Distributed intelligence

Core objective

This research direction investigates a paradigm where cognition and decision-making are not centralized within a single model, but are instead distributed across multiple autonomous agents that operate independently while communicating in natural language. In this setup, each agent relies on its own local knowledge, perspective, and computational resources, allowing global intelligence to emerge organically from their interactions. A guiding analogy for this system is a network of sensors where each sensor functions as a chatbot-like agent that can be queried directly in English.

Fundamental advantages

Studying distributed intelligence provides several key benefits over monolithic, centralized systems: * Energy efficiency: computation is distributed, meaning each agent performs only localized reasoning. * Explainability: the internal states, reasoning, and rationales of the agents are explicitly verbalized through natural language communication. * Modularity and hot-plugging: new agents can seamlessly join or leave the system on the fly. * Robustness: the failure of a single agent does not compromise the operational integrity of the entire system. * Scalability: the collective intelligence and capability of the system grows as the number of participating agents increases.

Research Focus: Controlling Multi-Agent Collaboration

Unlike traditional LLM research that focuses either on single-agent reasoning or fully aligned collaborative agents, this project seeks to understand how independent agents coordinate under conditions of partial alignment, differing perspectives, or even adversarial incentives. The research studies how agents use natural language channels to negotiate uncertain positions, selectively share information, and resolve conflicting preferences to converge toward shared goals.

Primary Use-Cases

1. Multi-Agent Chess Environment

As an initial, controlled test-bed, the project utilizes a multi-agent chess environment where each individual piece acts as an independent agent that debates with other pieces to decide which move the team should execute. This environment offers a fully observable, deterministic domain with clear rules, allowing researchers to precisely evaluate how agent-level reasoning scales into coherent, high-level group behavior. It provides a clean platform to conduct systematic experiments on communication protocols, negotiation, and conflict resolution before extending the framework to partially observable domains.

2. Automotive Sensor Networks

To transition this concept to complex, real-world applications, the project explores multi-modal perception in autonomous driving. Instead of relying on traditional centralized late-fusion pipelines, individual sensors (such as cameras, LiDAR, and radar) operate as autonomous agents that describe their observations in natural language. These sensor-agents maintain their own belief states, report local measurements, and dynamically exchange hypotheses with neighboring agents to negotiate uncertainty and build global situational awareness.

This decentralized approach allows the system to reach a consensus about objects, trajectories, and hazards through structured debate (e.g., arguing whether a shape is a pedestrian or a static obstacle). Ultimately, this makes the perception system more robust to occlusions, highly adaptive to sensor degradation, and fully interpretable under tight latency constraints.