The benchmark said pass. The parts it did not check failed

Passing where someone looked

A benchmark that checks only part of the output will pass a model that fails elsewhere, and a preprint on image-conditioned video generation shows the failure in numbers. Models in the study achieve high fidelity at anchor points, scoring 0.90 to 0.94, while fabricating text in peripheral regions at rates of 36 to 54 percent. Existing reference-based consistency scores do not track that peripheral fabrication, and string-matching metrics cannot catch fabricated text that looks legible but has no reference counterpart. The lesson generalizes beyond video: wherever an evaluation samples a few checked locations and assumes the rest, a model can learn to be faithful exactly where the checker looks. m:…4898

A second preprint offers a way to locate failures instead of averaging over them. Its diagnostic protocol for vision-language models, called See, Think, Use, traces an answer to a perception failure, a reasoning failure, or a usage failure, without needing access to model weights. When the model's localization boxes are correct, localization-first prompting repairs reasoning failures, and allocating compute based on the measured failure stage beats uniform compute policies at lower cost. For a buyer, that distinction is actionable: a perception failure calls for better sensing or input, a reasoning failure calls for different prompting or compute, and a usage failure calls for a different interface. A single accuracy number hides all three, and fixing the wrong stage wastes the budget. m:…4144

The scorer is part of the score

Even when the output is fully visible, the scorer can move the result. In a preprint auditing multimodal benchmark graders, the authors find that scoring rules shift accuracy by up to 22 points on identical model outputs. Parsers misinterpret or discard valid answers, and benchmarks with high non-compliance rates produce artifacts where the scorer abstains instead of guessing, skewing results. The authors' proposed remedy, a blinded human-plus-language-model adjudication protocol, reaches a Cohen's kappa of 0.956, and they recommend reporting a non-compliance rate alongside results under at least two distinct scoring rules. Until that practice spreads, a buyer comparing two models on different benchmarks may be comparing two graders, not two models. m:…8800

Benchmark compression adds a quieter version of the same risk. Random sampling for compressing benchmarks produces a long-tailed error distribution, with rare but large errors, according to a preprint proposing a method called BestRand. That method cuts high-error risk by 50.2 percent and estimation error by 16.4 percent compared with random sampling, and reduces out-of-distribution estimation error by 21.8 percent on average across shifts in score range, model family, and time. The implication is that a cheap benchmark is not automatically a misleading one, but a benchmark compressed by naive sampling can misrank models precisely on the cases where the ranking matters most. m:…5688

Right words, wrong place

A drone-agent benchmark closes the set with a failure that sounds abstract until it is deployed. Evaluating 45 multimodal models on grounded agent-world reasoning, the authors find that agent-state reasoning is consistently weaker than world-state reasoning, and that semantic answers do not reliably coincide with spatial or temporal grounding in aerial contexts. A model can therefore describe a scene correctly while being wrong about where it is, when something happened, or what state it is in. A lightweight optimization probe improves dual-cognition performance in the study, which suggests the gap is at least partly trainable. But the deployment lesson stands: for embodied agents, a correct-sounding answer is not evidence of correct grounding, and evaluations must test the agent's state separately from its description of the world. m:…4345