Why multimodal evaluation is treacherous
Evaluating a vision-language model is harder than evaluating either a vision model or a language model alone, because the model can be right for the wrong reason. Many benchmark questions are answerable from language priors with the image ignored — "what color is the banana?" — so a model that never looks can score well. A real evaluation must isolate the visual contribution, not the model's grasp of world trivia.
- Probe the language-prior shortcut: re-run with the image removed or replaced by noise; a big score is a red flag the benchmark is not testing vision.
- Counterfactually edit the image (swap the object, change the count) and check the answer follows the pixels, not the prior.
- Audit for contamination — many web-scale corpora already contain popular benchmark images and their answers.
Grounding and the modality gap, revisited
Visual grounding is the most honest probe we have, because it forces the model to commit to a location. A model can bluff a plausible caption, but a bounding box is either on the referent or it is not — there is nowhere to hide. Grounding accuracy correlates with whether the model is genuinely using its cross-modal attention rather than narrating from priors, which is why grounded benchmarks are increasingly the gold standard for perception.
The modality gap from guide 1 is not just a curiosity — it shapes evaluation and capability. Because image and text embeddings live in separated cones, retrieval thresholds tuned on one modality pair transfer poorly, and naive cross-modal arithmetic misbehaves. Researchers now study deliberately closing or exploiting the gap; the punchline is that a model's geometry, not just its accuracy, tells you how it will generalize.
Two vectors and the angle between them, with the dot product equal to the product of their lengths times the cosine of that angle.
Benchmarks and model judges
Open-ended multimodal answers resist exact-match scoring, so the field leans on the LLM-as-a-judge pattern — and increasingly a VLM as judge that sees the image too. It scales beautifully, but inherits every bias of LLM judging: position bias, length bias, and self-preference. A multimodal judge adds a new failure: it may share the same visual blind spots as the model under test, so both miss the same hallucination in unison.
Intersection-over-Union, the objective grounding metric the callout pairs with the judge: a predicted box A is scored against the ground-truth box B by overlap over union.
Multimodal hallucination
The signature failure of VLMs is object hallucination: confidently describing things not in the image — a non-existent person, a wrong color, an invented label on a chart. The roots trace straight back through this track: a strong language prior overrides weak visual evidence (guide 1's modality gap), the instruction data taught a confident style on uncovered cases (guide 3), and the image was too low-resolution to actually read (guide 3's diagnostic). Hallucination here is not a quirk; it is the seam where the modalities fail to bind.
Mitigations stack rather than solve: contrastive decoding that subtracts the no-image prediction to suppress the prior, grounding-aware tuning that rewards pointing before claiming, higher input resolution, and abstention training so the model can say "I can't tell from this image". Each chips at the rate; none closes it.
Visual contrastive decoding made explicit: amplify the image-conditioned logits and subtract the no-image (text-only) prediction ℓ(y∣x), suppressing the language prior that drives object hallucination.
The research frontier
Where is the field heading? Toward any-to-any models where understanding and generation share weights, so generating an image teaches the model to perceive better and vice versa. Toward native high-resolution and long-video perception without exploding token budgets. Toward document and GUI agents that act on screens, not just describe them. And toward grounded, faithful reasoning where every claim is tied to a visible region.