REVIEW 8 cited by
What's "up" with vision-language models? Investigating their struggle with spatial reasoning
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Recent vision-language (VL) models are powerful, but can they reliably distinguish "right" from "left"? We curate three new corpora to quantify model comprehension of such basic spatial relations. These tests isolate spatial reasoning more precisely than existing datasets like VQAv2, e.g., our What'sUp benchmark contains sets of photographs varying only the spatial relations of objects, keeping their identity fixed (see Figure 1: models must comprehend not only the usual case of a dog under a table, but also, the same dog on top of the same table). We evaluate 18 VL models, finding that all perform poorly, e.g., BLIP finetuned on VQAv2, which nears human parity on VQAv2, achieves 56% accuracy on our benchmarks vs. humans at 99%. We conclude by studying causes of this surprising behavior, finding: 1) that popular vision-language pretraining corpora like LAION-2B contain little reliable data for learning spatial relationships; and 2) that basic modeling interventions like up-weighting preposition-containing instances or fine-tuning on our corpora are not sufficient to address the challenges our benchmarks pose. We are hopeful that these corpora will facilitate further research, and we release our data and code at https://github.com/amitakamath/whatsup_vlms.
Forward citations
Cited by 8 Pith papers
-
JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles
With interlocking puzzle pieces, vision-language models mostly fail even at 4x4, and fine-tuned models that solve 4x4 fall to near-random by 12x12.
-
Synthetic Visual Genome
A GPT-4V/GPT-4o pipeline for completing and refining scene graph annotations yields a dense synthetic dataset that, after instruction tuning, gives a 3B model strong relationship understanding and grounding results.
-
SpatialThinker: Reinforcing Scene Graph-Grounded Spatial Reasoning via Dense Rewards
Dense scene-graph-grounded rewards let a 7B multimodal LLM trained on 7K synthetic questions beat SFT and sparse-RL baselines and outscore GPT-4o on average across 12 spatial/real-world benchmarks.
-
Beyond Accuracy: Metrics that Uncover What Makes a 'Good' Visual Descriptor
Two label-free metrics, DINO alignment and CLIP similarity, rank text descriptor sets in line with downstream accuracy and track iterative refinement.
-
A Good CREPE needs more than just Sugar: Investigating Biases in Compositional Vision-Language Benchmarks
Blind text-only likelihood models match or exceed CLIP on many compositionality benchmarks because positives and negatives differ systematically in length, plausibility, or image style.
-
Scene Graph-Guided Proactive Replanning for Failure-Resilient Embodied Agent
A robot replanner that compares scene graphs to successful demonstrations before each subtask, triggering LLM-based replanning on mismatch, raises task success in AI2-THOR.
-
Argus Inspection: Do Multimodal Large Language Models Possess the Eye of Panoptes?
A new 1,430-item multimodal benchmark shows that leading multimodal LLMs rarely notice small visual traps needed for commonsense safety reasoning, with top scores near 0.46 on a scale whose maximum is about 0.97.
-
IKIWISI: An Interactive Visual Pattern Generator for Evaluating the Reliability of Vision-Language Models Without Ground Truth
A visual heatmap tool lets people rate vision-language model reliability in video by inspecting patterns of green and red cells, with user ratings tracking objective F1 scores when those exist.
Discussion (0). Sign in to comment.