Brief summary: 

The Interspeech 2026 Audio Reasoning Challenge addresses the limited and unstable reasoning capabilities of current Large Audio Language Models (LALMs) by focusing on Chain-of-Thought (CoT) reasoning in complex acoustic scenarios. It builds upon an enriched MMAR benchmark with manually labeled CoT annotations and explicit reasoning cues. The evaluation criterion is strict: a prediction is correct only if both the reasoning path and the final answer are accurate. The goal is to push LALMs beyond surface-level accuracy toward logically consistent thought processes. The challenge features two complementary tracks: Single Model Track (focusing on intrinsic model reasoning via post-training of open-source models) and Agent Track (focusing on system-level orchestration and tool use with open-source models). 

A. Track 1: Single Model Track

Participants build a single, end-to-end Audio–Language Model that consumes the MMAR audio and produces. Rules and Restriction:

1) Open-source data only. Training must use datasets with audio publicly-available with research-permissive licenses; all external data sources must be declared.

2) Reproducibility. Provide full inference code and configuration. Finalists must submit training recipes and model weights (or a research-permissive checkpoint link) for on-site verification.

3) No external tools/APIs. No retrieval, web search, calculators, ASR/TTS calls, or plug-in agents at inference.

4) Transparency. The CoT must be emitted by the model; handcrafted post-hoc CoT editing is forbidden.

B. Track 2: Agent Track

Participants design an audio reasoning agent that may orchestrate multiple open-source models and tools 

Rules and Restriction:

1) Open-source models/tools only. All components (including music/audio tools) must be publicly available under research-permissive licenses; no commercial or hosted APIs. 

2) Full system + trajectory submission. Finalists must submit runnable code, environment specs, and complete execution logs (tool calls, prompts, intermediate outputs) for audit.

3) No human-in-the-loop. Inference-time human assistance or manual curation is strictly prohibited.

• Ziyang Ma (Shanghai Jiao Tong University & Nanyang Technological University)
• Yinghao Ma (Queen Mary University of London)
• Chao-Han Huck Yang (NVIDIA Research)
• Ruiyang Xu (Shanghai Jiao Tong University)
• Bohan Li (Shanghai Jiao Tong University)
• Jaeyeon Kim (Carnegie Mellon University)
• Jin Xu (Alibaba Group)
• Jinyu Li (Microsoft Corporation)
• Carlos Busso (Carnegie Mellon University)
• Kai Yu (Shanghai Jiao Tong University)
• Eng Siong Chng (Nanyang Technological University)
• Xie Chen (Shanghai Jiao Tong University)

Key dates for the challenge: 

Nov. 1st, 2025: Collect potential participants.

Dec. 1st, 2025: Challenge announcement with detailed rules, with challenge data available. 

Jan. 1st, 2026: Leaderboard is online and accepting submissions. 

(Registration Deadline: Jan 15th for participating the leaderboard ranking)

Feb. 1st, 2026: Leaderboard submission deadline. Top teams will be located by the ranking this time.

Feb. 10th, 2026: Check code and provide final rankings, announce the winners.

Feb. 25th, 2026: Paper submission deadline (with the same process as the main conference peer review). 

Link to challenge website:

https://audio-reasoning-challenge.github.io/