arXiv:2609.13225v1 Announce Type: cross Abstract: Benchmarks agree that vision-language models reason poorly about low-level manipulation, but an aggregate accuracy score does not say which step fail...
The paper introduces a counterfactual benchmark to evaluate whether vision models understand physical constraints or merely learn rendering shortcuts. Using Mitsuba 3, it generates 5,500 synthetic images across 500 scene families, each with a clean image and ten violations (shadows, reflections, support, surface response, occlusion). Diagnostic baselines (LLaVA‑1.5‑7B, Qwen2.5‑VL‑7B, InternVL3.5‑8B) achieve 64.0–67.8 % macro‑F1 on a held‑out synthetic test set, but performance drops to 40.8 % for LLaVA‑1.5‑7B under an intervention shift, indicating reliance on rendering cues rather than true physical reasoning.
arXiv:2608.21832v1 Announce Type: new Abstract: Computer-use agents ground natural-language instructions in screenshots to locate interface elements, yet existing benchmarks do not isolate whether mo...
arXiv:2608.16081v2 Announce Type: replace Abstract: Open-weight and frontier vision-language models (VLMs) perform well on general image understanding, but their ability to interpret fine-grained han...
The paper investigates a specific failure mode in vision‑language‑action models, termed instruction‑action binding, where models respond to language and vision separately but fail to combine them to select the correct action under counterfactual changes. Through behavioral analyses of fine‑tuned policies, the authors show that failed rollouts often preserve source behavior or switch to other demonstrated tasks, indicating that language is not ignored but mis‑bound. They propose Equivariant Counterfactual Training (ECT), which supplies counterfactual demonstrations and a paired loss to enforce correct action selection, achieving significant performance gains across simulated and real‑world benchmarks.
AMIGO (Agentic Multi-Image Grounding Oracle Benchmark) is a long-horizon evaluation framework for vision‑language models that tests hidden‑target identification across galleries of visually similar images. The benchmark requires a model to ask a sequence of attribute‑focused Yes/No questions, receiving Yes/No/Unsure feedback and penalizing invalid actions with Skip, thereby stressing question selection under uncertainty, constraint tracking, and fine‑grained discrimination. Using the Guess My Preferred Dress task, the study shows that final‑answer accuracy alone overstates performance, as models may guess correctly without verified evidence, waste turns, or violate the protocol, highlighting the need for combined visual discrimination, informative questioning, and robust protocol adherence.
arXiv:2609.26093v1 Announce Type: new Abstract: Vision-language models can answer spatial relation questions confidently even when the image supports an incompatible relation. We formulate relation-g...
arXiv:2607. 13305v1 Announce Type: cross Abstract: Benchmark accuracy in video large language models (LLMs) is often treated as evidence of visual understanding.
arXiv:2608. 06154v1 Announce Type: cross Abstract: Vision-language models (VLMs) are increasingly used as zero-shot controllers, but successful trajectories do not necessarily show that decisions are grounded in visual input: simulator dynamics and conservative action priors can produce favourable scores without meaningful perception.
arXiv:2609.36572v1 Announce Type: new Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has been extended to Large Vision-Language Models (LVLMs), and perception-aware methods further e...
arXiv:2604. 27720v2 Announce Type: replace Abstract: Vision-language models (VLMs) are increasingly applied to medical visual question answering (Med-VQA), yet whether they can \emph{localize} the evidence behind their answers---a prerequisite for clinical auditability---is poorly characterized.
VABench is a benchmark that tests general‑purpose multimodal large language models (MLLMs) on embodied spatial intelligence by requiring them to observe, reason, act, and revise based on visual demonstrations and active perception. The benchmark includes 14 task families, a fixed model‑agnostic controller, and evaluates models on target localization, spatial relations, and long‑horizon composition tracks without providing privileged object poses or learned action heads. Results show that while the best model achieves perfect target localization, overall task success remains modest, and active camera control and geometric transfer significantly influence performance.