RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models with a CHC-Grounded Benchmark

NeurIPS 2026

Hongyu Jin^1,*, Siyi Wang^1,*, Yang Xiao^1,*, Jiaheng Dong^1,*, Shihong Tan⁴, Kaiyuan Peng¹, Georgiana Juravle², Shanquan Chen³, Gongping Huang⁴, Hong Jia⁵, Eun-Jung Holden¹, James Bailey⁶, Ting Dang^1,†

¹The University of Melbourne, ²Alexandru Ioan Cuza University of Iași, ³The University of Hong Kong, ⁴Wuhan University, ⁵The University of Auckland, ⁶Monash University

^*Equal contribution. ^†Corresponding author.

Contact: Hongyu Jin (hongyuj1@student.unimelb.edu.au), Ting Dang (ting.dang@unimelb.edu.au)

Project page · arXiv · Paper on Hugging Face · Dataset on Hugging Face

TL;DR: RAIL evaluates large audio-language models the way cognitive science assesses human listeners. Grounded in Cattell–Horn–Carroll (CHC) theory, it organizes 5,306 audio questions into five core auditory capabilities and 32 subcapabilities, and compares 26 LALMs with a human baseline. Models do well on knowledge inherited from text pretraining, while fine-grained auditory perception and processing efficiency remain below human level.

Taxonomy and statistics

Capability	Code	Tasks	Samples	Total dur. (h)	Mean dur. (s)	Median dur. (s)	Config
Auditory Processing	Ga	7	1170	4.98	15.31	4.0	auditory_processing
Reasoning	Gf	3	322	3.49	39.04	25.04	reasoning
Memory	Gsm/Glr	6	1000	12.77	45.99	28.04	memory
Processing Efficiency	Gs	9	1800	2.01	4.02	1.81	processing_efficiency
Knowledge	Gc/Gkn	7	1014	7.18	25.48	10.49	knowledge
Overall		32	5306	30.43	20.65	8.0	all

Durations are summed over all clips of an item and computed from the audio files.

Tasks and subtasks

Every task is also a config, named by its CHC code (U1/U9 → U1_U9).

Capability	Task (code)	Samples	Answer format	Subtasks (samples)
Auditory Processing	Speech Sound Discrimination (US)	153	multiple_choice	Emotional Prosody Comparison (55); Confusable Word Identification (58); Minimal-Pair Discrimination (40)
Auditory Processing	Resistance to Auditory Stimulus Distortion (UR)	181	multiple_choice	Babble Noise (65); Band-Limiting (40); Background Music (36); Reverberation (40)
Auditory Processing	Maintaining and Judging Rhythm (U8)	160	multiple_choice	Beat Regularity Detection (40); Meter Identification (38); Tempo Change Detection (30); Tempo Comparison (52)
Auditory Processing	Absolute Pitch (UP)	200	multiple_choice	MIDI Pitch Identification (80); Note Name Identification (50); Relative Pitch Comparison (40); Frequency Estimation (30)
Auditory Processing	Musical Discrimination and Judgment (U1/U9)	102	multiple_choice	Aesthetic and Emotional Judgment (8); Genre Recognition (35); Instrument Identification (50); Musical Structure Analysis (9)
Auditory Processing	Sound Localization (UL)	182	multiple_choice	Direction Identification (82); Azimuth Estimation (50); Distance Estimation (30); Motion Trajectory (20)
Auditory Processing	Phonetic Coding (PC)	192	multiple_choice	Isolated Phoneme Identification (42); Concurrent Phoneme Segregation (50); Staggered-Onset Phoneme Identification (50); Stereo-Separated Phoneme Identification (50)
Reasoning	Induction (I)	68	multiple_choice	Pattern Abstraction (34); Rule Induction (34)
Reasoning	General Sequential Reasoning (RG)	154	multiple_choice	Sequential Reasoning with General Rules (124); Auditory Cognitive Puzzle (30)
Reasoning	Quantitative Reasoning (RQ)	100	open_ended, multiple_choice	Math Reasoning (50); Counting (50)
Memory	Memory Span (MS)	150	multiple_choice	—
Memory	Associative Memory (MA)	150	multiple_choice	—
Memory	Meaningful Memory (MM)	150	multiple_choice	Multi-Turn Dialogue (100); Narrated Passage (50)
Memory	Free Recall Memory (M6)	200	free_recall	—
Memory	Memory for Sound Patterns (UM)	200	closed_label, multiple_choice	N-Back (70); Speaker Tracking (70); Prosodic Contour Matching (60)
Memory	Working Memory (WM)	150	multiple_choice	—
Processing Efficiency	Perceptual Speed (P)	200	closed_label	—
Processing Efficiency	Rate-of-Test-Taking (R9)	200	closed_label	—
Processing Efficiency	Number Facility (N)	200	closed_label	—
Processing Efficiency	Reading Speed (RS)	200	closed_label	—
Processing Efficiency	Simple Reaction Time (R1)	200	closed_label	—
Processing Efficiency	Choice Reaction Time (R2)	200	closed_label	—
Processing Efficiency	Semantic Processing Speed (R4)	200	closed_label	—
Processing Efficiency	Mental Comparison Speed (R7)	200	closed_label	—
Processing Efficiency	Inspection Time (IT)	200	closed_label	—
Knowledge	General (Verbal) Information (K0)	143	multiple_choice	—
Knowledge	Language Development (LD)	181	multiple_choice	—
Knowledge	Listening Ability (LS)	150	multiple_choice	—
Knowledge	Foreign Language Proficiency (KL)	193	multiple_choice	—
Knowledge	Geography Achievement (A5)	65	multiple_choice	—
Knowledge	Mechanical Knowledge (MK)	141	multiple_choice	Anomaly Detection (47); Machine Identification (47); Machine Function Inference (47)
Knowledge	Knowledge of Behavioral Content (BC)	141	multiple_choice	—

Loading

from datasets import load_dataset
from huggingface_hub import snapshot_download

root = snapshot_download("AIMS-RAIL/RAIL", repo_type="dataset")  # audio + manifests
ds = load_dataset("AIMS-RAIL/RAIL", "all", split="test")        # or "memory", "UL", "U1_U9", ...

item = ds[0]
audio_files = [f"{root}/{p}" for p in item["audio_paths"]]
print(item["prompt"])

Fields

Field	Description
id	Unique item id, <capability>-<task>-<nnnn> (e.g. ga-ul-0001).
capability, capability_code	Core capability and CHC broad-ability code (Ga, Gf, Gsm/Glr, Gs, Gc/Gkn).
task, task_code	Sub-capability (CHC narrow ability) and its code, as named in the paper.
subtask	Finer task design within a sub-capability; null where the task has a single design.
answer_format	multiple_choice, closed_label (answer is one of choices, no letters shown), open_ended, or free_recall.
question	Question stem without options or output instructions.
choices	Option texts without letter prefixes (multiple_choice), or the allowed output labels (closed_label).
answer	Gold answer text. For free_recall, one target item per line.
answer_label	Gold option letter for multiple_choice; null otherwise.
prompt	Full model-facing prompt used in the paper's evaluation (question, options and output instruction).
audio, audio_paths, num_audio	Primary clip, all clips in presentation order, and clip count (paths are relative to the repo root).
duration_s	Total audio duration of the item in seconds.
source_id	Item id in the source manifest the item was generated from.
legacy_id, legacy_config	Id and config name in RAIL v1, for mapping earlier results.
metadata_json	Original source row as JSON text (generation parameters, conditions, provenance).

Evaluation

The evaluation toolkit (Hugging Face model inference, the paper's scoring rules and an English report) is in the eval/ folder of this repository and of the GitHub repository (https://github.com/AIMS-RAIL/RAIL):

cd eval
python run_rail_eval.py all --model Qwen/Qwen2-Audio-7B-Instruct --out runs/qwen2_audio

Models answer in the form Reason: ...; Answer: ... (reason capped at 20 words, except Reasoning items). Scoring follows the paper:

• Multiple choice / closed label: the answer must contain every token of the gold answer (or its letter) and no token unique to another option (tokenizer [A-Za-z0-9]+). An LLM-as-judge score (GPT-5.4) is reported alongside.
• Processing Efficiency (Gs): B-AUC, the normalised area under accuracy as a function of the reason-length budget (0–50 tokens).
• Free recall (M6): token recall of the target items in answer, pooled over items.
• Open-ended math (RQ, Math Reasoning): the final numeric answer is compared with answer.

Changes from v1

• Labels follow the paper: capability → task (32) → subtask; v1 benchmark/subset/task/ability are replaced.
• Configs are organised by capability and by task instead of by source manifest.
• Ids are unique (v1 had 35 duplicated ids); v1 ids are kept in legacy_id.
• Memory prompts and options are the exact ones used in evaluation (v1 lacked prompts for Memory Span, options for Working/Meaningful Memory prompts, and options for prosodic matching).
• Speaker-tracking items (Memory for Sound Patterns) use the evaluated question wording and options; v1 carried an earlier wording for the same audio.
• Options no longer carry letter prefixes; efficiency tasks list their allowed output labels.
• duration_s is filled for every item from the audio files.
• Audio is stored as audio/<capability>/<task>/<id>[_k].<ext>.
• v1 remains available: load_dataset("AIMS-RAIL/RAIL", revision="v1.0").

Citation

@inproceedings{jin2026rail,
  title     = {RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models
               with a CHC-Grounded Benchmark},
  author    = {Jin, Hongyu and Wang, Siyi and Xiao, Yang and Dong, Jiaheng and
               Tan, Shihong and Peng, Kaiyuan and Juravle, Georgiana and Chen, Shanquan and
               Huang, Gongping and Jia, Hong and Holden, Eun-Jung and Bailey, James and
               Dang, Ting},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.11260}
}

License

CC-BY-4.0. Some items are derived from existing corpora; use them within the terms of those sources.