Directed interaction changes the output dynamics of two Generative Pretrained Transformer-2 agents
System–bath interactions are central to physics, including regimes with negligible back-action on the bath. One-way AI communication offers a new setting for this idea: an autonomous AI drives another through generated text without receiving its replies. We study how directed text exchange changes the output dynamics of two parameter-identical GPT-2 agents with fixed decoding settings. In 200-round runs on an Earth-shape topic, one-way interaction increases the receiver’s fraction of outputs assigned to an operational unusable-text category, although the autonomous sender rarely produces that category. Reversing communication changes which agent is affected; under two-way interaction, both agents have elevated fractions. An independently generated prerecorded sender produces a comparable response. Additional self-history produces a smaller increase in the fraction over all rounds, but a similar change in composition among nonfallback outputs, distinguishing generation failure from output composition. A three-state phenomenological kinetic model describes the qualitative occupancy changes under explicit assumptions on transition rates and illustrates how message order can affect a finite-time response. These results characterize directed interaction in a controlled base-model setting; generalization to other models and tasks remains open.
Authors
- Frank Yingjie Huo (ORCID: https://orcid.org/0009-0008-2840-7204)
- Bella Xinrui Li
- Neil F Johnson
Publication Details
- Journal
- APS Open Science
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1103/nqld-61xd
- Primary Topic
- Advanced Thermodynamics and Statistical Mechanics
- Type
- article
- Field-Weighted Citation Impact
- 0.00