anonymous-submission-RIG-bench/RIG-Bench-Qualitative

Extended Qualitative Examples for RIG-Bench

This page provides additional qualitative examples complementing the paper. It is created in response to reviewer feedback requesting more generated examples illustrating both successful reasoning and common failure modes across model families.

These examples visualize the RIG-Bench setting: a model receives visual context and a short instruction, infers the latent rule or target outcome, and synthesizes the answer directly as an image. The selected cases span the four task families in the benchmark: Concept-based, Transformation-based, Pattern & Structure, and Scenario-based reasoning.

The examples illustrate both successful outputs and characteristic failure modes. Some models successfully generate a new scene that preserves the intended interaction or visual concept. However, many visually plausible outputs still fail the underlying reasoning requirement: paths are traced incorrectly, chemical products are hallucinated or mechanistically invalid, grid transformations preserve surface appearance but not the inferred rule, and matrix tasks often reproduce the full context rather than generating only the missing cell. Video-generation models may copy or minimally alter the visual context while failing to perform the required symbolic, spatial, or structural transformation.

These examples are intended as qualitative illustrations rather than an additional quantitative comparison. Together, they highlight the central RIG-Bench distinction between visual plausibility and reasoning-grounded image generation.