Technical Summary: MCBench – A Multicontext Safety Assessment Benchmark for Omni Large Language Models

1. Problem Definition

Current multimodal safety benchmarks primarily focus on visual inputs, evaluating scenarios where visual cues alone suffice for safety assessment. This limitation prevents the evaluation of Omni Large Language Models (Omni LLMs), which are designed to process vision, audio, and speech simultaneously. Existing omnimodal benchmarks generally assess general reasoning capabilities rather than safety-specific reasoning. Furthermore, real-world safety assessment often requires synthesizing information across multiple modalities (visual, acoustic, and linguistic), a complexity not captured by single-modality or single-context benchmarks.

The core problem addressed is the lack of a comprehensive benchmark to evaluate the safety-awareness of Omni LLMs in multicontext settings, where accurate judgment depends on the integration of image, audio, and speech data.

2. Methodology: MCBench Construction

The authors introduce MCBench, a benchmark comprising 1,196 scenarios designed to test safety assessment across four coarse-grained categories:

1. Physical Harm (e.g., violence, environmental hazards, medical emergencies).
2. Social Harm (e.g., hate speech, misinformation, harassment).
3. Illegal Harm (e.g., crimes against persons, cyber crimes).
4. Property Damage (e.g., gas leaks, structural threats).

Data Collection Pipeline

The construction of MCBench follows a two-phase framework:

• Phase 1: Scenario Generation: Leveraging Claude-Sonnet-4.5, the authors generated fine-grained subcategories and specific safety scenarios. Crucially, unsafe scenarios were constructed using an If-Then logical structure (e.g., IF <condition 1> AND <condition 2> THEN UNSAFE). For each unsafe scenario, a corresponding safe counterpart was created by minimally modifying one or two conditions. This pairing allows for the assessment of model sensitivity to subtle contextual changes.
• Phase 2: Multimodal Data Generation: Textual descriptions from Phase 1 were used to synthesize multimodal inputs.
  • Images: Generated using Gemini-Flash-2.5.
  • Audio: Generated using Stable Audio 1.0. Complex audio descriptions were decomposed into individual sound events and recombined based on temporal relationships (sequential or simultaneous).
  • Speech: Included as part of the scenario context.
• Quality Control: Human experts verified scenarios for realism and logical consistency, and filtered out low-quality generated media.

Evaluation Protocol

The study evaluated both open-source (e.g., Qwen-Omni2.5, AnyGPT, InternOmni) and proprietary (e.g., Gemini-Flash-2.5, GPT-4o-mini) Omni LLMs.

• Metric: Accuracy in classifying scenarios as safe or unsafe.
• Judging: The GPT-4o model served as an "LLM-as-a-judge" to interpret the chain-of-thought reasoning traces generated by the evaluated models.
• Ablation Studies: The authors conducted experiments replacing visual and audio inputs with their textual descriptions to isolate the impact of modality perception versus reasoning.

3. Key Results

The evaluation of state-of-the-art models on MCBench revealed significant challenges:

• Overall Performance: Even the best-performing models (Gemini-Flash-2.5 and Qwen-Omni-2.5-3B) achieved only ~64.5% average accuracy, indicating that MCBench poses a substantial challenge for current architectures.
• Category Disparities:
  • Models performed relatively well on Physical Harm and Property Damage, where salient visual and acoustic cues are present.
  • Models struggled significantly with Social Harm and Illegal Harm, which involve subtle or non-physical risks. For instance, some open-source models exhibited oversensitivity (classifying safe scenarios as unsafe) in Social Harm, while others showed insensitivity to unsafe scenarios in the same category.
• Modality Impact:
  • Visual Context: Replacing images with textual descriptions often improved model performance, suggesting that current Omni LLMs struggle to extract relevant safety cues directly from images.
  • Audio Context: Larger models (e.g., Gemini-Flash-2.5) performed better with textual audio descriptions than raw audio, while smaller models degraded. This indicates that capable models can better integrate textual context, whereas acoustic features provide limited additional discriminative power for safety judgment in current models.
• Reasoning vs. Perception:
  • Perception: Models generally demonstrated the ability to extract ground-truth predicates (relevant information) from multimodal inputs.
  • Reasoning: The primary failure mode was the integration of these cues. Models often failed to combine extracted information effectively, leading to incorrect safety judgments.
  • Oversensitivity: When provided with explicit reasoning contexts (ground-truth predicates), models became oversensitive, incorrectly classifying safe scenarios as unsafe. They tended to anchor judgments on a single concerning cue (e.g., a text notification) while ignoring contradictory evidence from other modalities (e.g., voice commands indicating safety).

4. Key Contributions

1. MCBench Benchmark: The first multimodal, multicontext safety benchmark tailored specifically for evaluating Omni LLMs, featuring 1,196 scenarios with paired safe/unsafe instances across four safety taxonomies.
2. Comprehensive Analysis: A detailed evaluation of state-of-the-art open-source and proprietary Omni LLMs, revealing that while models can extract modality-specific information, they lack robust cross-modal reasoning capabilities for safety-critical settings.
3. Failure Diagnosis: Identification of specific failure modes, including oversensitivity to isolated cues and a lack of balanced integration mechanisms, which leads to systematic false positives in safe scenarios.

5. Significance and Claims

The paper claims that MCBench serves as a rigorous testbed for advancing the development of safe Omni LLMs. The findings underscore a critical gap in current architectures: while models can process multiple modalities, they lack the robust cross-modal reasoning necessary to synthesize these inputs for accurate safety judgments in complex, real-world contexts.

The authors conclude that future research must focus on:

• Developing architectures and training strategies that enforce balanced multimodal reasoning.
• Improving mechanisms for information aggregation across modalities to prevent models from over-relying on single, potentially misleading cues.
• Addressing the specific challenges of intangible risks (social and legal harms) where visual and acoustic cues are less definitive.

The paper positions MCBench not as a final solution, but as a necessary step to expose the limitations of current Omni LLMs in safety-critical applications, thereby guiding the community toward more reliable multimodal safety systems.