Answering what-if queries about a scene with a VLM usually means injecting the assumption as text or repainting the scene with a generative model. We instead move the edit to the representation level, before the model input. The image is abstracted into a set of object-level tokens, and the original image never enters the VLM. This design rests on an open question: when do frozen VLMs actually respond to such token edits? We introduce an answer-key-free protocol: no post-edit answer is annotated. It scores edits whose answers are logically determined, and audits itself by reversing each scoreable choice. The protocol reveals three structures. The response is not free: explicit edit teaching, not ordinary VQA training, produces it in dense scenes and multiplies it in sparse ones, on all three operations. Once on, it is governed by token cleanliness and density, with deployable detector+segmenter tokens competitive with the oracle and outperforming it on VRSBench. And reading is a separable axis: the image-free token route preserves 92-96% of a matched patch-token baseline's free-text VQA, and the answers measurably depend on the tokens. The response, cleanliness, and reading structures are sign-preserved across two remote-sensing datasets (iSAID, VRSBench) and three frozen LM backbones. We release the probe generator, records, judge logs, and code.
AI systems are increasingly evaluated for legally accountable settings, where correct outputs must also be justifiable against an applicable legal standard. Existing legal-AI benchmarks and LLM-as-judge protocols provide important infrastructure for measuring task performance and open-ended response quality. We contribute one additional evaluation signal: a dual-judge protocol that pairs a standard 0-10 quality judge with a strict binary semantic-equivalence judge against a human-curated reference. We study a controlled, visually grounded regulatory task - UK traffic-sign interpretation, whose meaning is a codified question with a known reference for every input - and measure not merely whether the two judges disagree (by construction they must) but how much and where. On 4,680 evaluations under seven visibility levels and two occlusion modes, the two judges are moderately associated (point-biserial r = 0.644), while revealing an asymmetric Type II pattern affecting 8.0% of all evaluations. Its distribution is instructive: the marginal rate peaks at high visibility (14.2% at v = 0.8) simply because high-scoring answers are common there, but conditioned on the answer already scoring above 7, the rate is highest under heavy occlusion (54-63% at v <= 0.3), so a high quality score is least trustworthy when the input is most degraded. We are explicit that the signal is a property of this judge and reference: a 49-row human check shows the 0-10 judge aligns closely with everyday-reader judgement (Pearson r = 0.81; r = 0.80 with the LLM accuracy sub-score), while the equivalence judge is fairly but one-directionally stricter. The protocol adds one LLM call per evaluation and surfaces a signal single-judge protocols do not report. We release the prompt template, occluded variants, and full evaluation results.
Reasoning-augmented text-to-image models such as GoT-R1 emit an explicit textual plan - object names, attributes, and bounding boxes - before generating image tokens. When such a model fails a compositional prompt, is the plan wrong, or is the plan right and the decoder unfaithful? Because the plan is machine-readable it can be edited before decoding, which makes the two separable. We first validate the ruler. Swapping the two bounding boxes inside the model's own chain demonstrably flips the generated layout: detector-based accuracy falls 0.75 -> 0.48 (p<1e-3), while a widely used VQA-based spatial metric rises. A five-rater human study agrees with the detector on 81% of items and with the VQA judge on 57%. All spatial results therefore use geometric scoring. Under sound measurement the decoder is a faithful executor: 94% of generated layouts realize the planned relation, and object-box binding survives reordering of the plan's object segments. The planner is the bottleneck. It writes wrong relations for phrasing-dependent reasons - 98% accuracy on "left" against 54% on "right" for semantically identical layouts, a raster-order bias we isolate with a mention-order control - and cluttered geometry that the decoder faithfully reproduces. Editing the plan therefore fixes the image without retraining: symbolic verification with resampling gives +5.0 points (p<1e-3), minimal in-place repair +6.0 (p=.02), rewriting only box geometry +10.7 (p<1e-4), and replacing the plan outright +13.3 (p=1e-4). Gains are indifferent to the plan's prose style and to its likelihood under the planner, but not to its geometry. Modular planner-decoder designs are therefore viable, provided the plan is internally consistent: box-text contradictions induce object duplication and identity fusion. We release the plan-fidelity evaluation protocol, all plans, and 12k generated images.
Text-to-image and multimodal generative models are increasingly used to produce scientific figures such as mechanism diagrams, experimental-design schematics, conceptual frameworks, and graphical abstracts. Yet existing image-generation benchmarks (e.g., GenEval, T2I-CompBench, DPG-Bench) evaluate natural images and measure compositionality, object counting, or photorealism. None of them measure what makes a generated scientific figure usable: correct and legible text labels, faithful depiction of entities and their relations, coherent diagrammatic structure, and adherence to disciplinary drawing conventions. We introduce SciDraw-Bench, a benchmark of 32 structured scientific-figure generation tasks spanning eight figure types and ten disciplines, where each task pairs a natural-language prompt with a machine-checkable specification of required labels, relations, components, conventions, and negative constraints. We propose a four-dimensional evaluation protocol: Text Fidelity (OCR-based label recall and character error rate), Semantic Correctness (vision-language-model judging against the specification), Structural Quality, and Convention Adherence, together with a meta-evaluation protocol and a preliminary inter-judge reliability analysis (human-rating validation is ongoing). We evaluate a domain-specific system, SciDraw AI, against representative general-purpose text-to-image models, and outline a code-to-figure baseline as a planned extension. In a pilot over all eight figure types, the domain-specific system substantially outperforms the general-purpose baselines on every dimension and figure type, with the largest gaps on semantic correctness and convention adherence; text fidelity remains the hardest dimension for all systems.
Static visual question answering (VQA) benchmarks age quickly: Once the items leak into training corpora, scores can reflect memorization rather than genuine visual ability, thus obscuring real progress. Rebuilding high-quality benchmarks such as V*Bench requires substantial human annotation, yet each static release can quickly become another leaked artifact. We propose ReKey, a live benchmark protocol that randomly regenerates the answer-bearing local detail, or visual key, in real images at evaluation time. Using human-validated edit slots, ReKey samples fresh instances with new answers, construction-grounded labels, and controlled visual-search difficulty. On V*Bench, the ReKey regenerated benchmark reveals a sharp score jump across eight frontier vision-language models (VLMs): The original items score 9.5--18.8 percentage points higher than the regenerated variants. By making the visual key renewable, ReKey keeps evaluation fresh as models and training data evolve.
Spatial reasoning is fundamental to robotics, autonomy, and embodied AI, yet modern vision-language models (VLMs) remain unreliable on metric distance queries. A common assumption is that consistent predictions across viewpoints reflect geometric grounding. We test this assumption and find the opposite: leading VLMs often produce view-invariant and consistent answers even when those answers are incorrect, indicating weak coupling between predictions and viewpoint-specific visual evidence. We introduce \textbf{ViewDiag}, a controlled multi-view evaluation protocol built from Hypersim, ScanNet, and KITTI360, comprising 176 object-pair tracks across 80 scenes with 2--10 views per track. The protocol evaluates models along three axes: metric accuracy, distributional concentration, and internal collapse, the last of which is assessed using a latent feature probe. Across diverse models, we observe a consistent pattern of high prediction stability paired with substantial error, clustering in a regime characterized by strong consistency but low accuracy. \noindent These results challenge the common use of cross-view consistency as a proxy for geometric understanding. Instead, we show that stable predictions may reflect prior-driven collapse rather than evidence-sensitive reasoning. ViewDiag provides a controlled benchmark and diagnostic framework for evaluating whether spatial VLMs are not only accurate, but also meaningfully coupled to visual evidence.