This year IJCAI-ECAI came to our home turf – Bremen – and Constructor Labs made the most of the workshop days: 2 posters, a contributed talk, and a hands-on role in shaping a European research agenda on trustworthy AI.
1. Poster – Causal Algorithm Design (GenAIK: Generative AI & Knowledge Graphs workshop)
“Multi-Resolution Causal Knowledge-Graph Extraction and Paper–Code Grounding for ML Algorithms”, presented by Eduard-Antonio Zippenfenig (with Andrey Ustyuzhanin; Constructor Labs / Constructor University / IFIM NUS).
Every ML method exists twice — in the paper and in the code — and the 2 copies disagree. Our system renders both into typed causal graphs at three resolutions (micro/meso/macro) and turns the comparison into graph matching instead of manual reading. Highlights: 85–96% hierarchy consistency across three LLM judges; paper–code coverage analysis showing that what papers omit is mostly conceptual, not plumbing (36.5% vs 22.4%); and a constructive pipeline that is 56% cheaper than direct extraction. Project: https://constructorlabs.org/projects/knowledge-discovery/
2. Poster — Friction-Augmented Drifting Models (GLOW: Generalizing from Limited Resources workshop).
“Friction-Augmented Drifting Models for Resource-Efficient Domain Translation” — Arkadii Kazanskii, Tatiana Petrova, Andrey Ustyuzhanin, Konstantin Bagrianskii, Aleksandr Puzikov, Radu State (University of Luxembourg · Constructor University · Constructor Knowledge Labs · IFIM NUS).
A physics-flavoured fix for a newly identified failure mode in drifting models: the repulsive regime behaves as an underdamped oscillator, so we add dissipation — one line of code, no extra parameters, no extra forward passes. Result: DM-matched generation quality at 29× faster training than OFM (6.8 vs 197.5 min) on domain translation.
Talk — ML²B (GeCoIn: Generative Code Intelligence workshop).
“ML²B: Benchmarking LLMs on Cross-Lingual ML Pipeline Generation” — presented by Ekaterina Trofimova (Constructor Labs, Bremen; joint work with HSE University and NUS).
The first multilingual benchmark for end-to-end ML pipeline generation: can code-generating LLMs build working ML pipelines when the task is specified in languages other than English? Benchmark and code: github.com/enaix/ml2b
TRUST-AI workshop — co-authoring a research agenda.
At the 2nd European Workshop on Trustworthy AI (TRUST-AI), Andrey Ustyuzhanin joined the working group on “Construction of New Knowledge” — researchers from LUMC/NeLL, SnT Luxembourg, SINTEF, University of Piraeus/NTUA, NKUA Athens and Togo AI Lab — and presented the group’s report: 3 key challenges for knowledge construction in the age of LLMs (shared conceptualization, deriving genuinely new knowledge, and the “vanishing human” in data pipelines), with a descriptive-then-normative response plan and an open invitation to team up around the upcoming HORIZON-RAISE calls on Automated Scientific Discovery.
