New method clarifies what multimodal AI sees when generating image descriptions

Researchers have developed a new approach to make multimodal large language models more interpretable by improving how we trace which visual information the systems use when generating text. The challenge has been that while these models blend visual and textual information together internally, current methods for explaining their choices treat each image token independently—producing blurry, fragmented attribution maps that don't reveal the true decision-making process.

The core problem#

Vision-language models like LLaVA, Qwen2-VL, and InternVL excel at understanding images and generating captions, but their internal reasoning remains opaque. Recent attribution techniques project hidden states from image tokens into the word space to explain what each token contributes to the output. The trouble is that this approach creates a mismatch: the model processes image tokens in context with their neighbors, but the attribution method reads out each one as if it existed in isolation. Additionally, the generated text itself can interfere—earlier words in the caption influence later hidden states, muddying the attribution maps.

ERCR: combining two techniques#

The new framework, called ERCR (Evidence Recomposition and Predictive Context Residualization), addresses both problems with two complementary strategies.

Evidence Recomposition (ER) aggregates visual evidence across multiple grids with different ways of mapping tokens to image regions, then combines these perspectives to reduce fragmentation. Rather than relying on a single token-to-region assignment, this approach gathers the signal from multiple views.

Predictive Context Residualization (PCR) estimates how much the preceding text tokens contribute to explaining the generated word, then subtracts this estimated effect from the attribution map. This isolates the truly visual evidence from the text-token noise.

Results and implications#

Across standard benchmarks (COCO Caption, GranDf, and OpenPSG), ERCR improved visual attribution quality on multiple model families. For instance, on the Qwen2-VL-2B model, the TAM F1-IoU metric—which measures alignment between attributed regions and ground-truth object labels—rose from 39.10 to 44.45 on COCO Caption and from 30.83 to 37.20 on GranDf. These gains suggest that researchers can now more reliably see which image regions actually influence model outputs, making it easier to audit multimodal systems for biases and errors.