A new benchmark shows that conflicting scene text degrades MLLM geolocation and shifts predictions toward the injected location in every tested model.
Priority: street/road signs > city/district names > store names with location cues > other text
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SIGNPOST-Bench: Benchmarking Text-Vision Conflict Resolution in Multimodal Large Language Models
A new benchmark shows that conflicting scene text degrades MLLM geolocation and shifts predictions toward the injected location in every tested model.