3 September 2026
Three papers accepted at MUWS@ACM Multimedia 2026
We are pleased to announce that CP will be present at ACM Multimedia 2026 https://2026.acmmm.org/ in Rio de Janeiro, Brazil 🗓️10–14 November and its associated workshops!
3 papers accepted at the 5th International Workshop on Multimodal Human Understanding for the Web and Social Media (MUWS@ACM Multimedia 2026):
A Multimedia Perspective on the Study of Cognitive Biases in Large Language Models
by Markus Schedl, Shahed Masoudian, Antonela Tommasel
In a nutshell: Enhancing research that considers text-only LLMs, we discuss whether cognitive biases also emerge in multimodal large language models (MLLMs). Our position paper provides three major contributions: 1) We review cognitive biases identified in psychology whose study in (M)LLMs requires text and image modalities; 2) We formulate corresponding research hypotheses and provide preliminary evidence for the identified biases; 3) We present challenges we envision for the study of cognitive biases in MLLMs.
When Model Outputs Become Social Evidence: Challenges for LLM-Mediated Web and Social Media Analysis by Antonela Tommasel, Markus Schedl
In a nutshell: This position paper argues that LLM-mediated Web and social media analysis is never just a matter of sending content to a model and reading off the result. Model-generated labels, scores, summaries, and rationales become meaningful evidence only through the pipeline of choices that produces, evaluates, and interprets them. The paper calls for treating these outputs as fallible, context-dependent evidence, and for making LLM-based analysis more transparent, cautious, and accountable.
Who Judges the Frame? Auditing Multimodal LLM Judges for News Framing Across Event-Level Perspectives by Antonela Tommasel, Markus Schedl
In a nutshell: This paper audits multimodal LLMs as judges of news framing. Using left-, center-, and right-oriented articles covering the same events, we test whether framing assessments change across text, image, text-image, and metadata conditions. The key finding is that modality matters most: text-based and multimodal judgments are fairly stable, while image-only judgments differ substantially and are more sensitive to metadata. The takeaway is simple: LLM-based framing scores are not context-free measurements; they depend on how the content is shown to the model.
Congratulations to the authors! A huge thank you to all the co-authors and partner institutions for the collaboration! We look forward to the academic exchange at the famous town of Copacabana! 🏖️✨
