Our paper “The Trigger Is in the Embedding: Supervision-Free Discovery and Diagnosis of Universal Adversarial Triggers in Fact Verification”, by Michele Luca Contalbo, Matteo Paganelli, Paolo Papotti, and Francesco Guerra, has been accepted as a Main Conference paper at EMNLP 2026, taking place in Budapest, Hungary, from October 24–29, 2026.

The paper introduces FactFlip, a supervision-free approach for discovering universal adversarial triggers in fact verification models. These triggers are words that, when inserted into otherwise unchanged claims, can systematically bias model predictions toward a target class.

Unlike existing gradient-based approaches, FactFlip requires no labeled data and no gradient optimization. It identifies candidate triggers directly from the model’s embedding space by measuring their alignment with the classification weights.

Experiments across six datasets and different encoder- and decoder-based models show that FactFlip can identify effective and compositional triggers while achieving results close to gradient-based methods. The approach also provides a diagnostic tool for investigating systematic model biases.

The code is publicly available on GitHub.

We are pleased to announce that three DTALab contributions have been accepted at CIKM 2026, including two demo papers and one resource paper.

TabgenQA — Demo Track
TabgenQA is an interactive platform for generating non-relational, multi-table numerical question-answering benchmarks and evaluating LLMs on the resulting datasets. The platform supports configurable benchmark generation, instance inspection and editing, and model evaluation.
Authors: Michele Luca Contalbo, Matteo Paganelli, Paolo Papotti, Francesco Guerra
Code: GitHub – TabgenQA

AEQUITAS — Demo Track
AEQUITAS (An Evaluation toolkit for QUantItative and Trading AnalySis) is an interactive environment for stock market prediction and portfolio evaluation, built on top of the FinBench benchmarking framework. It supports the complete evaluation workflow from reusable dataset construction to model execution and portfolio-level result analysis and exposes it through a configurable backend and a Gradio-based graphical interface with LLM-assisted result interpretation.
Authors: Sara Pederzoli, Marta Santacroce, Francesco Guerra, Marco Bergianti, Francesco Del Buono
Code: GitHub – AEQUITAS

CFX-Bench: A Benchmark for Counterfactual Explanations — Resource Track
CFX-Bench provides a benchmarking framework for the systematic evaluation of counterfactual explanation methods across datasets and explainers, including support for LLM-based evaluation.
Authors: Riccardo Benassi, Federico Gualandri, Matteo Paganelli, Gennaro Aloe, Enrico Bagli, Mario Calò Carducci, Maurizio Vincini, Francesco Guerra
Code: GitHub – CFX-Bench

DTALab is taking part in ACM KDD 2026, held in Jeju, Korea, from August 9 to 13, with Sara Pederzoli representing the group.

The DTALab contribution, FinBench: A Benchmarking Framework for Stock Market Prediction and Portfolio Allocation, presents a benchmarking framework for the systematic evaluation of methods for stock-market prediction and portfolio allocation.

The code and benchmark resources are publicly available on GitHub.

Authors: Sara Pederzoli, Marta Santacroce, Francesco Guerra, Marco Bergianti, and Francesco Del Buono.