Publications
Selected Publications and Prior Work
This page currently highlights selected publications by the PI. Work produced after the launch of LUMI Lab will be marked separately as LUMI Lab research.
2026
How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure
A benchmark study of how instruction-hierarchy compliance deteriorates as multiple conflicts accumulate.
Language Shapes Instruction Hierarchy Compliance in Multilingual LLMs
A study of how the language a prompt is written in affects whether multilingual models respect instruction priority.
Don't Judge Code by Its Cover: Exploring Biases in LLM Judges for Code Evaluation
A systematic study of whether superficial code variations bias LLM-based code evaluation.
A Benchmark for the Generation and Evaluation of Scientific Architecture Diagrams in HTML/SVG
A benchmark for generating and evaluating scientific architecture diagrams in HTML and SVG.
2025
Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation
An analysis of visual biases that distort large vision–language model judgments.
Can You Trick the Grader? Adversarial Persuasion of LLM Judges
A study of how persuasive language can shift LLM-judge decisions on correctness-based evaluation tasks.
LLMs can be easily Confused by Instructional Distractions
A study of inputs that resemble instructions and divert models from the user's intended task.
Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the Effect of Epistemic Markers on LLM-based Evaluation
An evaluation of whether uncertainty expressions systematically influence LLM judges.
An asterisk after an author name indicates equal contribution.