Language and vision models (VLMs) have demonstrated impressive capabilities in tasks such as image description, answering visual questions, and multimodal reasoning. However, a recurring and puzzling flaw is its poor performance on basic counting tasks: asking how many objects are in an image often leads to incorrect answers, even in simple situations. Recent research reveals that the problem lies not in a lack of internal knowledge of the model, but in a disconnect between what the model knows internally and what it is able to verbalize. This finding opens up new avenues for understanding and correcting these errors, and has profound implications for the development of more reliable AI applications.
The baseline study analyzes four VLM models against five counting datasets. By training linear and nonlinear probes on the models' internal activations, the researchers found that nonlinear probes can detect when the model is going to make a counting error. That is, the model internally encodes the correct number of objects, but that knowledge does not translate correctly into the final answer. This phenomenon is similar to when a student knows the answer but makes a mistake in expressing it. The SVCCA (Singular Vector Canonical Correlation Analysis) analysis shows that probes trained with true values and those trained with model outputs occupy a partially shared activation subspace, but the read directions are misaligned. This suggests that the internal representations of the correct count exist, but they do not align with the representations that the model uses to generate the output text.
The definitive confirmation comes through causal steering interventions: by reinforcing the directions of the probes that identify the correct count, the performance of the model improves significantly. This shows that inner knowledge is accessible and modifiable. Based on this finding, the authors propose a detector-guided self-correction method: the model generates a response, an internal probe predicts whether that answer will be incorrect, and only then is the model asked again differently. This inference-time intervention, without the need to update parameters, improves counting accuracy by up to 15.6 absolute percentage points.
For companies that integrate artificial intelligence into their processes, these conclusions are crucial. A model that knows how to count but doesn't say it well can lead to costly errors in applications such as automated inventory, visual quality control, medical image analysis, or people counting in security environments. Understanding that the flaw is in verbalization and not in cognition allows us to design more effective mitigation strategies than simply training with more data. For example, a double-check system can be implemented: the model generates a response, then an internal detector evaluates confidence in that answer, and if it is low, an alternative reasoning process is used or a human is consulted.
At Q2BSTUDIO, as a software and technology development company, we apply these principles in our artificial intelligence solutions for enterprises. Our team designs systems that not only use pre-trained models, but incorporate layers of internal verification and correction to improve reliability. We work on the development of bespoke applications that integrate AI agents capable of self-assessment and correction, which is especially useful in environments where error has a high cost.
In addition, for these systems to work robustly in production, a solid infrastructure is necessary. That's why we offer AWS and Azure cloud services that enable you to deploy vision and language models with low latency and high availability. Data management and security are also critical; Our cybersecurity solutions ensure that the sensitive data used in these models is protected. And so that companies can make decisions based on these results, we integrate business intelligence services such as power BI that visualize the performance metrics of the models and the errors detected, facilitating continuous improvement.
Ultimately, research on the gap between internal knowledge and verbalization in VLMs is not only relevant to academia, but has immediate practical applications. It allows us to build more transparent, controllable, and accurate AI systems. The next time a VLM fails to count, remember that you probably do know the answer; He just needs us to help him express it. And at Q2BSTUDIO we are prepared to design these aids, either through custom software or by integrating process automation techniques with intelligent agents.
.jpg)
.jpg)
.jpg)
.jpg)