Research map / Vision-language and multimodal / Multimodal LLMs

Audio-Thinker: Guiding Large Audio Language Model When and How to Think via Reinforcement Learning

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Audio-Thinker, a GRPO-based RL framework for large audio language models that learns when to think (adaptive think accuracy reward) and how to think (consistency and model-judged think rewards), reaching SOTA on MMAU, MMAR and AIR-Bench.

Problem

Explicit chain-of-thought reasoning has not yielded clear gains for large audio language models on audio question answering, and existing LALMs lack adaptive thinking and supervision of their reasoning process.

Approach

Combine an adaptive think accuracy reward (four think/no-think correct/incorrect cases with batch-level soft penalty factors that decay over training) with a consistency reward from a Qwen3-8B-Base judge and a model-judged think-quality reward, all within GRPO.

Key results

• Audio-Thinker (Qwen2.5-Omni) reaches 73.70 average on MMAU test-mini versus 66.30 for the Qwen2.5-Omni baseline and 71.3 for Omni-R1 (per Table 2).
• Audio-Thinker (Qwen2.5-Omni) reaches 65.30 on MMAR, outperforming Omni-R1 (63.4) and Qwen2.5-Omni-7B (56.70).
• Audio-Thinker (Qwen2.5-Omni) averages 67.1 on AIR-Bench foundation, above GPT-4o Audio (62.3), though below Gemini 2.5 Flash (67.4).

Surprising findings

• Explicit reasoning variants (grpo-b, grpo-d with CoT) did not outperform implicit-reasoning GRPO counterparts, suggesting explicit reasoning alone gives little benefit without supervision.
• Prompt-only forcing of thinking shows no sensitivity to question difficulty, while the trained model's no-think rate tracks difficulty.

Limitations

• Training uses only the AVQA-derived audio QA data (40,176 samples), limiting diversity.
• Relies on a Qwen3-8B-Base model as both consistency and think-quality judge, adding dependence on a judge model.
• Still trails Gemini 2.5 Flash on AIR-Bench foundation average.

Experimental setup

Datasets: AVQA

Benchmarks: MMAU test-mini MMAU test MMAR AIR-Bench foundation MMAU v05.15.25

Metrics: accuracy

Baselines: Qwen2-Audio-7B-Instruct Qwen2.5-Omni-7B R1-AQA SARI Omni-R1 Audio-Reasoner Audio-CoT Gemini 2.5 Flash GPT-4o Audio Mellow Audio Flamingo 2 Kimi-Audio

Builds on: GRPO DeepSeek-R1 R1-AQA AutoThink R1-Reward SophiaVL-R1

Closest papers by meaning

• Incentivizing Consistent, Effective and Scalable Reasoning Capability in Audio LLMs via Reasoning Process Rewards (ICLR 2026 · similarity 0.93)
• Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models (ICLR 2026 · similarity 0.91)
• MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix (NeurIPS 2025 · similarity 0.90)
• Thinker: Learning to Think Fast and Slow (NeurIPS 2025 · similarity 0.89)
• Mellow: a small audio language model for reasoning (NeurIPS 2025 · similarity 0.89)
• Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities (ICML 2025 · similarity 0.89)