Multimodal reasoning demands that models combine multiple pieces of evidence into decisions that satisfy all relevant constraints. Conventional evaluations focus on final‑answer accuracy, leaving it unclear whether failures stem from reading individual facts or from integrating them. This work centers on the “multi‑evidence integration gap,” linking the correctness of local fact recovery to the validity of global decisions on controlled subset‑selection and path‑finding tasks. The evaluation suite includes conditional performance given correct local answers, same‑grid local queries, text‑oracle inputs, retrieval controls, and constraint ablation studies. We also probe hidden states and perform activation replacement across network depth, token positions, and model families. Findings reveal that accurate local recovery does not guarantee global correctness. For instance, Gemma‑3‑27B recovers every isolated item fact yet achieves only 19.75% global validity on eight‑item visual instances. Probes can decode solution‑membership information from global‑prompt states, and clean‑state replacement restores failed decisions within model‑specific depth ranges. These results indicate that merely recovering facts and representing solution‑related information is insufficient for reliable joint decisions. The study urges the development of evaluations and model improvements that explicitly target how evidence is combined, alongside advances in perception and model scaling.
Review