Antoine Edy

A. Edy⋆, M. Conti⋆, V. Xing⋆, M-A. Allard, N. Benhamdane, G. Viaud

Preprint, 2026

Self-improving agents that build their own memory by practicing on self-generated tasks, without training tasks or oracle verifiers. On AppWorld, τ²-bench and AutomationBench, strongest method without training tasks and competitive with those that use them; memory banks transfer across model families.

A. Loison⋆, Q. Macé⋆, A. Edy⋆, V. Xing, T. Balough, G. Moreira, B. Liu, M. Faysse et al.

ACL 2026 Main Conference, Oral Presentation

Multimodal RAG benchmark: 26,000 pages, 18,594 queries across 10 domains in 6 languages, 12,000h of human annotation. Work made in collaboration with NVIDIA; ∼500k dataset downloads/year on Hugging Face.

A. Edy⋆, M. Conti⋆, Q. Macé⋆

ECIR 2026 Late Interaction Workshop, Poster

Analysis of length bias in multi-vector scoring and MaxSim similarity distribution on NanoBEIR benchmark.

• First author of DAEDALUS: designed a self-improving agent that builds its own memory from self-generated practice tasks for in-context learning on long-horizon tool-use tasks
• Co-led the creation of ViDoRe V3, establishing the gold standard benchmark for multimodal document retrieval evaluation; contributor to the MTEB ecosystem
• Conducted research on Multimodal Retrieval and late interaction dynamics (ColBERT/ColPali-like models)
• Fine-tuned LLMs for tool calling in agentic RAG contexts; fine-tuned speech-to-text models (Whisper, NVIDIA Parakeet) for domain-specific industrial applications
• Conducted 50+ technical interviews as part of the hiring process for Data Scientist positions

• Studied zero-shot semantic segmentation using multimodal foundation models; improved ZegCLIP baseline with multi-scaling techniques and inner encoder layers
• Research led to Master's thesis: Exploring Zero-Shot Capabilities of Multimodal Foundation Models for Semantic Segmentation, supervised by Prof. Muhammad Awais Rana and Dr. Srinivasa Rao Nandam

• Developed a Python library for tabular model explainability and cluster analysis
• Presented at SpringTech Paris-Saclay 2023 at HEC Paris

• International Excellence Scholarship recipient. Graduated with Highest Honors
• NLP, CV, Audio Processing, Symbolic Knowledge Representation and Logical Reasoning, People-Centered AI
• Master's Thesis: Zero-shot and Few-shot Adaptation of Multimodal Foundation Models

• Specializations: Mathematics, Machine Learning, Probability Theory, Chaos and Fractals
• Project Leader: Unsupervised MCU Classification with Deep Learning for STMicroelectronics

Invited talk, Neurons & Peppers (Datacraft, Paris), 2026

Hugging Face blog, 2026

AI-vidence blog, 2023