Foundation models

Large pretrained models are now used where a rephrased question or a slightly different camera image can change their answer. We study such failures and how to bound them: adversarial methods that elicit hallucinations with queries that mean the same as the original, validation of vision-language and vision-language-action models over entire continuous regions of image perturbations rather than over samples, what guardrail certificates actually guarantee, and adversarial but intent-preserving user simulation for evaluating multi-turn agents.

Task-specific models, such as image classifiers and object detectors, are under Robust and verifiable machine learning.

Clean versus adversarial accuracy of LLMs under semantically equivalent attacks (COLM 2026)
Clean versus adversarial accuracy of LLMs under semantically equivalent attacks (COLM 2026)

Active researchers

2026

NeurIPS WS 2026

Verifying the Verifier: What a Guardrail Certificate Pays For

S. Varghese, A. Lomuscio

OpenReview
BibTeX
@inproceedings{SAIL:2026-verifying-the-verifier-what-a-guardrail,
  author     = {S. Varghese and A. Lomuscio},
  title      = {{Verifying the Verifier: What a Guardrail Certificate Pays For}},
  booktitle  = {NeurIPS26 Workshop Verify-Agents. Sydney, Australia. Poster},
  year       = {2026},
  url        = {https://openreview.net/forum?id=6RIDw4SNGo}
}
NeurIPS WS 2026

Intent-Preserving Adversarial User Simulation for Multi-Turn Agent Evaluation

A. V. Sharma, B. Formento, A. Lomuscio

OpenReview
BibTeX
@inproceedings{SAIL:2026-intent-preserving-adversarial-user-simulation-for,
  author     = {A. V. Sharma and B. Formento and A. Lomuscio},
  title      = {{Intent-Preserving Adversarial User Simulation for Multi-Turn Agent Evaluation}},
  booktitle  = {NeurIPS26 Workshop on Grounded User Simulation for Model Evaluation and Training: Diversity, Fidelity, and Validity (UserSim). Paris, France. Spotlight presentation},
  year       = {2026},
  url        = {https://openreview.net/forum?id=Ml0KxdEl5Q}
}
arXiv 2026

Validating, Not Sampling: Region-Level Robustness of Vision-Language and Vision-Language-Action Models

B. Aron, C. Brix, B. Brückner, Y. Zhang, P. Kouvaros, A. Lomuscio

arXiv
BibTeX
@article{DBLP:journals/corr/abs-2609-22293,
  author     = {Bogdan Aron and Christopher Brix and Benedikt Brückner and Yanghao Zhang and Panagiotis Kouvaros and Alessio Lomuscio},
  title      = {{Validating, Not Sampling: Region-Level Robustness of Vision-Language and Vision-Language-Action Models}},
  journal    = {CoRR},
  volume     = {abs/2609.22293},
  year       = {2026},
  eprinttype = {arXiv},
  eprint     = {2609.22293},
  doi        = {10.48550/ARXIV.2609.22293}
}
COLM 2026

Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks

A. V. Sharma, B. Formento, A. Lomuscio

arXiv
BibTeX
@article{DBLP:journals/corr/abs-2608-04286,
  author     = {Atri Vivek Sharma and Brian Formento and Alessio Lomuscio},
  title      = {{Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks}},
  journal    = {CoRR},
  volume     = {abs/2608.04286},
  year       = {2026},
  eprinttype = {arXiv},
  eprint     = {2608.04286},
  doi        = {10.48550/ARXIV.2608.04286}
}

Other areas