Researchers investigate why vision-language models (VLMs) struggle with object counting despite encoding correct counts internally. They find that nonlinear probes can detect these errors, revealing a misalignment between internal representations and verbalized outputs.
- Training probes on activations from four VLMs across five datasets shows that models often encode the correct count even when outputting the wrong answer.
- SVCCA analysis reveals that probes trained on ground-truth counts and model outputs occupy a partially shared activation subspace but read out along misaligned directions.
- Causal steering intervention proves that strengthening the direction of count-identified probes improves counting performance.
- A detector-guided self-correction method selectively re-prompts the model only when an internal error detector predicts failure.
This inference-time intervention improves counting accuracy by up to 15.6 absolute percentage points without any parameter updates, establishing activation-based error probing as a practical tool for improving VLMs.