Anne Fouilloux and Jean Iaquinta explore "nanopublications": small, cryptographically verifiable, machine-readable claims you can cite, link, and trust.

Rows untouched by an edit fall back into the exact same pages as before, remain byte-identical, and deduplicable nicely. Simple examples first: append, insert, delete, and the actual bytes saved on each. Then the real numbers, straight from that 100PB+ platform.

At Compute! Paris, Quentin Lhoest, maintainer of the popular datasets library, introduces Content Defined Chunking to Parquet, ending a page based on the row data itself rather than counting rows. This makes page boundaries, and thus Parquet file byte content, more robust to change.