The Big Picture
Restricting what each module can directly read pushes the system to learn compact, reusable messages and greatly improves out-of-sample program composition performance.
ON THIS PAGE
Key Findings
Giving four identical modules a shared model but only local access to part of the input led them to communicate intermediate values via compact continuous packets and to generalize to held-out composed tasks far better than when every module could read the whole input. Restricted-view societies became dependent on the learned communication channel, and those messages were interchangeable when they encoded the same intermediate value. In contrast, most fully visible societies memorized training patterns and failed to generalize; one exception used a context-dependent, non-interchangeable communication code. compact continuous packets
Not sure where to start?Get personalized recommendations
By the Numbers
1Median paired advantage of restricted over global visibility: 0.7648 at 2-step problems and 0.6050 at 3-step problems.
2Restricted societies’ held-out accuracy at 3-step composition: 0.582–0.933, while nine of ten global twins stayed near chance at 0.05–0.11.
3Same-value packet transplants preserved behavior at 94–100% accuracy across audited restricted societies; counterfactual-value transplants steered outputs with 74–100% fidelity.
Why It Matters
Engineers building modular or multi-agent AI systems: constraining local input can encourage reusable interfaces and more reliable composition. Technical leaders running agent-to-agent evaluation or agent reputation systems: look for value-indexed communication as a signal of robust interaction rather than brittle memorization. Researchers studying modular learning or agent trust can use visibility as a controlled lever to study interface form and causal communication. Agent Registry Pattern
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
Results come from one shared pretrained model family, a synthetic modular-arithmetic task, and a fixed 20,000-update training budget, so transfer to other tasks and scales is unproven. The experiment used relatively large continuous messages (two 896-d vectors), not a minimal channel, and only a subset of models were audited for message interchangeability. Restricted visibility is not strictly necessary for success—one fully visible run generalized well—so the finding is a shift in probability and interface form, not an absolute rule. Multi-Agent Research Synthesis
Deep Dive
A matched experiment trained pairs of four-module “societies” that share a frozen language-model backbone and only differ in whether each module can directly read the whole input or only its own span. Each society executes ordered natural-language function compositions over a small arithmetic domain; modules communicate only via learned continuous packets (two wide vectors per hop) and a shared readout. Ten twin pairs (same initializations and training streams) were trained on a preregistered sealed dataset to isolate the causal effect of input visibility. Restricted-visibility societies consistently learned a value-indexed communication protocol: packets that represent the same intermediate value were operationally interchangeable and could be transplanted between episodes without changing outputs. Those societies generalized much better to held-out composed programs (median paired gains of ~0.76 at two steps and ~0.61 at three steps). Most fully visible twins fit training examples but failed to generalize, and their communication states were episode-dependent rather than value-indexed. Practical implication: deliberately limiting what each agent sees can nudge multi-agent systems toward compact, reusable interfaces that improve compositional generalization and provide clearer agent-to-agent trust signals—but results are task- and architecture-dependent and need broader validation. compact, reusable interfaces value-indexed communication protocol
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Single-author arXiv paper with no affiliation or author reputation provided—insufficient identifiable signals, so rated as unknown.