VLMs encode correct counts but fail to output them
Vision-language models consistently fail at object counting despite encoding the correct information internally, according to new mechanistic analysis. Researchers used activation probes and causal steering to show that VLMs possess accurate count representations but output wrong answers due to misaligned readout directions between ground-truth and predicted activations. This finding reframes a core VLM weakness as a decoding problem rather than a knowledge gap, opening pathways for targeted interventions that could improve reliability on basic visual reasoning tasks without full retraining.
Modelwire context
ExplainerThe practical implication buried in the methodology is that causal steering corrections can be applied at inference time without retraining, meaning the fix is potentially cheap to deploy on already-shipped models rather than requiring a new training run.
This connects directly to the framing in 'Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI' from the same day, which argued that current AI systems are constrained by their representational readout mechanisms rather than their underlying knowledge. The counting paper is almost a concrete case study of that thesis: the model knows the answer but cannot surface it through its default output pathway. Both papers point toward the same architectural frontier, which is the gap between what a model encodes and what it can actually express. That framing matters because it shifts where researchers should be looking for reliability improvements, away from data scaling and toward the interface between internal representations and generation.
Watch whether any VLM benchmark suite, particularly MMStar or CountBench, begins tracking probe-corrected versus standard inference scores separately. If that split becomes standard practice within the next two benchmark cycles, it signals the field has accepted the decoding-versus-knowledge distinction as a real and measurable axis.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsVision-language models · SVCCA · Activation probes · Causal steering
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.