Speech-IFEval: Evaluating Instruction-Following and Quantifying Catastrophic Forgetting in Speech-Aware Language Models

TL;DR

Speech-IFEval tackles the problem that speech-aware language models degrade textual instruction-following after speech-text training, and existing benchmarks fail to separate perception from instruction-following. The authors propose Speech-IFEval, a framework with a cascade reference system and constraint-based outputs, plus a forgetting-rate metric $\Delta$ to quantify catastrophic forgetting. Results show most SLMs lag far behind text-based LLMs in instruction-following, with significant sensitivity to prompt variations; some models like DeSTA2 mitigate forgetting through speech-text alignment, and test-time LoRA scaling offers a partial remedy at the cost of task-level accuracy. The work provides a new standard for evaluating SLMs, guiding future research toward robust, instruction-following capabilities in multi-modal models.

Abstract

We introduce Speech-IFeval, an evaluation framework designed to assess instruction-following capabilities and quantify catastrophic forgetting in speech-aware language models (SLMs). Recent SLMs integrate speech perception with large language models (LLMs), often degrading textual capabilities due to speech-centric training. Existing benchmarks conflate speech perception with instruction-following, hindering evaluation of these distinct skills. To address this gap, we provide a benchmark for diagnosing the instruction-following abilities of SLMs. Our findings show that most SLMs struggle with even basic instructions, performing far worse than text-based LLMs. Additionally, these models are highly sensitive to prompt variations, often yielding inconsistent and unreliable outputs. We highlight core challenges and provide insights to guide future research, emphasizing the need for evaluation beyond task-level metrics.