Beyond Perception Errors: Semantic Fixation in Large Vision-Language Models
For researchers and developers of VLMs, this work provides a clean diagnostic for separating perception from rule-mapping failures and demonstrates that semantic fixation is a robust, editable bias in late representations.
The paper identifies and isolates 'semantic fixation' in VLMs—a tendency to default to familiar interpretations even when prompts specify alternative valid mappings. Using a controlled benchmark (VLM-Fix) over abstract games, they show a consistent accuracy gap favoring standard rules, which can be narrowed by neutral alias prompts and partially corrected via late-layer activation steering.
Large vision-language models (VLMs) often rely on familiar semantic priors, but existing evaluations do not cleanly separate perception failures from rule-mapping failures. We study this behavior as semantic fixation: preserving a default interpretation even when the prompt specifies an alternative, equally valid mapping. To isolate this effect, we introduce VLM-Fix, a controlled benchmark over four abstract strategy games that evaluates identical terminal board states under paired standard and inverse rule formulations. Across 14 open and closed VLMs, accuracy consistently favors standard rules, revealing a robust semantic-fixation gap. Prompt interventions support this mechanism: neutral alias prompts substantially narrow the inverse-rule gap, while semantically loaded aliases reopen it. Post-training is strongly rule-aligned: training on one rule improves same-rule transfer but hurts opposite-rule transfer, while joint-rule training improves broader transfer. To test external validity beyond synthetic games, we evaluate analogous defamiliarization interventions on VLMBias and observe the same qualitative pattern. Finally, late-layer activation steering partially recovers degraded performance, indicating that semantic-fixation errors are at least partly editable in late representations. Project page, code, and dataset available at https://maveryn.github.io/vlm-fix/.