While Chain-of-Thought (CoT) reasoning has revolutionized text-based LLMs, its application to the audio modality reveals a counter-intuitive thinking paradox: forcing models to “think” via prompting or post-training often degrades performance compared to direct answering. In this work, we investigate the bottleneck of audio reasoning and reveal a critical insight: the primary constraint is not the model’s logical deduction capability, but its inability to accurately perceive acoustic details. We demonstrate that keeping perceptual cues effectively activates the LALM’s internal reasoning potential. Building on this finding, we propose Audio-ReaPer, a post-training framework and corresponding model designed to unlock Reasoning potential by explicitly fortifying Perception. Our approach reforms the post-training paradigm through three key strategies: (1) utilizing agentic data generation pipeline to synthesize perception-enhanced data for SFT; (2) curating perception-focused data for RLVR; and (3) introducing a novel perceptual density reward during RL to stabilize the learning process and balance observation with deduction. Extensive experiments show that Audio-ReaPer establishes a new state-of-the-art on benchmarks such as MMAU, MMAU-Pro, and MMAR. Notably, our 7B model outperforms previous open-source models by significant margins and surpasses proprietary models like GPT-4o Audio and Gemini 2.5 Flash. The model and code will be open-source to facilitate further research for audio reasoning.

Building upon advancements in Large Language Models (LLMs), the field of audio processing has seen increased interest in training speech generation tasks with discrete speech token sequences. However, directly discretizing speech by neural audio codecs often results in sequences that fundamentally differ from text sequences. Unlike text, where text token sequences are deterministic, discrete speech tokens can exhibit significant variability based on contextual factors, while still producing perceptually identical audio segments. We refer to this phenomenon as Discrete Representation Inconsistency (DRI). This inconsistency can lead to a single speech segment being represented by multiple divergent sequences, which creates confusion in neural codec language models and results in poor generated speech. In this paper, we quantitatively analyze the DRI phenomenon within popular audio tokenizers such as EnCodec. Our approach effectively mitigates the DRI phenomenon of the neural audio codec. Furthermore, extensive experiments on the neural codec language model over LibriTTS and large-scale MLS dataset (44,000 hours) demonstrate the effectiveness and generality of our method. The demo of audio samples is available at https://consistencyinneuralcodec.github.io.

Recently, instruction-following audio-language models have received broad attention for human-audio interaction. However, the absence of benchmarks capable of evaluating audio-centric interaction capabilities has impeded advancements in this field. Previous models primarily focus on assessing different fundamental tasks, such as automatic speech recognition, and lack an assessment of the open-ended generative capabilities centered around audio. Thus, it is challenging to track the progression in the Large Audio-Language Models (LALMs) domain and to provide guidance for future improvement.In this paper, we introduce AIR-Bench (Audio InstRuction Benchmark), the first benchmark designed to evaluate the ability of LALMs to understand various types of audio signals (including human speech, natural sounds, and music), and furthermore, to interact with humans in the textual format. AIR-Bench encompasses two dimensions: foundation and chat benchmarks. The former consists of 19 tasks with approximately 19k single-choice questions, intending to inspect the basic single-task ability of LALMs. The latter one contains 2k instances of open-ended question-and-answer data, directly assessing the comprehension of the model on complex audio and its capacity to follow instructions. Both benchmarks require the model to generate hypotheses directly. We design a unified framework that leverages advanced language models, such as GPT-4, to evaluate the scores of generated hypotheses given the meta-information of the audio. Experimental results demonstrate a high level of consistency between GPT-4-based evaluation and human evaluation. By revealing the limitations of existing LALMs through evaluation results, AIR-Bench can provide insights into the direction of future research. Dataset and evaluation code are available at https://github.com/OFA-Sys/AIR-Bench.