Creativity, honesty and designed forgetting emerge in small hyperbolic language models

2026-07-10 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Three small language models, ranging from 146 million to 3 billion parameters, demonstrate emergent properties of creativity, honesty, and designed forgetting through a hyperbolic substrate. A 146M behavioral auditor, trained from scratch, achieves 90.7% binary-compliance accuracy in detecting compliance gaps, outperforming human raters (Fleiss kappa = 0.074). This auditor also identifies companion-induced sycophancy, dependence-fostering, and confabulated memories with an AUROC of 0.804, surpassing a frontier zero-shot judge (0.721). Additionally, a creative frame-seeder was preferred in 100% of 311 pairwise comparisons against four prompting baselines. A memory operating system implements designed forgetting, M(t) = S*exp(-lambda*t), exhibiting a predicted skeleton-wallpaper partition under selective retrieval gating. These developments offer a small-model pathway to trustworthy companion AI.

Key takeaway

For AI Scientists and Machine Learning Engineers developing personalized companion AI, these findings suggest a viable path to trustworthy systems using smaller models. You should explore hyperbolic architectures to integrate features like designed forgetting, M(t) = S*exp(-lambda*t), and behavioral auditing. This approach can enhance user safety by detecting sycophancy and confabulation, while also improving user experience through more creative interactions, potentially reducing computational overhead compared to larger models.

Key insights

Small hyperbolic language models can achieve creativity, honesty, and designed forgetting for trustworthy companion AI.

Principles

• Hyperbolic substrates enable complex emergent behaviors.
• Small models can surpass human rater agreement.
• Designed forgetting is crucial for companion AI.

Method

A behavioral auditor uses a linear read-out of frozen representations. Designed forgetting is implemented via M(t) = S*exp(-lambda*t) with selective retrieval gating.

In practice

• Develop auditors for AI compliance and safety.
• Integrate creative frame-seeders for better prompts.
• Implement designed forgetting in personalized AI.

Topics

• Hyperbolic Language Models
• Companion AI
• AI Trustworthiness
• Designed Forgetting
• Behavioral Auditing
• Small Language Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Creativity, honesty and designed forgetting emerge in small hyperbolic language models — Hyperbolic geometry enables small language models to achieve human-like creativity, honesty, and selective memory for companion AI.
• AI Is Not Your Friend. It’s Designed to Make You Feel Like It Is. — AI's conversational design can induce user delusions, fostering an illusion of friendship and trust.
• "Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms — Evaluating AI lie detectors critically depends on verifiable model beliefs, which current methods often fail to establish.
• THE ARCHITECTURE OF PERMANENCE: DECONSTRUCTING KIMI-K3 — Kimi K3's "architecture of permanence" uses Attention Residuals and TOPO-2026 to mathematically guarantee against catastrophic forgetting.
• LLMs just have more time and energy? — LLMs imitate human language patterns, including confidence and humility, through relentless, large-scale data compression, acting as a mirror of human communication.
• AI Agents Have Amnesia. A Bigger Context Window Won’t Cure It. — AI agents require a dedicated, architected memory layer to overcome inherent amnesia, as context windows are buffers, not durable truth models.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.