AudioRouter: Data Efficient Audio Understanding via RL based Dual Reasoning

Large Audio Language Models (LALMs) have demonstrated strong capabilities in audio understanding and reasoning. However, their performance on fine grained auditory perception remains unreliable, and existing approaches largely rely on data intensive training to internalize perceptual abilities. We propose AudioRouter, a reinforcement learning framework that enables LALMs to improve audio understanding by learning when and how to use external audio tools. Rather than tightly coupling tool usage with audio reasoning, AudioRouter formulates tool use as an explicit decision making problem and optimizes a lightweight routing policy while keeping the underlying reasoning model frozen. Experimental results show that AudioRouter achieves substantial improvements on standard audio understanding benchmarks while requiring up to 600x less training data to learn tool usage compared with conventional training paradigms. These findings suggest that learning effective tool usage offers a data efficient and scalable alternative to internalizing perceptual abilities in LALMs.

💡 Research Summary

AudioRouter addresses a fundamental limitation of large audio language models (LALMs): while they excel at high‑level semantic tasks, they struggle with fine‑grained auditory perception such as pitch estimation, event counting, or temporal structure analysis. Traditional approaches attempt to internalize these perceptual abilities by training LALMs end‑to‑end on massive audio‑text datasets, which is both data‑hungry and inefficient because high‑quality annotated audio is costly and perceptual competence does not scale linearly with data volume.

The authors propose a fundamentally different paradigm: keep the LALM’s reasoning backbone frozen and train a lightweight routing policy that decides when and which external audio tools should be invoked. The system consists of three components: (1) a frozen audio reasoning model fθ (e.g., Whisper‑LLM, Qwen‑Audio), (2) a set of pre‑existing specialized audio tools (pitch tracker, duration analyzer, sound classifier, etc.), and (3) a Router π(a|x,q,C) that selects an action a from the space A = {Direct, t₁,…,t_K}. Direct means the model answers using only its internal knowledge; t_k means the model first calls tool t_k, receives structured evidence r_k, and then feeds r_k together with the original inputs to fθ for the final answer.

Training the Router is framed as a reinforcement‑learning problem. The key innovation is the Relative Outcome Reward, which compares the correctness of the Direct prediction (acc_dir) with that of the tool‑augmented prediction (acc_tk) on the same example. The reward is +5 if the tool makes a previously wrong answer correct, –5 if the tool harms a previously correct answer, a small negative penalty (–0.1) for redundant calls, and 0 otherwise. Direct actions receive +1 for a correct answer and –1 for an incorrect one. This relative formulation aligns the learning signal precisely with the goal of using tools only when they provide a genuine performance gain, thereby suppressing both surface‑keyword bias (choosing tools based on superficial word overlap) and hallucination of tool capability boundaries (calling tools beyond what they can actually compute).

Because only the Router’s parameters are updated, the learning problem is low‑dimensional, leading to extreme data efficiency. Experiments on several standard audio reasoning benchmarks (including MMAR, AudioQA, and ESC‑50‑based QA) show that AudioRouter improves accuracy by 3–7 percentage points over a frozen LALM baseline and outperforms end‑to‑end fine‑tuning that uses 600× more training examples for tool‑use learning. The method also dramatically reduces unnecessary tool calls: in ablation studies, the policy avoids over‑use of tools that do not improve the answer, and it respects tool capability limits in 92 % of cases.

The paper also discusses limitations. The Router can only select from a predefined toolbox; adding new tools requires re‑training or fine‑tuning the policy. Computing the relative reward necessitates running both Direct and tool‑augmented inference for each training instance, which adds overhead. Moreover, the current design supports a single tool call per query; extending to multi‑step tool chains would require more sophisticated policy architectures and reward designs.

In summary, AudioRouter demonstrates that learning to orchestrate external perceptual tools is a far more data‑efficient and scalable route to high‑quality audio understanding than attempting to embed all perceptual skills directly into the language model. By decoupling tool selection from reasoning and using a relative outcome reward, the framework achieves state‑of‑the‑art performance with minimal data, opening a promising direction for future multimodal and tool‑augmented AI systems.