Researchers · Interviewed at ICLR 2026

Sara Papi

Fondazione Bruno Kessler

ICLR 2026 in Rio de Janeiro. Why Do Multimodal LLMs Get Worse When You Give Them More Information?

ICLR 2026 · June 5, 2026 · 13 min

ICLR 2026 in Rio de Janeiro. Why Do Multimodal LLMs Get Worse When You Give Them More Information?

Are today's multimodal LLMs really integrating information across speech, video, and text?

At #ICLR 2026, Sara Papi (Fondazione Bruno Kessler) presents MCIF, a new benchmark for Multimodal Cross-Lingual Instruction Following that evaluates 23 state-of-the-art models—including Gemini, Qwen, speech-language models, video-language models, and multimodal LLMs.

More from ICLR 2026

All event interviews →