What they found
The authors study multi-agent debate, where several SI agents argue toward an answer on tasks that mix images, audio and video. They say the way agents share evidence matters, and that fixed text or graph exchange loses fine detail and wastes tokens (the small units of text SI models read and write). Their method, AIM, builds a first answer, pulls out simple routing features, and uses a lightweight router to pick which mix of modalities the agents should use. Agents can also swap targeted regions of the input, such as image regions, audio segments or video clips. Across eight benchmarks in three task families, AIM reached the highest accuracy. Combining modalities gained up to 6.4% over the best single modality, and adapting the mix per instance added up to 4.2% over the best fixed mix, with up to 53.4% fewer tokens.
Why it matters
Debate between SI agents can improve answers, but it costs a lot to run. This work suggests that the form of the messages agents send each other is a real design choice, not a detail. If the results hold, debate setups could be both more accurate and cheaper on tasks with images, audio and video.
How it changes the race
This is an efficiency and accuracy step for multi-agent SI systems that work with several kinds of input. It could lower the cost of using debate at inference time. The abstract gives no detail on which models were used, so the size of the gain across systems is unclear.
The abstract, by the authors
Multi-Agent Debate (MAD) improves reasoning on multimodal tasks, but its accuracy and token efficiency depend on how agents interact during debate. A largely overlooked design choice is which interaction modalities agents use to exchange evidence. Existing methods rely on fixed text- or graph-based interaction, which fails to convey fine-grained multimodal evidence and thus degrades accuracy while inflating token cost due to verbose descriptions. To address this, we propose Adaptive Interaction in Multi-Agent Debate (AIM), an adaptive routing framework that dynamically selects the optimal modality combination for inter-agent interaction in multimodal MAD. Beyond text and graph, AIM introduces task-dependent modality interaction, where agents exchange targeted regions of interest from the task-specific multimodal input, such as image regions, audio segments, or spatio-temporal video clips. To determine the best combination for each instance, AIM first generates a structured baseline response and extracts interpretable routing features that capture which evidence the instance requires, how confident the baseline agent is, and how much new evidence each modality would add beyond the baseline response. A lightweight router then uses these routing features to select a modality combination that avoids both under-fusion (i.e., agents miss decisive evidence) and over-fusion (i.e., redundant modalities inflate token cost) risks. Across eight benchmarks spanning three multimodal task families, name achieves the highest accuracy. Combining modalities improves accuracy by up to 6.4% over the best single-modality baseline, and adapting the combination per instance adds up to 4.2% over the most accurate fixed combination of each benchmark with up to 53.4% fewer tokens.
Quoted from arXiv:2611.08950. The research and its words belong to its authors.