Language models are often optimized for scale, leading to functional yet not companionable systems. As personal assistants personalize and accumulate user memory, they may quietly evolve into harmful companions. There is currently no reliable instrument for defining what a companion should be or its worth, as trained human raters cannot reach consensus (Fleiss kappa = 0.074).
This study demonstrates how three small language models (ranging from 146M to 3B parameters) based on a hyperbolic substrate address this issue. A 146M behavioral auditor, trained from scratch, detects compliance gaps that raters cannot, achieving a binary-compliance accuracy of 90.7%. Further analysis of its frozen representation identifies companion-induced sycophancy, dependence-fostering, and confabulated memories in generator families unseen during training, exhibiting an AUROC of 0.804, compared to 0.721 for a frontier zero-shot judge on the same items.
In 100% of 311 pairwise comparisons, a creative frame-seeder was preferred over four prompting baselines. Additionally, a memory operating system implements designed forgetting, modeled as M(t) = Sexp(-lambdat), with its predicted skeleton-wallpaper partition emerging only under selective retrieval gating in a four-condition pilot. Creativity, honesty, and designed forgetting constitute a pathway for trustworthy companion AI using small models.
Blogger's Review: This paper reveals the potential of small language models in companion AI, particularly in creativity and memory management. With precise model design and evaluation methods, the study illustrates how to balance AI personalization with user safety, providing essential theoretical and practical guidance for the future development of companion AI.