Toward Societies of Agents: Reasoning, Debate, and Collaborative Intelligence

Speaker: Anoop Cherian, Senior Principal Research Scientist, Mitsubishi Electric Research Labs (MERL), Cambridge, MA

Date: 06 July 2026

YouTube link: https://youtu.be/g9PStqCi5Zo

The next frontier of artificial intelligence may not be a single monolithic model, but a society of agents that can reason, disagree, coordinate, and learn through interaction. While large language and vision–language models have shown remarkable capabilities, they remain fragile as isolated decision makers: they hallucinate, miscoordinate with partners, and often struggle to adapt to collaborative tasks.

This talk asked a central question: how can agent interaction become a computational principle for building more reliable, reflective, and cooperative AI systems? Anoop Cherian presented two complementary lines of work toward this goal.

In WISE, reasoning is studied as structured debate among heterogeneous foundation models. Rather than treating disagreement as noise, WISE organises models into specialised roles—solvers, critics, and aggregators—and uses multi-round interaction with probabilistic solution aggregation to turn agent diversity and uncertainty towards robust multimodal inference.

In LLawCo, this perspective is extended from debate to embodied cooperation. In decentralised, partially observable environments, agents must not only reason correctly, but also coordinate their actions and communication with partners. LLawCo learns high-level behavioural laws of cooperation from failed interactions, transforming recurring miscoordination into explicit priors for future planning, communication, and action. Together, these works advance a broader technical perspective: moving from foundation models as isolated predictors to agentic societies that reason through mutual interaction. Across multimodal reasoning and embodied collaboration benchmarks, Anoop Cherian showed that this structured interaction substantially improves robustness and achieves state-of-the-art performance, pointing toward AI systems that are not merely larger, but more collaborative, adaptive, and reliable.

MENU