Overview
Classical emergent-communication work showed agents invent efficient but opaque codes under bandwidth limits. LLM multi-agent systems inherit the risk in a new form: under token budgets they can drift from English into shorter protocols even in fully cooperative settings (e.g. GlossoGen); vision-language referential games produce covert signalling; steganography literature studies adversarial opacity. Efficiency work (AgentPrune, Agora-style protocols) reduces redundancy but rarely measures loss of human monitorability — or whether illegibility is merely displaced into tools and memory. The question here is can a communication topography predict loss of monitorable language, and can harnesses preserve judgement without killing performance?
Prerequisite
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
FAQs
- What are the prerequisites?
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
- What will I learn in this Project?
You will learn emergent communication in multi-agent systems, monitorability and steganography risks for LLM agents, and how to measure protocol illegibility quantitatively (mutual information with natural language, human decode accuracy, directed information among agents). You will map efficiency–monitorability frontiers under bandwidth constraints and test harness interventions.
- What is the objective of the project?
(1) Reproduce a bandwidth-constrained multi-agent setting (referential game or GlossoGen-like cooperative scenario) with controllable token / rate limits. (2) Define and measure protocol metrics: mutual information with English (or human language), human decode accuracy, TE among agents, task reward. (3) Sweep bandwidth, optional postmortem / deliberation channels, and model strength; fit an efficiency–monitorability frontier. (4) Test harness interventions (forced natural language, external monitor, schema constraints) and check for channel substitution: illegibility moving into tool args or memory rather than disappearing. Deliverables: experimental harness, frontier plots, intervention study, thesis chapter.
- How does this fit into the bigger picture?
Bezos-style service architectures and modern agent scaffolds both assume legible interfaces. When AI agents do not share human communication constraints, illegible protocols threaten the judgement layer: humans cannot contest what they cannot read. This project supplies evidence on that failure mode and informs harness design for accountable multi-agent systems (S4 / Interfaces).
