CORD: Bridging the Audio-Text Reasoning Gap via Weighted On-policy Cross-modal Distillation

TL;DR

This paper introduces CORD, a novel cross-modal alignment framework that improves audio-text reasoning in large audio language models by using online self-distillation and multi-level alignment, significantly reducing the performance gap with text models.

Contribution

CORD presents a unified, on-policy cross-modal self-distillation method that aligns audio and text reasoning, improving reasoning capabilities with minimal synthetic data.

Findings

CORD enhances audio-conditioned reasoning across benchmarks.

It reduces the audio-text performance gap effectively.

The method achieves strong results with only 80k synthetic samples.

Abstract

Large Audio Language Models (LALMs) have garnered significant research interest. Despite being built upon text-based large language models (LLMs), LALMs frequently exhibit a degradation in knowledge and reasoning capabilities. We hypothesize that this limitation stems from the failure of current training paradigms to effectively bridge the acoustic-semantic gap within the feature representation space. To address this challenge, we propose CORD, a unified alignment framework that performs online cross-modal self-distillation. Specifically, it aligns audio-conditioned reasoning with its text-conditioned counterpart within a unified model. Leveraging the text modality as an internal teacher, CORD performs multi-granularity alignment throughout the audio rollout process. At the token level, it employs on-policy reverse KL divergence with importance-aware weighting to prioritize early and semantically critical tokens. At the sequence level, CORD introduces a judge-based global reward to optimize complete reasoning trajectories via Group Relative Policy Optimization (GRPO). Empirical results across multiple benchmarks demonstrate that CORD consistently enhances audio-conditioned reasoning and substantially bridges the audio-text performance gap with only 80k synthetic training samples, validating the efficacy and data efficiency of our on-policy, multi-level cross-modal alignment approach.