I am Diego Antognini, a Staff AI Researcher and Tech Lead at Google DeepMind, where I work on making Gemini models better at long interactions, from multi-turn conversations to long-horizon agentic tasks. My research focuses on reinforcement learning and context management for these settings. Previously, I designed multi-turn evaluations for text and image generation and, before that, contributed to Gemini pre-training, where I led an effort to improve the model's multimodal capabilities. Work I led and co-led has shipped in Gemini 2.5, 3, and 4, as well as Nano Banana 2.1. I am also a lecturer and module head at the Lucerne University of Applied Sciences (HSLU), where I teach advanced generative AI and deep learning for NLP at the M.Sc. level. Across my academic roles, I have taught 500+ students and supervised 90+ B.Sc. and M.Sc. projects and theses.

Before joining Google DeepMind, I was a research scientist at IBM Research, where I collaborated with the MIT-IBM Watson AI Lab. I worked on aligning large language models (LLMs), including turning multi-turn conversations into SQL queries over massive databases (used in IBM Watsonx.ai), personalizing LLMs to user preferences, automatically augmenting prompts to steer their behavior, and adapting retrieval-augmented LLMs for question answering over scientific literature. I also built efficient NLP models of only a few megabytes with millisecond latency that matched the quality of much larger models, and deployed them in IBM Deep Search.

I hold a Ph.D. in Computer Science from EPFL, where I worked in the Artificial Intelligence Laboratory (LIA) under the supervision of Prof. Boi Faltings. My thesis, "Textual Explanations and Critiques in Recommendation Systems" (available here), developed scalable, data-driven methods to infer high-quality explanations from text through selective rationalization and to make those explanations actionable through critiquing, with applications in conversational recommendation and natural language processing. I also worked on multi-objective recommendation and multi-document summarization, and was a visiting researcher in Prof. Julian McAuley's lab at UC San Diego.

I regularly give talks, for example on my work on efficient NLP at the MIT-IBM Watson AI Lab and at the NLP Meetup in Zürich, where I presented one of my past projects. I have also appeared in national media after taking part in challenges with students, notably winning a $10k prize at the IARPA Geopolitical Forecasting Challenge 2018 (press coverage: EPFL News, 24 Heures, RTS Radio (22:25)).

On this website, I present a selection of my publications and patents, as well as some of the most exciting projects I worked on before my Ph.D. If you have any questions, feel free to contact me.

News

Selected Publications

For the full list, you can consult my Google Scholar profile. Authors marked with * contributed equally.

32) Learning to Learn from Language Feedback with Social Meta-Learning Paper
Jonathan Cook, Diego Antognini, Martin Klissarov, Claudiu Musat, Edward Grefenstette
2026, COLM

TL;DR: A finetuning methodology that trains LLMs to solicit and learn from language feedback by transforming static tasks into interactive student-teacher dialogues. SML-trained models demonstrate an enhanced ability to solve complex problems over multiple turns.
31) Improving Interactive In-Context Learning from Natural Language Feedback Paper
Martin Klissarov*, Jonathan Cook*, Diego Antognini*, Hao Sun, Jingling Li, Natasha Jaques, Claudiu Musat, Edward Grefenstette
2026, Technical Report

TL;DR: A framework that treats interactive in-context learning as a distinct, trainable skill by transforming single-turn tasks into multi-turn "didactic interactions" between a student and a teacher.
30) Position: Introspective Experience from Conversational Environments as a Path to Better Learning Paper
Claudiu Musat*, Jackson Tolins*, Diego Antognini, Jingling Li, Martin Klissarov, Tom Duerig
2026, CoRR

TL;DR: Robust artificial intelligence reasoning is not merely a product of scaling compute, but emerges from internalizing "introspective experiences" derived from high-quality social and conversational interactions.
29) Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities Paper
..., Diego Antognini, ...
2025, Technical Report

TL;DR: We introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models.
28) Trans-LoRA: towards data-free Transferable Parameter Efficient Finetuning Paper
Runqian Wang, Soumya Ghosh, David Cox, Diego Antognini, Aude Oliva, Rogerio Feris, Leonid Karlinsky
2024, NeurIPS

TL;DR: We propose a nearly data-free transfer of LoRA modules between models using synthetic data, eliminating the need for original training data. This approach shows effective transfer across various models and tasks, improving performance in many cases.
27) Paraphrase and Solve: Exploring and Exploiting the Impact of Surface Form on Mathematical Reasoning in Large Language Models Paper
Yue Zhou, Yada Zhu, Diego Antognini, Yoon Kim, Yang Zhang
2024, NAACL

TL;DR: We investigate how the presentation or surface form of a mathematical problem affects its solvability by large language models. Then, we present Self-Consistency-over-Paraphrases (SCoP), which diversifies reasoning paths from specific surface forms.
26) MC Layer Normalization for Calibrated Uncertainty in Deep Learning Paper
Thomas Frick, Diego Antognini, Ioana Giurgiu, Benjamin Grewe, Cristiano Malossi, Rong Zhu, Mattia Rigotti
2024, TMLR

TL;DR: A drop-in replacement for Layer Normalization to endow neural networks with calibrated prediction uncertainty.
25) FlowPilot: An LLM-powered system for enterprise data integration
Enrico Toniato, Abdel Labbi, Katya Mirylenka, Christoph Miksovic Czasch, Thomas Gschwind, Paolo Scotton, Francesco Fusco, Diego Antognini
2023, NeurIPS Demo Talk

TL;DR: A system to generate training datasets and fine-tune LLMs tailored for customers to convert natural language into SQL queries for massive databases in a conversational manner. FlowPilot mitigates errors during both training and inference, and it can also generate Python code and charts.
24) ESG Accountability Made Easy: DocQA at Your Service Paper
Lokesh Mishra, Cesar Berrospi, Kasper Dinkla, Diego Antognini, Francesco Fusco, Benedikt Bothur, Maksym Lysak, Nikolaos Livathinos, Ahmed Nassar, Panagiotis Vagenas, Lucas Morin, Christoph Auer, Michele Dolfi, Peter Staar
2024, AAAI Demo

TL;DR: A question-answering system using large language models and retrieval-augmented generation, designed to generate accurate and relevant answers from a given corpus of domain-specific documents.
23) pNLP-Mixer: an Efficient all-MLP Architecture for Language Paper or Paper
Francesco Fusco, Damian Pascual, Peter Staar, Diego Antognini
2023, ACL

TL;DR: An embedding-free MLP-Mixer model for on-device NLP using a projection layer that relies on MinHash and counting Bloom filters. Our model occupies merely one megabyte and achieves 99% of the performance of mBERT.
22) Extracting Text Representations for Terms and Phrases in Technical Domains Paper or Paper
Francesco Fusco*, Diego Antognini*
2023, ACL

TL;DR: Meaningful word embeddings can be achieved using character-based models that are 5x smaller and 10x faster than BERT-based counterparts and do not suffer from out-of-distribution problems.
21) Assistive Recipe Editing through Critiquing Paper or Paper
Diego Antognini, Shuyang Li, Boi Faltings, Julian McAuley
2023, EACL

TL;DR: A framework for generating recipes and enabling users to edit them using critiques in an iterative manner. The system coherently rewrites recipes to satisfy users' feedback.
20) Towards Workflows for the Use of AI Foundation Models in Visual Inspection Applications Paper
Mattia Rigotti, Diego Antognini, Roy Assaf, Kagan Bakirci, Thomas Frick, Ioana Giurgiu, Klara Janouskova, Filip Janicki, Husam Jubran, Cristiano Malossi, Alexandru Meterez, Florian Scheidegger
2023, Journal (ce/papers)

TL;DR: A set of proof-of-concept workflows using foundation models in visual inspection applications for civil engineering.
19) Unsupervised Term Extraction for Highly Technical Domains Paper or Paper
Francesco Fusco, Peter Staar, Diego Antognini
2022, EMNLP

TL;DR: A fully unsupervised method for term extraction that generalizes across domains. Our setup improves predictive performance and decreases inference latency on both CPUs and GPUs.
18) Active Learning for Imbalanced Civil Infrastructure Data Paper or Paper
Thomas Frick, Diego Antognini, Mattia Rigotti, Ioana Giurgiu, Benjamin Grewe, Cristiano Malossi
2022, ECCV Workshop on Computer Vision for Civil and Infrastructure Engineering (CVCIE)

TL;DR: A method capable of operating on datasets suffering from heavy class imbalance, achieved by replacing the traditional active learning acquisition function with an auxiliary binary discriminator.
17) Textual Explanations and Critiques in Recommendation Systems Paper or Paper
Diego Antognini
2022, EPFL Ph.D. thesis

TL;DR: This dissertation focuses on two fundamental challenges. The first involves generating explanations: inferring high-quality explanations from text documents in a scalable and data-driven manner. The second challenge consists of making explanations actionable, which we refer to as critiquing. This dissertation examines two important applications in natural language processing and recommendation tasks.
16) Positive & Negative Critiquing for VAE-based Recommenders Paper or Paper
Diego Antognini, Boi Faltings
2022, CoRR

TL;DR: Fast negative and positive critiquing generalized for variational autoencoders, resulting in up to a 15% higher success rate compared to state-of-the-art models. The key lies in modeling positive and negative critiques as different modalities and employing a multi-modal VAE with weak supervision.
15) Interlock-Free Multi-Aspect Rationalization for Text Classification Paper or Paper
Shuangqi Li, Diego Antognini, Boi Faltings
2022, CoRR

TL;DR: Addressing the interlocking dynamics of multi-aspect rationalization, utilizing a novel self-supervised contrastive loss and multi-stage training to generate more semantically diverse rationales.
14) Interacting with Explanations through Critiquing (T-RECS) Paper
Diego Antognini, Claudiu Musat, Boi Faltings
2021, IJCAI

TL;DR: How to extract explanations significantly preferred by humans over those produced by state-of-the-art models and how to make them actionable, enabling users to interact with them iteratively to improve the recommendation.
13) Fast Multi-Step Critiquing for VAE-based Recommender Systems (M&Ms-VAE) Paper or Paper Video
Diego Antognini, Boi Faltings
2021, RecSys

TL;DR: Fast multi-step critiquing generalized for variational autoencoders, resulting in speeds up to 26x faster and a success rate 20% higher compared to state-of-the-art models. The key lies in modeling the problem using multi-modal VAE and weak supervision.
12) Multi-Step Critiquing User Interface for Recommender Systems Paper or Paper Video
Diana Petrescu*, Diego Antognini*, Boi Faltings
2021, RecSys Demo

TL;DR: We propose and demonstrate a new way of interacting with recommender systems to help users make decisions and find their ideal items.
11) Rationalization through Concepts (ConRAT) Paper or Paper Video
Diego Antognini, Boi Faltings
2021, ACL Findings

TL;DR: Generalization of MTM: how to extract interpretable multi-faceted concepts (i.e., rationales) for single-task classification problems. It generates concepts that align with human rationalization, and outperforms state-of-the-art methods trained on each aspect label independently.
10) Multi-Dimensional Explanation of Target Variables from Documents (MTM) Paper or Paper Video
Diego Antognini, Claudiu Musat, Boi Faltings
2021, AAAI

TL;DR: One model to extract interpretable, meaningful, and coherent multi-faceted rationales for multi-task text classification problems, which performs better than individual rationalization models.
9) Addressing Fairness in Classification with a Model-Agnostic Multi-Objective Algorithm Paper or Paper Video
Kirtan Padh, Diego Antognini, Emma L. Glaude, Boi Faltings, Claudiu Musat
2021, UAI

TL;DR: A model-agnostic multi-objective architecture that optimizes multiple fairness notions and sensitive attributes using a novel differentiable relaxation that approximates fairness notions through the hyperbolic tangent function.
8) Multi-Gradient Descent for Multi-Objective Recommender Systems Paper or Paper
Nikola Milojkovic, Diego Antognini, Giancarlo Bergamin, Boi Faltings, Claudiu Musat
2020, AAAI Workshop on Interactive and Conversational Recommendation Systems (WICRS)

TL;DR: An efficient stochastic multi-gradient descent approach for multi-objective recommender systems.
7) HotelRec: a Novel Very Large-Scale Hotel Recommendation Dataset Paper or Paper
Diego Antognini, Boi Faltings
2020, LREC

TL;DR: A new dataset with 50 million hotel reviews with meta-attributes, user information, and multi-aspect ratings.
6) Recommending Burgers based on Pizza Preferences: Addressing Data Sparsity with a Product of Experts Paper or Paper
Martin Milenkoski, Diego Antognini, Boi Faltings
2021, RecSys Workshop on Cross-Market Recommendation

TL;DR: We address data sparsity and generate recommendations in domains where there is limited knowledge about the user preferences.
5) Modeling Online Behavior in Recommender Systems: The Importance of Temporal Context Paper or Paper
Milena Filipovic*, Blagoj Mitrevski*, Diego Antognini, Emma L. Glaude, Boi Faltings, Claudiu Musat
2021, RecSys Workshop on Perspectives on the Evaluation of Recommender Systems

TL;DR: Omitting temporal context while evaluating recommender systems leads to false confidence. We propose an evaluation protocol and a model-agnostic training procedure to incorporate temporal context.
4) Momentum-based Gradient Methods in Multi-objective Recommender Systems Paper or Paper
Blagoj Mitrevski*, Milena Filipovic*, Diego Antognini, Emma L. Glaude, Boi Faltings, Claudiu Musat
2021, RecSys Workshop on Multi-Objective Recommender Systems

TL;DR: A coordinated multi-objective optimization method in which each objective is optimized using an algorithm similar to Adam.
3) GameWikiSum: a Novel Large Multi-Document Summarization Dataset Paper or Paper
Diego Antognini, Boi Faltings
2020, LREC

TL;DR: A non-news domain-specific dataset for multi-document summarization that is 100 times larger than commonly used datasets.
2) Learning to Create Sentence Semantic Relation Graphs for Multi-Document Summarization Paper or Paper
Diego Antognini, Boi Faltings
2019, EMNLP Workshop on New Frontiers in Summarization

TL;DR: How to leverage universal and domain-specific sentence embeddings using a graph structure for multi-document summarization.
1) Dataset Construction via Attention for Aspect Term Extraction with Distant Supervision Paper or Paper
Athanasios Giannakopoulos*, Diego Antognini*, Claudiu Musat, Andreea Hossmann, Michael Baeriswyl
2017, ICDM Workshop on Sentiment Elicitation from Natural Text for Information Retrieval and Extraction (SENTIRE)

TL;DR: How to utilize large corpora for improved aspect term extraction using distant supervision.

Patents

Additional applications pending.
5) Layer Normalization for Calibrated Uncertainty in Deep Learning Paper
Thomas Frick, Mattia Rigotti, Diego Antognini, Ioana Giurgiu, Cristiano Malossi
2025
4) Domain-specificity prediction for natural language processing Paper
Diego Antognini*, Francesco Fusco*
2024
3) Determining specificity of text terms in application contexts Paper
Francesco Fusco*, Diego Antognini*
2024
2) Self-supervised term encoding with confidence estimation Paper
Francesco Fusco*, Diego Antognini*
2024
1) Updating window representations of sliding window of text using rolling scheme Paper
Francesco Fusco*, Diego Antognini*
2024

Projects (prior to Ph.D.)

From Relation Extraction to Knowledge Graphs - M.Sc. thesis

My master's thesis at Iprova. A system that extracts concepts from large corpora and builds interactive knowledge graphs to provide invention developers with new insights. View more

NeoBrain - B.Sc. thesis

A research project focused on optimizing the processing of neuronal activity maps using massively parallel technologies. View more

Hurricane

A scalable, decentralized system that aggregates secondary storage devices in a cluster with the aim of supporting parallel scans of data stored across them. View more

Optimized flocking algorithm for e-pucks

I implemented, tested, analyzed and optimized a flocking algorithm for e-pucks. The objective was for the robots to avoid obstacles within the arena while maintaining their collective formation. Done as part of a multidisciplinary team. View more

PokerFace

A complete Texas Hold'em Poker game with artificial intelligence. View more

Starfighter 4K

Shoot 'em up game utilizing motion recognition with Kinect and Wiimotes for spaceship movement, inclination, and shooting. View more

CUDA

Several mini-projects for learning about GPGPU technologies, primarily CUDA. View more

Image classification

Classifier that recognizes the object present in an image using advanced models. The objects could be classified as a horse, airplane, car, or something else. View more

Social Recommendation System

Recommender system for events based on user data and Facebook profiles. View more

Facial recognition among profiles

Detect whether a person is wearing sunglasses using a collection of profile pictures of different individuals. Each person has pictures taken from different head angles, displaying different emotions, and with or without sunglasses. View more

Pattern classification and machine learning project 1

Project on regression and classification using linear models. One dataset is provided for each task without any accompanying information. View more

Recommender System challenge

Third task of the European Semantic Web Conference challenge on Top-N book recommendation (ESWC-14 Challenge). GitHub Report

EPFL-IMDB

A large database movie directory using real data from IMDb. View more

Star2

Planetarium software that displays the current view of the sky at the present location. View more