Official Review · ICLR 2026

This paper introduces a benchmark for omnidirectional embodied intelligence, aiming to evaluate multi-view and multi-task embodied reasoning. The benchmark integrates panoramic and multi-view visual inputs into embodied reasoning tasks such as scene understanding, visual question answering, and navigation. The authors evaluate multiple multimodal large language models (MLLMs), including Qwen2.5-VL-7B, to demonstrate the benchmark’s potential for analyzing 3D reasoning and perception capabilities. The paper claims to provide a step toward unified evaluation of omnidirectional embodied understanding. 1. Relevant topic. The paper addresses an emerging direction in embodied AI—omnidirectional and multi-view reasoning—which is timely given recent interest in panoramic perception and 3D-aware MLLMs. 2. Benchmark perspective. The attempt to unify multi-view, multi-task embodied reasoning under a single benchmark setup is conceptually interesting and could contribute to evaluation consistency in future research. 3. Comprehensive evaluation coverage. The authors test several models, including large MLLMs and baseline vision-language systems, which provides a broad empirical overview of current model capabilities. 1. Unclear motivation for panorama vs. multi-view reasoning. The paper fails to clarify why omnidirectional (panoramic) images provide distinct advantages over standard multi-view or 3D representations for scene reasoning. The conceptual difference between panoramic projection and multi-view fusion is not discussed or experimentally analyzed. This weakens the benchmark’s motivation and contribution. 2. No data quality assessment. As a benchmark paper, it is essential to report dataset statistics, diversity metrics, and annotation quality analysis. None of these are provided. Without data quality validation, it is difficult to evaluate the credibility or difficulty of the proposed dataset. 3. Missing human evaluation or validation. There is no human evaluation to verify whether model outputs align with human-level reasoning. Human scores or expert annotations are crucial for validating embodied reasoning benchmarks and ensuring that high model scores reflect genuine understanding rather than dataset biases. 4. Benchmark not sufficiently challenging. The reported results show very high performance (e.g., 67.78 by Qwen2.5-VL-7B), which suggests that the benchmark is too easy for current MLLMs. If existing models already achieve strong performance, the benchmark fails to expose meaningful limitations or reasoning gaps—undermining its purpose as a challenging evaluation tool. 1. What is the conceptual or practical difference between panorama-based and multi-view representations for embodied reasoning? 2. How was data quality ensured or validated during benchmark creation (e.g., annotation consistency, scene diversity)? 3. Were any human evaluations conducted to confirm that the benchmark tasks align with real-world embodied reasoning difficu