GroundBench: A Factorized, Counterfactual Benchmark for Locating VLM Affordance Failures

Benchmarks agree that vision-language models reason poorly about low-level manipulation, but an aggregate accuracy score does not say which step fails. We separate two steps that affordance questions conflate: identifying which part of an object to act on, and knowing what action that part requires. Across 19 articulated objects we asked eight models, spanning three developers, what motion a robot should apply. Under an open prompt, push was produced once in 64 evaluations where it was correct, despite being correct for 8 of 19 objects and appearing in the offered label set every time. Inspecting the outputs showed why: models described a different part than the one being scored, e.g. explaining how to pick up a camera rather than press its button. Naming the target part raises action accuracy by 0.32 to 0.63 for every model, from 0.158-0.474 to 0.684-0.947, and push recall from 0-1/8 to 7-8/8. No model beats a constant answer that ignores the image under the open prompt; once the part is named, all eight do. Asked to describe the same part in free prose with no label set, models produce pressing language for 6 to 8 of 8. These results are hard to reconcile with missing action knowledge, and instead point to part grounding as the dominant bottleneck, a pattern that holds across all three model families and does not diminish with capability. Naming the part supplies the grounding variable, so this bounds what a perfect part detector would offer rather than demonstrating a general model of mechanics. Two supporting results agree: on real photographs only three of eight models localize grasp points better than a constant baseline, and on rendered objects none do. We also document two measurement errors of our own, a threshold that let a constant baseline score 0.929 and a labelling rule wrong on 4 of 19 objects, both caught only by testing our numbers against trivial alternatives.

Task-conditioned manipulation requires grounding instructions to task-relevant functional parts rather than object categories. This setting is scene-dependent and often one-to-many in cluttered scenes: the same object may afford different interactions across tasks, while a single task may correspond to either one functional region or multiple valid functional regions, depending on the scene layout. Existing affordance datasets and benchmarks remain misaligned with this setting, as they typically focus on grasping or object-level affordances, rely on synthetic scenes, or assume a single instruction-region correspondence. We present Affordance2Action (A2A), a benchmark-centered learning framework for scene-level, task-conditioned part affordance grounding. At its core is A2A-Bench, a manipulation-oriented benchmark that covers both single-region and multi-region instruction correspondences in everyday scenes, with the latter highlighting the ambiguity and diversity of affordance grounding in realistic multi-object environments. To construct it at scale, we build A2A-AffordGen, an agent-assisted annotation pipeline that combines language-model filtering, interactive part segmentation, instance-level mask-out refinement, task-reasoning instruction generation, and human verification. A2A-Bench's supervision further supports diverse downstream applications, with real-time affordance grounding and affordance-conditioned manipulation policies as two representative examples. Experiments show that A2A exposes substantial gaps in generic segmentation, VLM-based grounding, and affordance distillation baselines, while improving task-level localization and providing useful spatial priors for downstream manipulation.

Final-answer accuracy reports whether a vision-language model was right, but not whether its answer was grounded in the image, whether its stated reasoning reflects that evidence, or what the answer cost to produce. Those are the properties that decide whether a system can be deployed. We present PRISM-BENCH, a benchmark of 1,000 expert-curated questions spanning five under-represented reasoning categories, in which every item carries an idealized reasoning trace written by one human annotator and validated by another. The traces support two diagnostics that accuracy cannot provide. A perception probe, in which the model answers from a description it wrote itself, attributes roughly 85% of frontier-model failures to misperception and places the bottleneck at grounding. Trace alignment predicts answer correctness for all 25 systems tested (macro AUROC 0.66) and survives length control, so it measures reasoning content and not verbosity. We further introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length. Across 25 open and closed systems, including GPT-5 and Claude Opus 4.6, the best accuracy is 66.0%. On the four categories with a human baseline the best system trails humans by 18.4 points, and recent frontier gains fall on recognition and not on spatial grounding.