Aurelia: Test-time Reasoning Distillation in Audio-Visual LLMs

2025-03-29

AI Summary

Existing inference optimization methods struggle to adapt audio-visual large language models (AVLLMs) for complex multimodal reasoning. To address this, we propose AURELIA, the first framework introducing test-time inference distillation—enabling efficient transfer of structured, stepwise reasoning capabilities to AVLLMs without additional training or fine-tuning. For systematic evaluation, we introduce AVReasonBench, the first fine-grained audio-visual reasoning benchmark comprising 4,500 questions across six task categories, including the novel AV-GeoIQ dataset. Integrating Actor-Critic reinforcement learning, multimodal prompt engineering, and structured reasoning path modeling, AURELIA achieves up to 100% relative performance gain on AVReasonBench. Our evaluation comprehensively exposes critical reasoning deficiencies across 18 state-of-the-art AVLLMs and significantly enhances their practical deployability in real-world multimodal scenarios.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Recent advancements in reasoning optimization have greatly enhanced the performance of large language models (LLMs). However, existing work fails to address the complexities of audio-visual scenarios, underscoring the need for further research. In this paper, we introduce AURELIA, a novel actor-critic based audio-visual (AV) reasoning framework that distills structured, step-by-step reasoning into AVLLMs at test time, improving their ability to process complex multi-modal inputs without additional training or fine-tuning. To further advance AVLLM reasoning skills, we present AVReasonBench, a challenging benchmark comprising 4500 audio-visual questions, each paired with detailed step-by-step reasoning. Our benchmark spans six distinct tasks, including AV-GeoIQ, which evaluates AV reasoning combined with geographical and cultural knowledge. Evaluating 18 AVLLMs on AVReasonBench reveals significant limitations in their multi-modal reasoning capabilities. Using AURELIA, we achieve up to a 100% relative improvement, demonstrating its effectiveness. This performance gain highlights the potential of reasoning-enhanced data generation for advancing AVLLMs in real-world applications. Our code and data will be publicly released at: https: //github.com/schowdhury671/aurelia.

Problem

• Enhances audio-visual LLMs' reasoning without training
• Addresses multi-modal input processing complexities in AVLLMs
• Introduces benchmark for evaluating AV reasoning capabilities

Innovation

• Actor-critic based AV reasoning framework
• Test-time distillation of step-by-step reasoning
• AVReasonBench with 4500 audio-visual questions