TinyGiantALM: A Compact Audio-Language Model for Intent-Aware Reasoning under Resource Constraints

TinyGiantALM is a compact 1.5B parameter audio-language model designed for efficient intent-aware reasoning in resource-constrained environments. Utilizing an Instruction-Aware Feature Refinement framework with a Query-guided Projector and Semantic Gating, it achieves 46.4% zero-shot accuracy on the MMAR benchmark, outperforming larger 7B-13B models. This model presents a viable solution for practitioners needing robust audio reasoning capabilities without the resource demands of larger models, particularly in mixed-modality scenarios.