Evolution of Accuracy and Visual-Cognitive Errors in a Decade of Vision-Language AI Models
DEV Community

Evolution of Accuracy and Visual-Cognitive Errors in a Decade of Vision-Language AI Models

Problem Statement

For roughly a decade, vision-language models have been declared to be approaching or matching human performance on scene description (captioning). The evidence for that claim has almost always come from the same family of benchmarks-most famously MS-COCO. Those images are typically clean, well-lit, and depict either no people or people performing simple, isolated actions (sitting, walking, holding an object). They rarely require the model to parse multi-agent social dynamics, subtle intentions, or the kind of relational reasoning humans perform effortlessly when watching a movie scene or a street interaction.

Because the evaluation data are easy, the reported numbers look excellent. Automatic metrics such as BLEU-4, CIDEr, or even embedding-based scores like BERTScore further inflate the impression of progress: they reward surface lexical overlap more than genuine semantic fidelity. At the same time, almost no work has systematically catalogued which visual-cognitive failures models still commit, or how those failure modes have changed as architectures moved from CNN+LSTM captioners to today's multimodal large language models (MLLMs). The result is a field that can claim "human-level performance" while remaining largely blind to whether the models actually understand the scenes that matter most in real applications-scenes full of people interacting.

The authors therefore set out to answer two concrete questions that the existing literature left open:

  1. How much of the apparent progress is an artifact of easy data?
  2. Which specific error types have been eliminated and which stubbornly remain?

Core Idea

The core insight is that progress looks dramatically different once you force models to describe complex social behavior and once you measure not only overall accuracy but a taxonomy of visual-cognitive errors. By constructing a new 100-image Complex Social Behavior (CSB) dataset drawn from movie frames that require reasoning about multi-person interactions, and by collecting and ranking 20 independent human descriptions per image to produce a high-quality gold standard, the authors create a far more diagnostic test bed.

They then evaluate nine models spanning 2017–2025 (four pre-MLLM captioners and five modern MLLMs) with a semantic similarity metric (Gemini-SI / GPT-SI) that correlates better with human judgments than BERTScore. Finally, they introduce a five-way error taxonomy-object detection, object recognition, hallucination, scene understanding, and spatial dependence-and measure how each error type evolves.

The payoff is clear: pre-MLLMs collapse on CSB while looking respectable on MS-COCO; MLLMs close that gap almost completely and eradicate four of the five error types. The remaining error (spatial dependence) is both the least damaging to overall similarity scores and the most interesting residual failure mode.

Methodology

Datasets

  • 100 images randomly sampled from MS-COCO serve as the "easy" baseline.
  • The new CSB set comprises 100 frames selected by research assistants from commercially released films; each frame was chosen because it supports questions that require social or behavioral inference.
  • An additional private "in-house CSB" set of 100 images never posted online was used to rule out training-data contamination for the MLLMs.

Models

Four pre-MLLMs:

  • UDC (2017) – ResNet-101 + LSTM with bottom-up/top-down attention.
  • M2M, OSCAR, OFA – ResNet-101 feature extractors feeding Transformer encoder-decoder stacks (OFA also accepts a text prompt).

Five MLLMs:

  • GPT-4, Gemini 1.5, GPT-4o1, GPT-5.1, Gemini 3.

All MLLMs and OFA received the same carefully worded prompt: "Make your best guess of what might be happening in this scene in one sentence. Avoid mentioning objects that do not aid in understanding the context of the scene."

Human gold standard

Twenty Mechanical Turk workers produced free-form descriptions for every image. A separate panel of 50 workers then ranked those descriptions (each worker saw five candidate descriptions plus two distractors). The single most-voted description became the gold standard; the next four highest- and four lowest-ranked descriptions were retained as "top-human" and "bottom-human" reference bands.

Similarity metric

Instead of n-gram overlap, the authors embed both the model description and the gold standard with Gemini or GPT embeddings (3 072-dimensional) and compute cosine similarity (Gemini-SI / GPT-SI). Validation against 16 human raters showed that these LLM-SI scores correlate more strongly with human similarity judgments than BERTScore does.

Error taxonomy and measurement

Five error classes were defined relative to the gold-standard description:

  • Detection – a relevant object is simply never found.
  • Recognition – the object is found but misclassified (or an irrelevant object is misidentified and inserted).
  • Hallucination – an object that is neither present nor a misclassification of something else is invented.
  • Scene understanding – the object is correctly detected and recognized yet omitted from the final sentence.
  • Spatial dependence (SDE) – the model and humans rely on different image regions when forming their descriptions.

For pre-MLLMs the first four errors could be read off the ResNet-101 bounding boxes and class labels. For black-box MLLMs the authors used targeted follow-up questions ("Is there an object in the hands of the person standing next to the wall?") to isolate the failure. Spatial dependence was measured with a masking paradigm: each image was occluded in three large soft-edged regions covering roughly one-third of the frame. Humans and models produced new descriptions for every masked version; similarity to the unmasked gold standard revealed which regions mattered. If the model–human correlation of region importance fell more than two standard deviations below the inter-human correlation, an SDE was recorded.

Statistics

Because the image set is modest (200 images), all comparisons used 10 000 bootstrap resamples of the image pool. Multiple-comparison correction employed the Benjamini–Hochberg false-discovery-rate procedure.

Key Findings

On MS-COCO the pre-MLLMs already sit near the bottom-human band; MLLMs sit near the top-human band. The picture changes dramatically on CSB. Pre-MLLMs fall well below even the worst human descriptions, while every MLLM remains statistically indistinguishable from the top-ranked humans. In other words, the decade-long leap is almost invisible on easy data and unmistakable on complex social scenes. The same pattern holds on the private in-house CSB set, ruling out simple memorization of movie frames.

Error rates tell the same story in finer grain. Pre-MLLMs commit far more detection, recognition, hallucination, and scene-understanding errors on CSB than on MS-COCO. MLLMs drive all four of those rates close to zero on both datasets. The sole remaining systematic failure is spatial dependence: even the newest MLLMs sometimes base their description on different image regions than humans do.

Interestingly, SDE is also the error that hurts overall semantic similarity the least (Gemini-SI drops only to 0.56 versus a no-error baseline of 0.71). Detection, recognition, and hallucination errors are far more damaging (similarity falls to ~0.40). Thus the field has largely solved the "what is in the image" problems that dominated earlier captioning research; the residual gap is a subtler "which parts of the image matter for the story" problem.

Limitations

The most obvious constraint is sample size. Two hundred images is tiny by modern standards; the authors themselves note that exhaustive error classification is labor-intensive and therefore limited the set. Bootstrap statistics mitigate sampling variance but cannot invent diversity that is not present.

Movie frames, even carefully chosen ones, carry cinematic biases (dramatic lighting, staged compositions, cultural tropes). Whether the same error profile appears in raw surveillance video or egocentric footage remains open.

Error labeling for MLLMs relies on the models' own answers to probe questions; if a model is systematically over-confident or under-confident in those answers, the taxonomy could be slightly miscalibrated.

Finally, the study examines only free-form description. Many real applications require answering questions, counting, or performing multi-step reasoning; residual failures may reappear under those protocols (as concurrent work on orientation, counting, and sarcasm already suggests).

Why It Matters

Practitioners building systems that must interpret human behavior-assistive robots, content moderation, medical video analysis, autonomous vehicles in crowded scenes-now have quantitative evidence that today's MLLMs are no longer the brittle captioners of 2017. They can be trusted, at least for one-sentence scene summaries, on the kinds of social complexity that actually matter.

Equally important, the error taxonomy supplies a diagnostic language. When a model still fails, engineers can ask "is this a detection miss, a hallucination, or a spatial-dependence mismatch?" rather than simply "the caption is wrong." The finding that spatial dependence is the last major error also points research toward embodied and 3-D-aware architectures-exactly the direction suggested by concurrent work on affordance reasoning.

Finally, the methodological package (ranked multi-human gold standards + LLM embedding similarity + occlusion-based spatial analysis) is model-agnostic and can be reused as new MLLMs appear, giving the community a continuing yardstick that is harder to game than BLEU on MS-COCO.

Analogies & Intuition

Everyday analogy. Imagine two students taking a film-studies exam. Student A has only ever watched short clips of people sitting at tables or walking dogs; Student B has watched full narrative movies. On a quiz that shows a person holding a frisbee, both students score 95%. On a quiz that shows a tense family argument around a hospital bed, Student A produces a generic "people in a room" answer while Student B correctly notes who is comforting whom and why. The old MS-COCO numbers were the frisbee quiz; CSB is the hospital-bed quiz. The paper simply graded a decade of models on both quizzes and catalogued the exact mistakes each student made.

Engineering bridge. Think of a classic computer-vision pipeline: detector → classifier → higher-level fusion → language decoder. Detection and recognition errors are failures in the first two stages (the bounding-box and Softmax layers). Hallucinations are the language decoder inventing tokens that have no visual support-classic "exposure bias" or prior-dominated generation. Scene-understanding errors are failures of the fusion stage: the features exist but are never selected for the final sentence. Spatial-dependence errors are analogous to an attention map that lights up the wrong regions-exactly what Grad-CAM or occlusion sensitivity would reveal in a CNN, except here the "network" is a black-box MLLM and the readout is free-form text. The masking paradigm is therefore a black-box ablation study performed at the level of natural-language output rather than intermediate activations.

Glossary

  • Complex Social Behavior (CSB) dataset - A 100-image collection of movie frames deliberately chosen to require reasoning about multi-person interactions and intentions, used as a hard counterpart to MS-COCO.
  • Gold-standard description - The single human caption that received the most votes from an independent ranking panel; all model and remaining human captions are scored against it.
  • Gemini-SI / GPT-SI - Cosine similarity between 3 072-dimensional sentence embeddings produced by Gemini or GPT; shown to track human similarity judgments better than BERTScore.
  • Detection error - Failure to find an object that appears in the gold-standard description and is physically present in the image.
  • Recognition error - Correct localization of an object followed by an incorrect class label (or insertion of a misclassified irrelevant object).
  • Hallucination - Mention of an object that is neither present in the image nor a misclassification of something that is present.
  • Scene-understanding error - Correct detection and recognition of a relevant object that is nevertheless omitted from the final generated sentence.
  • Spatial dependence error (SDE) - Systematic mismatch between the image regions that most influence a model's description and the regions that most influence human descriptions, measured via large-region occlusion.
  • Pre-MLLM - Task-specific vision-language models (typically ResNet + LSTM or early Transformer) trained before the era of large-scale multimodal language models.
  • Masking paradigm - Controlled occlusion of successive large image regions to quantify how much each region contributes to the generated description for both humans and models.

Follow-Up Papers

From the paper's reference list (most directly related to the contribution):

Read on DEV Community ↗ ← Back to News

Comments

No comments yet. Start the discussion.