Research map / Vision-language and multimodal / Multimodal LLMs

ODI-Bench: Can MLLMs Understand Immersive Omnidirectional Environments?

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ODI-Bench, a 2,000-image omnidirectional image QA benchmark across 10 general and spatial tasks evaluated in close- and open-ended settings on 20 MLLMs, plus Omni-CoT, a training-free chain-of-thought method that boosts MLLM accuracy on ODIs.

Problem

Existing MLLM benchmarks do not adequately evaluate comprehension of omnidirectional (360°) images, which are high-resolution, distorted in ERP format, and require immersive spatial reasoning; prior ODI benchmarks are low-resolution, indoor-only, or automatically generated with textual biases.

Approach

A high-resolution, indoor-and-outdoor ODI benchmark with human-verified QA in both close-ended and open-ended formats, paired with a training-free CoT pipeline that adds viewpoint captions from six cubemap views and refined narrow-FoV crop cues grounded by the MLLM itself.

Key results

• o3 achieves the top close-ended overall score of 62.62 on ODI-Bench, while the best open-source model InternVL3-78B reaches 59.43.
• Omni-CoT improves GPT-4o overall close-ended accuracy from 55.79 to 62.08 and Gemini-2.0-flash from 57.12 to 63.89 without training.
• Egocentric view orientation for GPT-4o rises from 43.24 to 71.53 with Omni-CoT, while spatial tasks remain hard with allocentric view orientation near 33-40.

Surprising findings

• Multiple-choice options can act as hints: several models score higher in close-ended than open-ended settings on counting and OCR, and some answer correctly open-ended but pick wrong options (Option Anchoring Bias).
• Simple resizing of ODIs to 512x1024 hurts general-level tasks substantially while spatial-level tasks are less affected.

Limitations

• Close-ended results are sensitive to answer options, which can bias evaluation (option anchoring).
• Spatial tasks remain near random chance for many models, and Omni-CoT gains are uneven on some tasks (e.g., relative direction drops for several models).

Experimental setup

Datasets: ODI-Bench

Benchmarks: ODI-Bench

Metrics: accuracy LLM-based evaluator score

Baselines: GPT-4o Gemini-2.0-flash Qwen2.5-VL-72B InternVL2.5-8B Zero-shot CoT Blind GPT-4o multi-view input video input

Builds on: Zero-shot CoT GroundedSAM

Closest papers by meaning

• OpenView: Empowering MLLMs with Out-of-view VQA (NeurIPS 2026 · similarity 0.88)
• OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs (ICLR 2026 · similarity 0.87)
• Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.87)
• OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding (NeurIPS 2025 · similarity 0.87)
• MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness (NeurIPS 2025 · similarity 0.86)
• ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning (ICML 2026 · similarity 0.86)