Small hyperbolic language models exhibit creativity, honesty, and designed forgetting

A study demonstrates that small language models (146 M to 3 B parameters) with a hyperbolic substrate can exhibit creativity, honesty, and designed forgetting, addressing the lack of reliable instruments for evaluating companion AI.

• A 146 M behavioral auditor detects compliance gaps with 90.7% binary-compliance accuracy, outperforming trained human raters who showed low agreement (Fleiss kappa = 0.074).
• The auditor identifies companion-induced sycophancy and confabulated memories on unseen generator families with an AUROC of 0.804, compared to 0.721 for a frontier zero-shot judge.
• A creative frame-seeder was preferred in 100% of 311 decided pairwise comparisons against four prompting baselines.
• A memory operating system implements designed forgetting using the formula M(t) = S*exp(-lambda*t), with its predicted skeleton-wallpaper partition emerging under selective retrieval gating.

The authors conclude that these traits provide a small-model route to trustworthy companion AI.