The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs

2026-07-10

AI Summary

This study addresses the underperformance of vision-language models (VLMs) on basic object counting tasks despite their strong capabilities in other multimodal settings. Investigating whether this failure stems from missing internal knowledge or a misalignment between representations and outputs, the authors employ nonlinear activation probing, SVCCA subspace alignment, and causal intervention analyses. Their findings reveal that counting information is indeed encoded in internal representations but fails to be properly routed through the output pathway. To bridge this gap, they propose a detector-guided, inference-time self-correction method that leverages a re-prompting mechanism without updating model parameters. This approach yields substantial improvements—up to 15.6 percentage points in accuracy across multiple datasets—demonstrating that activation probes not only serve as practical tools but also effectively expose the discrepancy between a model’s internal knowledge and its final predictions.

Abstract

Despite strong performance on many multimodal tasks, vision-language models (VLMs) still struggle with basic object counting. We investigate whether this reflects missing internal knowledge or a gap between internal representations and verbalized outputs. Training simple probes on activations from four VLMs across five counting datasets reveals that nonlinear probes can reliably detect counting errors, suggesting that VLMs often encode the correct count even when they output the wrong answer. SVCCA analysis shows that probes trained on ground-truth counts and probes trained on model outputs occupy a partially shared activation subspace but read out along misaligned directions. We further validate our findings using a causal steering intervention, proving that strengthening the direction of count-identified probes does improve model counting performance. Motivated by this result, we propose a detector-guided self-correction method that selectively re-prompts the model only when an internal error detector predicts failure. This simple inference-time intervention improves counting accuracy by up to 15.6 absolute percentage points, without any parameter updates. Our results establish activation-based error probing as both a practical tool for improving VLM counting and a mechanistic lens on the gap between internal knowledge and model outputs.

Problem

• vision-language models
• object counting
• counting errors
• internal representations
• output misalignment

Innovation

• error probing
• activation alignment
• detector-guided self-correction
• vision-language models
• counting accuracy