VLM Chain-of-Thought Doesn't Reliably Track Visual Evidence, EMNLP Paper Finds

oanacamb · x · 2026-09-30

Key points

A paper accepted at Findings of EMNLP 2026 adapts counterfactual tests for CoT faithfulness to visual inputs, introducing vCT and vCCT.
The authors benchmark eight recent open-source VLMs on two datasets using image edits that remove objects.
Finding: CoTs do not reliably track the visual evidence driving predictions — a removed object may be omitted even when its removal causes a large prediction shift, and mentioned when the shift is small.
Predict-then-Explain explanations align better with perturbation-induced probability shifts than pre-answer CoTs; binary vCT scores are often nearly saturated.
A reconstruction control confirms object removal induces larger shifts than editing alone. The team also releases the Counter-SNLI-VE dataset.

More from Models