Bias in Multimodal AI:
Representation, Risk, and Repair

• Invited Speaker: Dr. Kyungreem Han

• Keynote Talk: Prof. Eduard Hovy, Prof. Sang Wook Yi

• Panel Discussion: Representation, Risk, and Repair — global experts from academia and industry

• Contributed Paper Presentations: Selected short and position papers

Morning Session

09:30 AM Opening Remarks

09:45 AM Invited Speaker: Dr. (Assoc. Prof.) Kyungreem Han (Korea Institute of Science and Technology, University of Science and Technology)

Title: Unnecessarily intricate multimodal interpretations: understanding machine bias through lived experiences beyond neuro-representationalism
Abstract: Recent advances in multimodal large language models (MLLMs) will open new opportunities for versatile human-machine collaboration, provided safety and ethical concerns are properly managed. This talk aims to acquaint AI researchers with phenomenological, explainable approaches that rely on the physical entities that machines experience from training to inference, rather than biology- and psychology-based traditional symbolic or a new symbolic (i.e., neuro-representationalism) account. Treating MLLMs as complex systems, the numerical operations within the transformer (e.g., self-attention, feedforward, and cross-attention) are interpreted as dynamical phenomena manifesting non-linear, self-organized, and adaptive properties.

The Cybernetics (derived from the Greek ‘κυβερνητικός’, meaning ‘good at steering’) concept for modern computations was created primarily by neuroscientists—the Ratio Club was founded by the neurophysiologist John Bates (1918-1993) and his colleagues, and then some mathematicians, including Alan Turing (1912–1954), had joined the club. It is no small coincidence: the mechanisms of the physical brain-body can shape important features of human intelligence; thus, design principles of (human intelligence-mimicking) AI naturally follow the mechanisms of the human brain-body, establishing an ‘embodied AI’ foundation for the next generation. However, there are no demons representing external reality; only the power of lived experiences rooted in the modern philosophy developed by Edmund Husserl, Martin Heidegger, Maurice Merleau-Ponty, and others.

This talk presents physics-based models that capture the semantic network structure, self-attention, and cross-attention mechanisms to analyze the dynamics of cross-modal bias in MLLMs. A diagnostic analysis using Qwen2.5-Omni and Gemma 3n is combined with a dynamical multi-agent model interpretation. Some fundamental ideas about the computational principles behind bias generation are also covered. Throughout the talk, the ontologically neutral term “physical entity” is used, as it does not matter for any future AI realizations or interpretation doctrines.

10:30 AM Morning Tea Break (15 mins)

10: 45 AM Accepted Paper Presentations

• Expert Collapse and Compositional Failure in Simple Multimodal MoE

• Cultural Representation Bias and Alignment Divergence in Large Language Models

• Language as a Latent Control: Cross-Lingual Bias Inversion and Modality Gaps in Medical Text-to-Image Generation

• Physics-based phenomenological characterization of cross-modal bias in multimodal models

12:00 PM Lunch Break (1.5hrs)

Afternoon Session

01:30 PM Keynote Talk: Prof. Eduard Hovy (University of Melbourne, Carnegie Mellon University)

Title: Understanding LLM inference for bias removal in multimodal processing
Abstract: Bias is the assignment of an undesired feature or characteristic to a topic of interest. Usually this assignment arises from data asymmetry, namely when the characteristic is associated with the topic more than is typical in the data. Since all (interesting) data is asymmetric, biases will always occur. When a bias is deemed undesirable, you have two options to remove it:
• Prevention: Edit the data to remove the asymmetry. However, this destroys data fidelity. Further, prevention often doesn’t work because machine learning systems are good at finding proxy characteristics that co-occur with the problematic ones and simply uses them instead.
• Repair: Edit the system to remove the discovery or reporting of the bias. Since making the system internally ‘blind’ to asymmetries risks debilitating it in other unforeseen ways, the best alternative is to add a post-hoc filter to the system that recognizes problematic characteristics and removes or alters them before they are reported.  

Ideally, instead of simply filtering out all mentions of the problematic characteristic (even mentions that are not in context problematic), one would identify the machine learning model’s states that are most specific to the problem, and respond to just those with a repair. This avoids ‘repairing’ output that does not really require repair due to its current context, even when it contains characteristics that in other cases would indeed be problematic.  

Identifying such ‘root cause’ of problematic biases in a Large Language/Vision Model (LVLM) is quite difficult because of the distributed and uninterpretable nature of the internal embedding representations.  Since multiple meanings/topics may superpose into a single node, one cannot simply mask out (sets of) nodes or change connection parameter weightings. One has to let the LVLM identify the problematic node patterns itself and respond to them. 

This talk is about finding and repairing the hidden root causes of bias using Steering Vectors. These vectors, developed within mechanistic interpretation approaches to LVLMs, have proven effective at pinpointing and changing outputs at a quite granular level.  

02:15 PM Keynote Talk: Prof. Sang-Wook Yi (Hanyang University)

Title: Ethics as Enabler: Bias, Neutrality and Fairness
Abstract: Ethics is often regarded as perhaps inevitable but still cumbersome companion of scientific and technological research. A widely-held view among scientists and engineers is that ethical considerations are important, but should be moderated as they are hindering equally (or perhaps more) important innovations. I claim that this common view is inadequate at least when we deal with frontier science and technologies such as AI, in particular when we try to solve a critical problem such as how to mitigate bias in multimodal AI. I shall elaborate my claim from two directions; model-dependency of optimization and cost of not-doing. Finally I will suggest that ethics in AI research can and should be ‘enabler’, nurturing innovation in a more comprehensive way. I shall illustrate my point with examples discussing crucially related but distinct concepts such as bias, neutrality, fairness, equality and equity.

03:00 PM Panel Discussion

03:45 PM Awards and Closing Ceremony

04:00 PM Conference Coffee Break

Panel Moderator