A graph-based AI agent scored 65.3% on inferential questions drawn from AMI meeting recordings, compared with 24.6% for a text-retrieval system, according to an arXiv preprint. These questions asked the system to work out an answer from the conversation, rather than simply locate a directly stated fact. The result comes from a benchmark evaluation, not a test of whether the system improved real meetings.
The system, called GRGA, is designed to address acoustic information loss and the difficulty of keeping track of long-term context in meeting audio. It represents a recording as a multi-dimensional graph, linking semantic, temporal and speaker relationships, and combines query decomposition, planning, execution, synthesis and reflection. The benchmark spans factual, inferential, temporal and acoustic-aware questions rather than a single kind of answer.
A benchmark built from three audio collections
LongAudioQA was built from AliMeeting, AMI Meeting and DailyTalk. The source corpora contained 14.91 hours, 18.24 hours and 21.59 hours of audio, respectively, and 3,013, 3,243 and 10,200 question-answer pairs. The paper reports those figures at corpus level; it does not give one pooled participant or item total.
The benchmark uses timestamp evidence to anchor answers to particular moments in the recordings. The required overlap with the target span was greater than 90%, and reported inter-annotator agreement was 0.91. In plain terms, the people checking the evidence showed a high level of consistency in the reported quality-control measure.
The system plans before it answers
GRGA’s workflow is built around deciding what to retrieve before composing a response. It breaks a question into parts, plans a retrieval route, executes the search, synthesizes the findings and reflects on the result. Its graph preserves connections between meaning, timing and speaker identity, giving the system several ways to locate relevant evidence in a long exchange.
Automated scoring used a high-capability language model as a judge to decide whether an answer was semantically equivalent to the reference. Accuracy was calculated as the percentage of answers judged correct. A separate blind human evaluation randomly sampled 150 queries across five question types from the three corpora; three graduate-student annotators rated them, with average agreement of 0.78.
The strongest gap came on inference
Across the displayed datasets and question categories, the authors report GRGA performance against several other systems. On AMI, they report that GRGA outperformed MiMo-Audio. The category-level inferential score was 65.3% for GRGA versus 24.6% for Text RAG. No confidence interval or formal significance test was reported for that comparison.
GRGA also reported citation scores between 69.7% and 84.2% across six displayed precision and recall columns. The six values were 84.2%, 77.2%, 76.2%, 69.7%, 79.1% and 74.1%. These measures assess the match between cited evidence and the benchmark’s timestamp spans, so they speak to grounding rather than replacing the separate answer-accuracy score.
Human ratings pointed in the same direction. Compared with Qwen3-Omni, GRGA’s reported overall improvements were 0.52 points in correctness, 0.97 in groundedness and 0.30 in coherence. The comparison used the randomly sampled queries and the three annotators described in the evaluation.
Removing pieces weakened the result
Component-removal tests on AMI gave full GRGA an overall accuracy of 49.49%. Without reflection, the corresponding score was 39.64%; without semantic search, 16.28%; without graph traversal, 38.21%; and without audio access, 34.71%. The version without semantic search had the lowest score in this set.
These figures show how the complete pipeline compared with versions missing a feature, but they do not establish that one component alone caused the difference. GRGA combines graph structure, planned retrieval, reflection and audio access, so the ablation changes are best read as evidence about the package tested.
A tougher test, with a caveat
Under the paper’s High/High simulated-noise condition, GRGA scored 39.54%, compared with 24.67% for Text RAG. The noise analysis used a 3 x 3 grid of simulated transcription and speaker-separation errors. That result suggests better retention under the tested stress condition, but it does not establish robustness to every naturally occurring audio problem.
What the preprint leaves open
The evidence remains limited to a constructed benchmark based on three meeting and dialogue corpora. The evaluation does not establish real-world deployment outcomes or generalization beyond structured meeting audio. The iterative planning and reflection workflow may also carry higher computational cost than a single-turn retrieval system, which could matter for real-time use.
Automated language-model judging means the results remain partly dependent on the evaluation model, alongside the human check. The document is marked arXiv version 1 and dated 25 August 2026. It says the work was supported by Jiangsu Province Frontier Program Project BF2025036 and Hong Kong RGC grant GRF #15611021.
Paper data and sources
Original title: Don't Just Listen, Try Planning: Graph-based Retrieval-Generation Agent for Long-form Audio Meeting Understanding
Authors: Quanwei Tang, Dong Zhang, Shoushan Li, Guodong Zhou
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text