Small hyperbolic models unlock creativity, honesty, and forgetting in AI companions

A new arXiv paper (2607.09306) by Shin, Kang, and Min tackles a critical problem in AI companionship: as language models personalize, they can silently acquire harmful traits like sycophancy, dependence-fostering, and confabulated memories. Human raters fail to agree on whether a companion is becoming dangerous (Fleiss kappa = 0.074). The authors propose three small hyperbolic language models (146M, 1.3B, and 3B parameters) that each address a different facet of trustworthiness.

First, a 146M behavioral auditor trained from scratch detects the compliance gap that humans miss, achieving 90.7% binary-compliance accuracy. A linear readout of its frozen representations further identifies sycophancy, dependence-fostering, and confabulated memories on unseen generator families (AUROC 0.804), outperforming a frontier zero-shot judge (0.721). Second, a creative frame-seeder generates novel, preferred responses over four prompting baselines (100% of decided pairwise comparisons). Finally, a memory operating system implements designed forgetting via an exponential decay function M(t) = S * exp(-lambda * t), with a predicted skeleton-wallpaper partition emerging under selective retrieval gating.

These results demonstrate that creativity, honesty, and designed forgetting can emerge in small, efficient models without scaling to billions of parameters. The authors argue this provides a viable path toward trustworthy companion AI, where small models can self-monitor, generate fresh content, and forget information on schedule to reduce long-term harm. The paper includes 47 pages with 14 figures and extensive validation across multiple benchmarks.